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.1005285 | Necrotic Cells Actively Attract Phagocytes through the Collaborative Action of Two Distinct PS-Exposure Mechanisms | Necrosis, a kind of cell death closely associated with pathogenesis and genetic programs, is distinct from apoptosis in both morphology and mechanism. Like apoptotic cells, necrotic cells are swiftly removed from animal bodies to prevent harmful inflammatory and autoimmune responses. In the nematode Caenorhabditis elegans, gain-of-function mutations in certain ion channel subunits result in the excitotoxic necrosis of six touch neurons and their subsequent engulfment and degradation inside engulfing cells. How necrotic cells are recognized by engulfing cells is unclear. Phosphatidylserine (PS) is an important apoptotic-cell surface signal that attracts engulfing cells. Here we observed PS exposure on the surface of necrotic touch neurons. In addition, the phagocytic receptor CED-1 clusters around necrotic cells and promotes their engulfment. The extracellular domain of CED-1 associates with PS in vitro. We further identified a necrotic cell-specific function of CED-7, a member of the ATP-binding cassette (ABC) transporter family, in promoting PS exposure. In addition to CED-7, anoctamin homolog-1 (ANOH-1), the C. elegans homolog of the mammalian Ca2+-dependent phospholipid scramblase TMEM16F, plays an independent role in promoting PS exposure on necrotic cells. The combined activities from CED-7 and ANOH-1 ensure efficient exposure of PS on necrotic cells to attract their phagocytes. In addition, CED-8, the C. elegans homolog of mammalian Xk-related protein 8 also makes a contribution to necrotic cell-removal at the first larval stage. Our work indicates that cells killed by different mechanisms (necrosis or apoptosis) expose a common “eat me” signal to attract their phagocytic receptor(s); furthermore, unlike what was previously believed, necrotic cells actively present PS on their outer surfaces through at least two distinct molecular mechanisms rather than leaking out PS passively.
| Necrosis is a type of cell death often caused by cell injury and is linked to human diseases including neuron degeneration, stroke, and cancer. Necrotic cells undergo distinct morphological changes, including swelling, before being engulfed and degraded by engulfing cells. The clearance of necrotic cells from animal bodies is important for wound healing and for preventing harmful inflammatory and autoimmune responses. However, the mechanisms by which necrotic cells are removed remain elusive. We study the recognition of necrotic neurons in the nematode C. elegans. There is a common belief that the plasma membrane of necrotic cells are ruptured, allowing the detection of phosphatidylserine (PS), a so-called “eat me” signal molecule, by specific transmembrane receptors on the surface of engulfing cells. Contrary to this belief, we found that necrotic neurons actively present PS to their outer surface through two parallel molecular mechanisms, one of which is shared by cells undergoing apoptosis, a “cell suicide” event, whereas the other is unique to necrotic cells. Ca2+-influx, a key factor that triggers necrosis, is implicated in activating a unique PS-scramblase. Our findings reveal novel necrotic cell-specific “eat me” signal-exposure mechanisms and indicate that cells that die through different mechanisms (necrosis and apoptosis) utilize both common and unique mechanisms to attract engulfing cells. They further demonstrate that C. elegans is an effective model system for studying the fate of necrotic cells.
| Cell death during animal development and under pathological conditions is important for removing unwanted cells that are often harmful. Necrosis and apoptosis are two morphologically distinct types of cell death events. Whereas cells undergoing apoptosis display features such as cytoplasm shrinkage, chromatin condensation, nuclear DNA fragmentation, and well-maintained plasma membrane integrity, necrotic cells display cell and organelle swelling, excessive intracellular membranes, and the eventual rupture of intracellular and plasma membranes (reviewed in [1,2]). Necrosis is most frequently observed during cell injury, and is closely associated with diseases such as stroke, neurodegeneration, chronic inflammation, and cancer [3–7]. Although necrosis was historically considered an uncontrolled cell death event caused by acute damage, recent discoveries made in multiple organisms demonstrated that in addition to injury-induced necrosis, cells possess genetic pathways that specifically trigger necrosis in response to extracellular or intracellular stimuli (reviewed in [8–11]). For instance, tumor necrosis factor (TNF) induces a necrosis pathway executed through Ser/Thr kinases [10]. In addition, hyperexcitation of neurons or glial cells induced by the massive release of neurotransmitters or constitutively active ion channels cause excitotoxic necrosis [7,12,13]. Unlike apoptosis, which relies on caspase-mediated death-triggering mechanisms, known necrosis-triggering pathways appear to be independent of caspase-activities (reviewed in [8,14]). On the other hand, like apoptotic cells, in many cases necrotic cells have been observed to be engulfed by phagocytes [15,16]. Efficient clearance of necrotic cells from animal bodies helps to resolve the wounded area; furthermore, cell-corpse removal is essential for reducing harmful inflammatory and auto-immune responses induced by the contents of necrotic cells [15,17]. It is currently unclear how necrotic cells expose the “eat me” signal molecules on their surfaces to attract engulfing cells.
Besides being an excellent model organism for studying the mechanisms of apoptosis and the removal of apoptotic cells [18], the soil nematode Caenorhabditis elegans has also been established as a model for studying necrosis [8,13]. In C. elegans, a number of mutations in the subunits of ion channels, the acetylcholine receptor, and trimeric GTPases induce specific neurons to undergo necrotic cell death that mimics the excitotoxic necrosis, which occurs during stroke, trauma, and neurodegenerative disorders in humans (reviewed in [8]). In particular, specific mutations in multiple genes trigger the necrosis of six mechanosensory (touch) neurons (AVM, PVM, ALML/R and PLML/R) required to sense gentle mechanical stimuli along the body wall [19–21]. Dominant (dm) mutations in mec-4, which encodes a core subunit of a multimeric, mechanically gated sodium channel belonging to the DEG/ENaC family specifically expressed in the touch neurons, lead to hyperactive channel conductivity of Na+ and Ca2+ and induce these neurons to undergo necrosis [19,22]. In mec-4(dm) mutants, the six dying neurons swell to many times their original sizes and develop cytoplasmic vacuoles and large membranous whorls, and are easily distinguishable from living or apoptotic cells under Differential Interference Contrast (DIC) optics by their giant sizes (Fig 1) [16,21]. This type of cell death does not require CED-3 caspase activity [23], and is instead triggered by the influx of Ca2+ into the cytoplasm [22,24]. Despite their distinct modes of triggering cell death, the seven ced genes needed for the engulfment of apoptotic cells are also required for the efficient removal of necrotic touch neurons [25], indicating the presence of certain common recognition and engulfment mechanisms for dying cells. On the other hand, the distinct cellular features observed during macrophage engulfment of necrotic mammalian cells imply that unique pathways exist to clear necrotic and apoptotic cells [26].
Phosphatidylserine (PS), a membrane phospholipid, is a known “eat me” signal presented on the surface of apoptotic cells to directly or indirectly attract phagocytic receptors such as C. elegans CED-1, Drosophila Draper, and mammalian Tim4 and BAI1, leading to the initiation of their engulfment [27–31]. In living cells, PS is almost exclusively localized to the inner leaflet of the plasma membrane, at least partially due to an ATP-dependent aminophospholipid translocase activity that selectively returns PS and PE (phosphatidylethanolamine) from the outer to the inner leaflet [32–34]. During the early stage of apoptosis, PS is detected on the outer leaflet, suggesting a process of trans-bilayer redistribution [32,33]. Phospholipid scramblases, by catalyzing the random, bi-directional “flip-flop” of phospholipids across the membrane bilayer, could potentially counter the aminophospholipid translocase activity [35]. The mouse transmembrane protein 16F (TMEM16F) was recently found to act as a novel Ca2+-activated phospholipid scramblase [36]. However, TMEM16F does not seem to be involved in exposing PS on apoptotic cell surfaces [37]. On the other hand, mouse Xk-related protein 8 and CED-8, its C. elegans homolog, mediate PS exposure in response to apoptotic stimuli [38,39]. These results suggest that different phospholipid scramblases function in different cell types and in response to different stimuli. In addition, the mammalian ATP-binding-cassette transporter A1 (ABCA1) has been implicated in the translocation of PS from the inner to the outer leaflet [40,41], although evidence to the contrary also exists [42].
Previously, using milk-fat-globule EGF8 (MFG-E8::GFP), a GFP-tagged, secreted PS reporter, we have detected the presentation of PS specifically on the surface of apoptotic cells during animal development [43]. We have further identified two alternative mechanisms that promote PS exposure in apoptotic somatic and germ cells, respectively [43]. The PS exposure on apoptotic cell surface during embryonic development, which is necessary for their engulfment, relies on the function of C. elegans CED-7, a homolog of mammalian ABCA1 transporters [43].
Considering the insights that have been made to understand how apoptotic cells are recognized and removed, the mechanisms by which necrotic cells are engulfed remain poorly defined. In particular, it is unclear whether necrotic cells are capable of the active presentation of “eat me” signaling molecules such as PS to attract engulfing cells. Rather, it was assumed that PS was detected on necrotic cell surfaces due to the rupture of necrotic cell membranes [44]. The work reported here establishes that necrotic C. elegans touch neurons actively present PS on their outer surfaces while maintaining plasma membrane integrity. It further defines two mechanisms that act in parallel to promote the exposure of PS on necrotic cell surfaces, one that is shared with apoptotic somatic cells, and another that is unique to necrotic touch neurons.
The dominant mutant allele e1611dm of C. elegans mec-4 results in the necrotic death of six touch neurons, which swell to several times of their original diameter (Fig 1A and 1B(a)), displaying a morphology distinct from somatic apoptotic cells, which undergo cytoplasmic shrinkage and nuclear condensation [19,21,45–47]. Previously, necrotic AVM and PVM were observed inside the hypodermis [16]. To visualize the engulfment status of all necrotic touch neurons, we expressed Pced-1 ced-1C::gfp, which produced a GFP reporter that is distributed evenly in the cytoplasm of engulfing cell types, including hypodermal cells [48]. We detected all six necrotic touch neurons as dark holes embedded inside the GFP-labeled engulfing cells in newly hatched mec-4(e1611dm) larva (Fig 1B), establishing that necrotic touch neurons are engulfed by hypodermal cells.
To examine whether necrotic cells expose PS on their cell surfaces, we expressed mfg-e8::gfp, a secreted PS reporter [43], in mec-4(e1611dm) mutant worms. In newly hatched L1 larvae, we detected GFP specifically enriched on the surface of necrotic touch neurons (Fig 1C). In contrast, when mfg-e8::gfp was expressed in mec-4(+) worms, no fluorescence was detected on the surface of living touch neurons (Fig 1D). These results suggest that PS is present specifically on the surface of cells that undergo necrosis.
Previously, it was generally believed that necrosis caused prominent plasma membrane rupture [1,2]. If that is the case, it is possible that MFG-E8::GFP molecules penetrate through the plasma membrane and associate with the PS molecules on the inner leaflet of the plasma membrane. To distinguish whether the enriched MFG-E8 signal is a result of PS exposure on the outer or inner surfaces of necrotic cells, we examined whether the necrotic touch neurons observed in the mec-4(e1611dm) mutants lost plasma membrane integrity. We observed the localization of GFP or mRFP reporters expressed specifically in touch neurons under the control of the mec-7 promoter (Pmec-7) [49] in mec-4(e1611dm) worms and found that the fluorescent signals were exclusively retained inside necrotic cells (S1A (a, b, e, f) Fig). In parallel, secreted GFP (ssGFP) reporters, which are tagged with a signal sequence from SEL-1 [50] and expressed specifically from hypodermal (Pcol-10) and body wall muscle (Pmyo-3) cells [28,51], two types of cells that neighbor the touch neurons, were not observed inside touch neurons (S1A (c, d, g, h) Fig). These GFP signals were detected inside coelomocytes, mobile cells that possess high endocytic activity (S1B Fig), indicating that they are indeed secreted into the close proximity of the touch neurons. No plasma membrane penetration of the touch-neuron reporter or neighboring-cell reporters was observed during a 36-hr observation period from the appearance of necrotic cell morphology in embryos to the mid-L4 stage. The above lines of evidence indicate that the plasma membrane of necrotic touch cells is not permeable to GFP or mRFP molecules (which are of sizes between 25 and 27 kD). Thus, it is unlikely that the same plasma membrane would be permeable to MFG-E8::GFP, which is substantially larger (78 kD). Furthermore, when wild-type and mec-4(e1611dm) worms were stained with propidium iodide, a small molecular weight (MW = 688 Da) fluorescent dye that is not permeable across the intact plasma membrane (Materials and Methods), we did not observe propidium iodide signal in the living or necrotic cells (S1C Fig). The only propidium iodide signal observed came from the intestinal track, inside which were ingested propidium iodide-stained bacteria cells (S1C Fig). Together, the above results indicate that, against the common belief that necrotic cells passively expose PS through plasma membrane rupture and in this manner attract engulfing cells, the C. elegans necrotic touch neurons maintain cell integrity and actively expose PS, which may function as a specific “eat me” signal, on their surfaces.
Using a live-cell recording protocol that we established for touch neurons (Materials and Methods), we monitored the dynamics of an MFG-E8::mCherry reporter during embryogenesis. Among the six touch neurons, four are born during mid embryogenesis, including PLML and PLMR, which were reported to arise at approximately 510 min post the 1st embryonic cell division (the 1st-cleavage), whereas AVM and PVM were reported to be born at the L1 larval stage [46,47]. At hatching, PLML and PLMR should have existed for 290 min. We observed that the enrichment of PS on necrotic PLML and PLMR was a gradual process after necrosis was initiated at a morphological level (Fig 1E and 1F) (S1 Movie). Among the following three events, (1) the differentiation of touch neurons, which is indicated by the expression of Pmec-7 gfp, (2) the swelling of touch neurons undergoing necrosis, which is visible under DIC optics, and (3) the exposure of PS on the outer surface, indicated by the enrichment of MFG-E8::mCherry, cell differentiation occurs the earliest, initiating approximately 480–525 min after the 1st cleavage (Fig 1E(b, j) and 1F) (S1 Movie). On average 120 min later, the swelling morphology of PLML/R starts to develop, indicating that the constitutively active Na+/Ca2+ channel starts to initiate necrosis. The first time point when the enrichment of PS is detected on PLML/R surface varies; yet in all 19 cases monitored, it occurs after the initial appearance of the necrotic morphology observed by DIC (S1 Movie). Subsequently, the PLML/R surface MFG-E8::mCherry intensity continues to increase until the time point of hatching (Fig 1E(p and q)). Our observations established the order of the initiation of these three events, and further suggest a causal relationship between the initiation of necrosis and the exposure of PS.
To determine whether PS-externalization is a general phenomenon occurring to different types of neurons that undergo necrosis, we monitored PS enrichment on non-touch neurons. u662, a gain-of-function mutation in deg-3, which encodes a subunit of an acetylcholine receptor ion channel [20], causes the necrosis of the six touch neurons and a few additional sensory and inter-neurons through hyper-activation of the acetylcholine receptor ion channel [20]. Cells undergoing necrosis in mec-4 and deg-3 dominant mutants display the same distinct morphology [19,20,52,53]. In deg-3(u662) animals, we detected PS on the surface of necrotic neurons, including touch neurons and other types of neurons (S2 Fig). This result suggests that PS-exposure is a general feature of neurons induced to undergo necrosis through excitotoxicity.
CED-1 is a phagocytic receptor that is localized on the surface of several types of cells, including all engulfing cell types and clusters around apoptotic cells in response to the neighboring “eat me” signals [28,43]. To determine the efficiency of necrotic cell clearance in the ced-1(e1735); mec-4(e1611dm) double mutants, we chose to score the dynamic presence of necrotic PLML and PLMR, two touch neurons in the tail that undergo necrosis during mid-embryogenesis, throughout all larval stages. The rationale is that the longer a necrotic touch cell persists during larval development, the less efficient its removal process must be. The same scheme was used to score the efficiency of necrotic cell clearance throughout this report (Materials and Methods). We found the ced-1(e1735) null mutation greatly reduced the efficiency of necrotic cell removal (Fig 2A), consistent with a previous report [25]. Prior results indicate that the extracellular domain of CED-1 is responsible for recognizing the surface feature(s) of apoptotic cells. We analyzed two truncated forms of CED-1 for their ability to recognize necrotic cells by monitoring GFP-tagged truncated forms expressed in larval hypodermal cells, which engulf necrotic touch neurons. We first found that CED-1ΔC::GFP, a truncated form of CED-1 (Fig 2B) that remains bound to the plasma membrane of the engulfing cells, is highly enriched on the phagocytic cup or phagosomal surface (Fig 2C and 2E) in comparison to other regions of the plasma membrane of the same cell. Quantification of GFP fluorescence intensity on phagocytic cups or phagosomes is on average 3.1 times of that on other plasma membrane regions of the same cell (Fig 2E(d)). We next found that CED-1Ex::GFP, a truncated and secreted form of CED-1 (Fig 2B), was specifically enriched on the surfaces of necrotic cells (Fig 2D). These results indicate that the extracellular domain of CED-1 directly recognizes necrotic cells, and that the high level of CED-1ΔC::GFP detected on the engulfing cell membrane region around necrotic cells is not merely a result of necrotic touch neurons being embedded inside hypodermal cells.
Since necrotic cells, like apoptotic cells, specifically expose PS on their surfaces, and because CED-1 recognizes both necrotic and apoptotic cells in vivo, we tested whether PS could act as a ligand for CED-1. The extracellular domain of CED-1 was expressed as a fusion protein to glutathione S-transferase (GST) (CED-1-GST) in an insect cell expression system, affinity purified by glutathione-sepharose chromatography (Fig 3A), and tested for its binding affinities to PS in vitro (Materials and Methods). We first employed an Enzyme-Linked ImmunoSorbent Assay (ELISA)-like reaction to test the interaction between CED-1-GST and PS or phosphatidylcholine (PC). The CED-1 protein displayed efficient association with PS but not PC, in a dose-dependent manner (Fig 3B). We next examined this association in a surface plasmon resonance assay [29]. In this assay, CED-1-GST was applied onto channels of the HPA chip, on which equivalent numbers of PS-containing liposomes and PC-only liposomes had been immobilized (S3 Fig). We found that the values in response units after the injection of CED-1-GST to the channel with PS-containing liposome were higher than those obtained for the control channel (S3B Fig). These two assays indicate that the extracellular domain of CED-1 interacts directly with PS. Furthermore, free PS-containing liposomes, but not PC-liposomes, could efficiently compete with PS-containing liposomes coated on the well for binding to CED-1-GST in the ELISA-like reactions (Fig 3C), suggesting that CED-1 specifically recognizes PS as a component of membrane bilayer. In the ELISA-like assay, binding between CED-1-GST with two other phospholipids, phosphatidylethanolamine (PE) and phosphatidylinositol (PI), was also detected (Fig 3D). All these results indicate that the extracellular domain of CED-1 directly associates with phospholipids containing an amino group or negative charge, including PS, and support the hypothesis that PS serves as a ligand for CED-1 for the recognition of necrotic and apoptotic cells.
Among the seven engulfment ced genes that act in two parallel pathways, ced-7, which encodes a member of the ABC transporter family (Fig 4A), is the only one required for the presentation of PS on the outer surface of apoptotic cells in C. elegans embryos [43]. The ced-7(n1996) null mutation severely impairs the removal of necrotic touch neurons (Fig 4C). We found that among the null or strong loss-of-function mutations of six ced genes, only the ced-7(n1996) mutation significantly reduced the percentage of necrotic touch neurons exhibiting surface PS (Fig 4D). These results indicate that CED-7 is essential for PS exposure on necrotic touch neurons. In support of this conclusion, a CED-7::GFP reporter expressed in touch neurons (Pmec-7 ced-7::gfp) is observed on the surface of necrotic touch neurons in mec-4(e1611dm) background (Fig 4B), consistent with a role of CED-7 in PS-flipping from one leaflet of the plasma membrane to the other.
CED-7 is broadly expressed in all cells [54]. To determine whether ced-7 functions to promote necrotic cell removal in the engulfing cells or in touch neurons, we tested the effect of cell type-specific expression of ced-7 in the rescue of ced-7 mutant phenotypes. The ced-7(n1996) null mutation severely delays necrotic cell removal, resulting in the persistent existence of necrotic touch neurons in more than 60% of L4 larvae (Fig 4C). Expression of ced-7 in either touch neurons (Pmec-7 ced-7) or neighboring engulfing cells (Pced-1 ced-7) each partially rescued the necrotic-cell removal defect (Fig 4C), indicating that the functions of CED-7 in necrotic and engulfing cells both contribute to the efficient removal of necrotic cells. We further found that the specific expression of ced-7 in touch neurons but not in neighboring engulfing cells primarily rescued the PS exposure defect of ced-7(n1996) mutants (Fig 4E). This result clearly indicates that the touch cell-specific role of CED-7 is responsible for promoting PS exposure.
Mammalian TMEM16F, a multispan transmembrane domain protein, is a Ca2+-activated phospholipid scramblase that triggers PS exposure in response to Ca2+-influx [36]. Given that the mec-4(dm)-induced touch neuron necrosis is mediated by Ca2+ influx [22,24], we examined whether ANOH-1, a close C. elegans homolog of TMEM16F (Figs 5B and S4), mediated PS exposure when necrosis occurred. We analyzed an anoh-1(tm4762) deletion allele (www.wormbase.org) for the removal of dying cells. The tm4762 allele carries a 202-bp deletion that results in a frameshift and a premature stop codon after amino acid 17 of the predicted ANOH-1b open reading frame and removes the start codon of the alternatively-spliced ANOH-1a open reading frame (S4 and S5B(b) Figs) (also see the next section), presumably generating a null allele. In anoh-1(tm4762) mutant embryos, the numbers of apoptotic cells are the same as that displayed in wild-type embryos at 5 different embryonic developmental stages (Fig 5C). Furthermore, the dynamics of the engulfment and degradation processes of three individual apoptotic cells, C1, C2, and C3, are normal comparing to wild-type embryos (Fig 5E), using previously established protocols [48,55]. Similarly, the number of apoptotic germ cells are virtually the same in wild-type and anoh-1(tm4762) mutant adult hermaphrodites at four time points (Fig 5D). These results indicate that anoh-1, unlike ced-7, is not involved in the removal of apoptotic cells. In contrast, in anoh-1(tm4762); mec-4(e1611dm) double mutant animals, the removal of necrotic touch neurons is siginificantly delayed: at L1 and L2 stages, the mean numbers of persistent necrotic PLML/R in anoh-1(tm4762) background are approximately 1.6-fold and 1.5-fold of that in wild-type larvae, respectively, whereas at L3 and L4 stages, the mean numbers are not significantly different from wild-type animals (Fig 5F). These results strongly suggest that anoh-1 specifically contributes to efficient removal of necrotic but not apoptotic cells.
In wild-type and anoh-1 mutant animals, we further measured the MFG-E8::GFP signal intensity and calculated the ratio of GFP intensity on necrotic PLML/R to that in an equivalent area of the neighboring live cells (Materials and Methods). Lack of PS enrichment on the surface of necrotic cells will result in a ratio of approximately 1.0. We first quantified the PS signal intensity at the early L1 stage, when the necrotic cell removal defect displayed by the anoh-1 mutants was the most prominent among all four larval stages (Fig 5F). The anoh-1(tm4762) mutation significantly inhibits PS enrichment on necrotic touch neurons, reducing the median value of this ratio from 1.34 in the wild-type animals to 1.13 (Fig 6B and 6C). This result indicates a unique function of ANOH-1 in promoting PS exposure on necrotic cell surfaces. At the young L2 larval stage (15–16 hrs after hatching), the average relative PS signal intensity value increases from the young L1 value in both the wild-type (from 1.36 to 1.71) and anoh-1(tm4762) mutant (1.20 to 1.51) strains (S6B Fig), probably as a result of the continuous accumulation of the GFP signal on the surface of necrotic neurons,. The PS signal increase in anoh-1 mutants could explain the reduced removal defect at later developmental stages (Fig 6A).
Based on its gene structure, anoh-1 is predicted to encode two splice variants, anoh-1a and anoh-1b (Fig 5A). The predicted ANOH-1b protein carries an additional 18 residues at the amino-terminus, as compared to the predicted ANOH-1a protein (S4 Fig). Using RT-PCR (Materials and Methods), we detected the existence of the anoh-1b transcript in whole worm extracts (S5B Fig). To determine which of the two splice variants and in which cell type anoh-1 is functional in the removal of necrotic cells, we individually expressed anoh-1a and anoh-1b in touch neurons (Pmec-7) or engulfing cells (Pced-1), in the anoh-1(tm4762); mec-4(e1611dm) background (Fig 5F and 5G). To monitor the subcellular localization of each isoform, anoh-1a and anoh-1b were each tagged with gfp. Among the four expression constructs tested, only anoh-1b, when expressed in touch neurons (Pmec-7 anoh-1b::gfp), rescued the necrotic cell removal defect of these mutants (Fig 5F and 5G). In addition, Pmec-7 anoh-1b::gfp also leads to the recovery of PS exposure on necrotic neuron surfaces (S7 Fig). These results indicate that anoh-1b is the functional form and that it acts in necrotic touch neurons to promote their removal.
To understand the function of the amino-terminal 18 amino acids present in ANOH-1b, which are absent from the predicted ANOH-1a protein, we characterized the subcellular localization of each form, as N- or C-terminal GFP-tagged proteins, expressed in touch neurons. GFP::ANOH-1b and ANOH-1b::GFP are both localized to the plasma membrane, consistent with the hypothesized role of ANOH-1 in promoting PS on cell surface (Fig 5H(a, b, e, and f)). In contrast, GFP::ANOH-1a and ANOH-1a::GFP are both observed inside touch neurons, enriched on the nuclear surface (Fig 5H(c, d, g, and h)). These results suggest that the plasma membrane localization is essential for the function of ANOH-1 in necrotic cell-removal.
To determine the expression pattern of anoh-1b, we cloned the 617 bp DNA fragment immediately 5’ to Exon 1 of anoh-1b, and tentatively regard this fragment as the Panoh-1b promoter. The nuclear localization sequence (NLS)-tagged GFP signal produced by Panoh-1b is expressed in touch neurons (Figs 5I and S5C(l, o, r)). This result supports the touch neuron-specific function of ANOH-1 (the b isoform) in promoting PS-exposure. In addition, by comparing the expression patterns of Panoh-1b NLS-GFP and a pan-neuronal expression reporter Prab-3 dsRed [56], we observed strong Panoh-1b activity in many neurons in the head and tail (S5C Fig). This result is consistent with a previous report detecting anoh-1 expression in sensory neurons [57]. In addition, anoh-1 expression is also observed in intestinal cells as previously reported [57] and pharyngeal muscles (S5C (g, h, i) Fig).
To investigate the functional relationship between CED-7 and ANOH-1 in the clearance of necrotic cells, we first compared the numbers of persistent necrotic cells throughout larval developmental in anoh-1(tm4762) ced-7(n1996) double mutants and in anoh-1 or ced-7 single mutant animals, in the mec-4(e1611dm) mutant background. We observed that inactivating ced-7 causes a stronger necrotic cell removal defect than that of anoh-1 (Fig 6A). Furthermore, inactivating both ced-7 and anoh-1 results in an enhanced necrotic cell removal defect starting at the L2 stage (Fig 6A), suggesting that CED-7 and ANOH-1 might perform partially parallel functions during removal. We further quantified the PS signal intensity at the early L1 stage. Inactivating both ced-7 and anoh-1 further reduces the PS signal on the surface of necrotic PLML/R, which is already significantly reduced by the ced-7 or anoh-1 single mutations compared to the wild-type background (Fig 6B and 6C) or the ced-6 mutation, which delays necrotic cell removal but does not affect PS exposure (S8 Fig). The above results suggest that CED-7 and ANOH-1 both contribute to the PS-externalization activity; furthermore, they may do so through partially parallel pathways. At the L2 stage, when the necrotic corpse-removal and PS-exposure phenotypes of anoh-1 mutants are greatly reduced, the PS-exposure phenotype of the anoh-1 ced-7 double mutants was similar to that of ced-7 single mutants (S6 Fig), again suggesting that inactivating anoh-1 delays but does not block necrotic-cell removal.
We examined whether a loss-of-function mutation of ced-8, which is proposed to encode a phospholipid scramblase essential for PS-exposure on the surface of apoptotic cells, also affects the removal of necrotic touch neurons. In a ced-8(n1891) mutant allele, a strong loss-of-function allele that carries a nonsense mutation, truncating CED-8 (458aa) after aa 219 [58], there is a significant necrotic touch neuron removal defect at the L1 larval stage (Fig 7A). However, this defect was not observed in any of the later larval stages (Fig 7A), suggesting that CED-8 merely delayed but did not block necrotic cell removal. We further analyzed the functional relationship between anoh-1 and ced-8. At the L1 stage, the anoh-1(tm4762); ced-8(n1891) double mutants display a more severe necrotic cell-removal defect as compared to each single mutant (Fig 7B). These results suggest that anoh-1 and ced-8 act in parallel and thus the necrotic cell removal function of ced-8 is independent of anoh-1.
The presentation and function of PS on the surface of cells undergoing necrosis have previously been overlooked, primarily because of the long-existing notion that these cells encounter injury and lose plasma membrane integrity [44,59,60]. However, in recent years, accumulating evidence has demonstrated that in addition to cell injury, necrosis could also be induced by genetic programs (reviewed in [9]). In short, multiple molecular mechanisms exist that induce and execute necrosis [9,14]; thus all necrotic cells do not necessarily lose plasma membrane integrity. The loss of membrane integrity of necrotic cells in culture, where there are no phagocytes to engulf them, does not necessarily represent what happens inside animal bodies, where engulfing cells target dying cells at early stages of their death [16,61] (our own observations). On the other hand, those necrosis events that occur inside animal bodies were rarely examined for plasma membrane integrity. In the case of neuronal excitotoxic necrosis, electron microscopic studies of rat brains and C. elegans touch neurons reported the swelling of necrotic cells and the presence of surrounding phagocytes, yet no loss of plasma membrane integrity [16,62]. By monitoring signals elicited from GFP or mRFP reporters either inside or outside necrotic cells and by incubating worms with propidium iodide, a small molecule dye that is not plasma membrane permeable, we observed that the necrotic C. elegans touch neurons induced by excitotoxicity maintain plasma membrane integrity throughout embryonic and larval development. These results indicate that the common notion that necrotic cells lose plasma membrane integrity is not necessarily true for all kinds of necrotic cells at all developmental stages, and further suggest that for PS to be present on the surface of necrotic touch neurons, an active PS exposure mechanism must exist. Further genetic studies reported here revealed that at least two separate PS-exposure activities act to promote PS exposure on the surface of necrotic cells.
PS is an evolutionarily conserved “eat me” signal exposed by apoptotic cells in metazoan organisms ranging from simple to complex and it is implicated in recruiting engulfing cells [27,29,43,63–65]. Here we report the active exposure of PS on the surface of necrotic touch neurons. The fact that both necrotic and apoptotic cells expose PS on their surfaces implies the existence of a conserved dying cell-recognition mechanism. Indeed, we further discovered the novel function of phagocytic receptor CED-1 in recognizing necrotic cells in addition to apoptotic cells. The extracellular domain of CED-1 (CED-1Ex) alone, without the transmembrane or intracellular domains, is sufficient for associating to the surface of necrotic cells, suggesting an extracellular ligand-receptor interaction. In support of this hypothesis, we have detected direct and selective in vitro interaction between CED-1Ex and acidic phospholipids including PS. On the other hand, CED-1Ex does not display affinity to phosphatidylcholine (PC), a neutral phospholipid without a net charge and is most abundant on the outer surface of the plasma membrane [32]. Consistent with this conclusion, Draper, the Drosophila orthologue of CED-1, also directly associates with PS exposed to the surface of apoptotic cells [29]. Together, the in vivo and in vitro observations indicate that the direct interaction between the CED-1 family phagocytic receptors and PS is an important mechanism that brings together phagocytes and their target cells, regardless of whether these cells die of apoptosis or necrosis.
As an “eat me” signal, PS is also known to attract phagocytic receptors via an indirect mechanism. Secreted bridging molecules such as mouse MFG-E8 bring dying and engulfing cells together by interacting simultaneously with both PS and phagocytic receptors [66]. C. elegans TTR-52, a transthyretin-like secreted protein, has been implicated as a bridging molecule that links PS on apoptotic cells to CED-1 on engulfing cell surfaces [67]. Our finding, together with that reported by Wang et al (2010) [67], indicate that direct and indirect interactions between PS and CED-1 provide two distinct molecular mechanisms to support the recognition of dying cells by CED-1.
The phosphatidylserine receptor (PSR) protein family was originally identified as PS receptors that promote apoptotic cell-removal yet was later reported to function in other developmental processes and possess several biochemical activities that are in conflict with the proposed role as PS receptors [68–71]. Recently, it was reported that C. elegans PSR-1 displayed an in vitro PS-binding affinity [72]. In vivo, psr-1 mutants display weak defects in apoptotic- and necrotic-cell removal [72,73]. PSR-1 might thus contribute to the recognition of dying cells in addition to CED-1.
In C. elegans, CED-7, a member of the ABC transporter family, was known to regulate PS exposure by apoptotic cells [43]. Mouse ABCA1 was also reported to participate in PS redistribution during apoptotic cell clearance [40,74]. Our discovery of CED-7 as a key factor in promoting the externalization of PS by necrotic touch neurons further demonstrates that the presentation of the “eat me” signal shares conserved mechanism(s) during different types of cell death. We further discovered that CED-7 has two distinct functions, one in necrotic and the other in engulfing cells. How CED-7 acts in necrotic cells to promote PS exposure remains to be elucidated. CED-7 is ubiquitously expressed. There thus must be dying cell-specific mechanisms that activate CED-7. Whether the CED-7 activation mechanisms are common or distinct in necrotic and apoptotic cells remains unknown. Moreover, the engulfing cell-specific function of CED-7 is a mystery and requires further investigation. Previous research suggests that engulfing cells might also externalize PS and that ABC transporters might be involved in this event [40,75]. The function of this event remains to be clarified.
Our observations indicate that ANOH-1, the C. elegans homolog of mammalian TMEM16F, functions in necrotic neurons to promote PS exposure. ANOH-1 is primarily expressed in neurons, including touch neurons [57] (Figs 5I and S5C). These lines of evidence indicate a cell type-specific function of ANOH-1 to facilitate necrotic-cell removal.
The vertebrates TMEM16 family of proteins, also known as anoctamins, are divided into two subfamilies based on two distinct Ca2+-dependent biochemical activities: Cl- channels and lipid scramblases [37,76]. In addition, TMEM16F possesses both biochemical activities [36,77]. Mammalian TMEM16F promotes cellular PS exposure in response to Ca2+ ionophore yet not to apoptotic stimuli [37]. The Ca2+-activated phospholipid scramblase activity of the TMEM16 subfamily provides an important clue towards revealing a necrosis-specific PS-exposure mechanism (Fig 7C). As an evolutionarily conserved feature, Ca2+ influx is known to be an effective trigger of the excitotoxic death of mammalian neurons [78]. For example, the activation of the NMDA receptor upon binding to excessive glutamate elicits an initial rise of cytoplasmic calcium that induces a subsequent calcium-dependent calcium release from the ER [12,79,80]. Elevation of cytoplasmic Ca2+ is also a critical trigger for excitotoxic necrosis of neurons in C. elegans, including that of touch neurons and other types of neurons [24,81]. Particularly in touch neurons, the dominant mutation in MEC-4, a subunit of a multimeric, mechanically gated DEG/ENaC channel, leads to an increased influx of Ca2+, resulting in necrosis [19,22] (Fig 7C). We propose that in touch neurons that undergo Ca2+-activated necrosis, Ca2+ further acts as an activating factor for the PS-exposure activity of ANOH-1 (Fig 7C). This Ca2+-dependent PS-exposure mechanism might apply to multiple kinds of necrotic neurons including but not limited to mechanosensory neurons. Moreover, the possibility remains that CED-7 or CED-8 might also be activated by Ca2+ in necrotic neurons (Fig 7C). As the disruption of Ca2+ homeostasis is closely associated with neuron degeneration conditions [82], the work reported here has a broader application in understanding the physiological role of the clearance of many kinds of degenerative neurons resulted from pathological conditions or aging.
Our finding that the anoh-1 ced-7 double mutants display more severe defects in PS-exposure and necrotic cell-removal than each single mutant alone suggests that ANOH-1 and CED-7 together provide the necessary activities for efficient PS-exposure on necrotic touch neurons. One possible model is that they act in two independent and partially redundant pathways. The common function of CED-7 in both necrotic and apoptotic cells and the necrotic cell-specific function of ANOH-1 in facilitating PS-exposure have established that cells die of different mechanisms employ both common and unique molecular activities to present a common “eat me” signal. Given that a necrotic C. elegans neuron possesses a surface area many times of that of an apoptotic cell, the cooperation of multiple molecular activities such as those represented by CED-7 and ANOH-1, are likely essential for the efficient and timely exposure of PS on the cell surface at a level high enough to attract engulfing cells (Fig 7C). We further found that CED-8, a homolog of the mammalian phospholipid scramblase Xk8 [38,39], also made a modest contribution to the removal of necrotic cells. ced-8 and anoh-1 act in two independent pathways to promote PS exposure. Currently, it is unknown whether CED-8 facilitates PS-exposure to the surface of necrotic cells or whether CED-8 acts in necrotic cells; moreover, the functional relationship between ced-7 and ced-8 is unknown. CED-8 might represent a third pathway that is in parallel to both the CED-7 and the ANOH-1 pathways (Fig 7C).
C. elegans was grown at 20°C as previously described [83] unless indicated otherwise. The N2 Bristol strain was used as the wild-type strain. Mutations are described in [84] and by the Wormbase (http://www.wormbase.org) unless noted otherwise: LGI, ced-1(e1735), ced-12(n3261); LGII, enIs46[Pmec-7 ced-7 and punc-119(+)]; LGIII, ced-7 (n1996), ced-6 (n2095), anoh-1(tm4762), unc-119(ed3); LGIV, ced-5(n1812), ced-10(n1993); LGV, unc-76(e911), deg-3(u662); LG X, ced-8(n1891), mec-4(e1611dm). The tm4762 allele was generated and provided by the National Bioresource Project of Japan and was outcrossed twice prior to analysis. The precise location of nucleotide deletion has been confirmed by allele-sequencing. Integrated transgenic arrays used are as follows: LGII, ttTi5605[mos] [85]; LGV, enIs33[Pdyn-1 mfg-e8::gfp and punc-76(+)] [43].
Extrachromosomal arrays were generated by microinjection [86] of plasmids with coinjection marker punc-76(+) [87] into strains carrying the unc-76(e911) mutant. Transgenic animals were isolated as non-Unc animals.
We obtained a single-copy insertion of Pmec-7 ced-7 in chromosome II in the ttTi5605 locus using the MosSCI insertion method [85], in strain EG4322 (ttTi5605; unc-119(ed3)) [86]. The transgenic array also carries the C. briggsae unc-119(+) genomic DNA that rescues the unc-119(ed3) phenotype. The single-copy insertion of the transgenic array into anticipated locus was confirmed by single-worm PCR analysis.
The Pmec-7 mrfp (pZL08) and Pmec-7 mCherry constructs were generated by replacing GFP in Pmec-7 gfp (pPD117.01, a gift from Andrew Fire) with mrfp [88] or mCherry [89]. Pcol-10 ssGFP is a secreted GFP reporter expressed by hypodermal cells under the control of Pcol-10, the promoter for col-10. It is generated by replacing the myo-3 promoter (Pmyo-3) in the Pmyo-3 ssGFP reporter [90] with Pcol-10, a gift from V. Ambros [28]. Pced-1 ced-7 was constructed by placing the 5.1kb CED-7 cDNA [54] behind the Pced-1 promoter [28]. Pmec-7 ced-7 and Pmec-7 ced-7::gfp were constructed by replacing Pced-1 from Pced-1 ced-7 with the Sph-1-ClaI fragment of Pmec-7 from pPD117.01, respectively. The Pmec-7 ced-7/unc-119(+) construct for single-copy MosSCI insertion was generated by cloning the Pmec-7 ced-7 fragment into the BssHII and SpeI sites of plasmid CFJ151 [85].
The anoh-1a cDNA was amplified from total RNA from mixed-stage C. elegans population through RT-PCR and cloned into pPD117.01 to generate Pmec-7 gfp::anoh-1a and Pmec-7 anoh-1a::gfp. The anoh-1b genomic-cDNA chimeric fragment was constructed by ligating exon 1 and intron 1 of anoh-1b genomic DNA with anoh-1a cDNA, and similarly cloned into pPD117.01 to generate Pmec-7 gfp::anoh-1b and Pmec-7 anoh-1b::gfp. Pced-1 anoh-1a::gfp and Pced-1 anoh-1b::gfp were constructed by replacing Pmec-7, respectively, with Pced-1 from pZZ829 [91]. The 617bp 5’ UTR of anoh-1b together with the first 297bp of exon 1 of the anoh-1b isoform was PCR-amplified from N2 worm extracts and cloned into pPD95.69 (a gift from Andy Fire), a promoter-less vector carrying a gfp cDNA tagged with a SV40 nuclear localization signal (NLS::gfp), between SphI and XmaI sites to generate Panoh-1b NLS::gfp, which allowed us to identify the cells in which anoh-1b was expressed.
Total RNA was isolated from mixed-stage C. elegans population using Trizol extraction with column purification (Qiagen, Inc.). cDNA was synthesized using the iScript cDNA Synthesis Kit (BIO-RAD, Inc.).
The cDNA encoding the extracellular region of CED-1 tagged with GST at its C-terminus (CED-1-GST) was expressed in insect Sf9 cells using Bac-to-Bac Baculovirus Expression System, a baculovirus-based vector system, (Life Technologies Japan, Tokyo, Japan), and the resulting protein was affinity-purified by glutathione-Sepharose chromatography (GE Healthcare Japan, Tokyo, Japan), essentially as described previously [92]. An ELISA-like solid-phase binding reaction was conducted virtually according to the published procedure [93]. In brief, varying amounts of CED-1-GST or GST, the latter serving as a negative control, were added to the wells of a 96-well culture container surface-coated with phospholipids and incubated for 1 h at room temperature. The wells were washed, successively incubated with anti-GST antibody (Millipore, Inc.) and anti-mouse IgG antibody conjugated with horseradish peroxidase, and then subjected to a colorimetric reaction with o-phenylenediamine followed by the measurement of OD490. An assay for surface plasmon resonance was done with Biacore 3000 (GE Healthcare Japan) using the HPA chip pre-bound by liposomes, as described previously [29]. Liposomes were prepared using PC alone (PC-only liposome) or a mixture of PC and PS at a molar ratio of 7:3 (PS-containing liposome) as described previously [94]. Phospholipids were purchased from Avanti Polar Lipids (Alabaster, USA).
According to the previous reports, among the six touch neurons, ALML and ALMR are born at ~450 min post-1st cleavage, PLML and PLMR are born at ~510 min post 1st-cleavage, whereas AVM and PVM are born at ~9 hrs after hatching, at the L1 larval stage [46,47]. To determine the efficiency of necrotic cell clearance during all four larval stages, we chose to score the presence of necrotic PLML and PLMR, two touch neurons in the tail. L1, L2, L3, and L4 larvae were staged as larvae collected within 1 hr, 15–16 hrs, 24–25 hrs, and 33–34 hrs after embryos hatching, respectively. The total number of necrotic touch neurons in the tail of 10 worms was scored, and the mean of three repeats was calculated. The number of apoptotic cells were scored in embryos of different stages, in the head of young L1 larvae hatched within 1 hr, and in the gonad of adult hermaphrodites 48 hrs after the mid-L4 stage as described [48].
Olympus IX70-Applied Precision DeltaVision microscope equipped with Photometris Coolsnap digital camera and Applied Precision Softworx 5.0 software was used to acquire serial Z-stacks of fluorescence images at 0.5 μm intervals and to deconvolve these images of embryos and larvae [48]. To quantify the MFG-E8::GFP signal intensity on the surface of necrotic cells (In), following deconvolution of z-stack images using the Applied Precision Softworx software, the necrotic cell surface was outlined by two closed polygons and the signal intensity in the area of the bigger polygon was subtracted with that of the smaller polygon. To normalize the signal intensity, the same two polygons were placed in the area neighboring the necrotic cell and the background fluorescence intensity (Ib) was measured using the formula similar to that applied to necrotic cell surface. The relative signal intensity (Ir) of MFG-E8::GFP enriched on the necrotic cell surface is represented as In/Ib. For each data point, at least 40 necrotic cells were quantified.
To monitor the dynamics of PS presentation during the necrosis of touch neurons in embryos via time-lapse recording, embryos were mounted on an agar pad on a glass slide in M9 solution [48]. The starting point of recording was at 460 min-post the 1st cleavage (the 1st embryonic cell division), when an embryo reached the 2-fold stage. Recording was performed in 5-min interval until the embryo hatched. At each time point, a Z-stack of images composed of 40 serial Z sections at 0.5 μm/section were captured. Since embryos continue to move inside the eggshell, PLML and PLMR were followed by monitoring both the touch neuron reporter Pmec-7 GFP and the distinct swelling morphology of necrotic cells.
For propidium iodide staining, mixed-stage worms were washed off plate using Hanks’ balanced salt solution buffer (HBSS buffer; with calcium and magnesium, Fisher Scientific) containing 10 μM propidium iodide and incubated for 2 hrs [95]. Worms were subsequently washed three times using HBSS buffer and mounted on an agar pad on a glass slide in 30 mM sodium azide for microscopic observation. Olympus IX70-Applied Precision DeltaVision microscope was used to acquire serial Z-stacks at 0.5 μm interval. Excitation and emission wavelengths used are ~540 and ~590 nm, respectively.
|
10.1371/journal.pcbi.1000511 | Estimating the Continuous-Time Dynamics of Energy and Fat Metabolism in Mice | The mouse has become the most popular organism for investigating molecular mechanisms of body weight regulation. But understanding the physiological context by which a molecule exerts its effect on body weight requires knowledge of energy intake, energy expenditure, and fuel selection. Furthermore, measurements of these variables made at an isolated time point cannot explain why body weight has its present value since body weight is determined by the past history of energy and macronutrient imbalance. While food intake and body weight changes can be frequently measured over several weeks (the relevant time scale for mice), correspondingly frequent measurements of energy expenditure and fuel selection are not currently feasible. To address this issue, we developed a mathematical method based on the law of energy conservation that uses the measured time course of body weight and food intake to estimate the underlying continuous-time dynamics of energy output and net fat oxidation. We applied our methodology to male C57BL/6 mice consuming various ad libitum diets during weight gain and loss over several weeks and present the first continuous-time estimates of energy output and net fat oxidation rates underlying the observed body composition changes. We show that transient energy and fat imbalances in the first several days following a diet switch can account for a significant fraction of the total body weight change. We also discovered a time-invariant curve relating body fat and fat-free masses in male C57BL/6 mice, and the shape of this curve determines how diet, fuel selection, and body composition are interrelated.
| The unrelenting obesity epidemic has resulted in intensive basic scientific investigation into the molecular mechanisms of body weight regulation—with the mouse being the organism of choice for such studies. We know that any mechanism of body weight regulation must exert its effect by influencing food intake, energy output, fuel selection, or some combination of these factors over extended time scales (∼weeks for mice). While food intake and body weight can be frequently measured in mice, current methods prohibit corresponding measurements of energy output or fuel selection on such long time scales. We address this deficiency by developing a mathematical method that quantitatively relates measurements of food intake, body weight and body fat to calculate the dynamic changes of energy output and net fat oxidation rates during the development of obesity and weight loss in male C57BL/6 mice. The mathematical model is based on the law of energy conservation, makes very few assumptions, and provides the first continuous-time estimates of energy output and fuel selection over periods lasting many weeks. Application of our methodology to various mouse models of obesity will improve our understanding of body weight regulation by placing molecular mechanisms in their whole-body physiological context.
| Mouse models of obesity have become critically important research tools for discovering molecular mechanisms of body weight regulation. But understanding these mechanisms in the context of whole-body physiology requires knowledge of food intake, energy output, and fuel selection [1]. Furthermore, measurements made at an isolated time point cannot explain why body weight has its present value since body weight is determined by the past history of energy and macronutrient imbalance [2]. While food intake and body weight changes can be measured frequently over several weeks (the relevant time scale for mice), correspondingly frequent measurements of energy output and fuel selection are not currently feasible.
Expensive indirect calorimetry systems can be used to measure energy expenditure and respiratory exchange over periods of a few days and most systems require removing mice from their normal environment which can alter their behavior [3]. Alternatively, the doubly labeled water method can give an estimate of average energy expenditure, but this method requires specialized equipment for sample analysis as well as prior knowledge of fuel selection as measured by the respiratory quotient (RQ) [4]. Furthermore, significant quantities of blood need to be collected which could impact the behavior of the mouse and makes repeat measurements untenable [4].
Here, we present a mathematical method that quantitatively relates food intake, body weight and body fat to calculate the dynamic changes of energy output and net fat oxidation rates during the development of obesity and weight loss in male C57BL/6 mice. The mathematical model is based on the law of energy conservation, makes very few assumptions, and provides the first continuous-time estimates of energy output and fuel selection over periods lasting many weeks. Our methodology also revealed the relationship between diet, fuel selection, and body composition change in male C57BL/6 mice by identifying a time-invariant curve relating body fat and fat-free masses.
As previously described [5], male C57BL/6 mice were given ad libitum access to standard chow (C), high fat diet (HF), or high fat diet plus liquid Ensure (EN) for 19 weeks, while some mice were fed the high fat or the high fat plus Ensure for 7 weeks before being switched back to chow for the remaining 12 weeks (HF-C and EN-C, respectively). Figure 1A shows the body weight changes of the various groups during the development of obesity on the HF and EN diets as well as the weight loss and persistent obesity of the HF-C and EN-C groups following a switch back to the chow diet at 7 weeks (error bars have been omitted for clarity). A single curve was able to describe the adjusted fat-free mass as a function of body fat mass for all groups at all time points (Figure 1B) and is analogous to the curve discovered by Forbes describing human body composition change [6]. Our mathematical model used this fitted curve along with the body weight data to compute the body fat mass changes (Figure 1C). Without adjusting any parameters, the model also accurately predicted the fat mass changes measured in a separate experiment with high-fat feeding of C57BL/6 mice followed by a switch to chow after 20 weeks (Figure 1D).
Our model calculated the first continuous-time estimates of the energy output dynamics underlying the observed body weight changes (Figure 2A). The 95% confidence interval surrounding the calculated energy output rates resulted primarily from variability of the measured energy intake rate (individual data points are depicted along with the average black curve used for each group) but also included the effect of body composition variability (Figure 1B). The HF and HF-C groups had a transient decrease of energy output at the onset of high fat feeding at 0 days. In contrast, the EN and EN-C groups did not show a significant transient reduction of energy output at the onset of the high energy diet. Energy output gradually increased with weight gain in all of the groups. Following the return to the chow diet, the HF-C group had a transient increase of energy output which was not seen in the EN-C group. Note that these transient changes account for significant fractions of the overall energy imbalances and would be difficult to detect using indirect calorimetry or doubly labeled water methods.
Net fat oxidation rates increased sharply at the onset of high fat feeding in the HF and HF-C groups, but did not rise sufficiently to match the increase of fat intake (Figure 2B). Interestingly, despite similar increases of fat intake in the EN and EN-C groups compared with the HF and HF-C groups, the initial increase of net fat oxidation was significantly attenuated. Net fat oxidation gradually increased in all the groups as body weight increased. Following the switch to chow, there was a transient increase of net fat oxidation in both HF-C and EN-C groups before falling to match the low level of fat intake after a few weeks.
A useful measure of fuel selection is the respiratory quotient, RQ, where a value of 0.7 reflects a state of pure fat oxidation whereas a value of 1.0 reflects a state of pure carbohydrate oxidation and intermediate values represent a fuel selection mixture (see Methods). The estimated 24 hour RQ (Figure 3) demonstrates the impact of both diet and body composition on fuel selection. The HF group had an immediate decrease of RQ due to the diet followed by a slow progressive decrease as body fat gradually increased. The EN group showed little initial change of RQ which then progressively decreased to an intermediate value. After switching to the chow diet, the HF-C group had a rapid increase of RQ towards that of the C group whereas the EN-C group had a transient decrease of RQ before increasing towards the C group.
The mouse has become the most popular organism for investigating molecular mechanisms of body weight regulation. But understanding the physiological context by which a molecule exerts its effect on body weight requires knowledge of energy intake, energy expenditure, and fuel selection. Our simple mathematical method calculates the dynamics of energy output and fuel selection over extended time periods using longitudinal measurements of body weight, food intake, and body composition. We showed that our method can detect both transient changes of energy expenditure and net fat oxidation rates as well as longer timescale changes found with weight gain and loss. Similar methodology has been previously developed by our group to relate human body-composition changes with dynamic adaptations of fuel selection in both adults [7] and infants [8]. The method is especially well-suited for mouse studies because food intake can be accurately measured over the extended time periods required to measure significant changes of body weight and body fat. While we have applied the model to data averaged within groups of mice, it would be also interesting to examine individual mouse trajectories as a way of investigating inter-individual variability.
Our equations extract information about energy output that is already present in the body weight and food intake data. Other than the law of energy conservation, the only assumption was that the relationship between changes of body fat and fat-free mass were described by a well-defined function in accordance with the Forbes theory of body composition change [6]. This assumption was confirmed in the present study for mature male C57BL/6 mice (Figure 1B) and we hypothesize that genetic manipulations can alter the shape of this function. However, once the function has been determined we showed that it provided accurate estimates of body fat changes in an independent feeding experiment using body weight measurements alone (Figure 1D). Therefore, knowledge of the Forbes curve for a given mouse model eliminates the need for frequent body composition measurements.
To estimate the net fat oxidation rate and RQ, an additional assumption regarding carbohydrate balance was required (see Methods). We found that the Forbes function (Figure 1B) determined the relationship between food intake, body composition change, and net fat oxidation rate [7]. While both humans and mice have Forbes functions that increase with body fat mass, the concavity of the curves is opposite [6]. Therefore, great caution should be exercised when extrapolating fuel selection results in mice to predict human responses. The physiological reason for this difference is presently unclear. Our research group is actively engaged in developing detailed models of the complex interactions between carbohydrate, fat, and protein metabolism in humans [9] to better understand the relationship between the physiological drivers of fuel selection and the Forbes body composition curve. We plan to develop similar models in mice to help understand these relationships and the differences between the species.
In contrast to our method, currently available techniques for estimating energy expenditure are expensive, involve a plethora of assumptions, and can impact the behavior of the mice [3],[4]. These factors make it common to find reports of energy expenditure rates that are quantitatively inconsistent with the measured energy intake and body weight changes found in mice that were not subjected to these procedures. As an illustrative example, consider the recent publication by Funato et al. where the energy intake rate of the wild type mice was at least 17 kcal/d and the energy expenditure measured by indirect calorimetry was less than 5 kcal/hr/(kg BW)0.75. This translates to an absolute expenditure rate of less than 10.7 kcal/d for a mouse that was at most 40 grams at the time of measurement [10]. Such a large positive energy balance would translate to a rate of weight change of at least 4.7 g/week (if all excess energy was deposited as fat) versus the measured weight gain which was less than 1 g/week. The purpose of this example is not to criticize the work of Funato et al., but rather to highlight how even careful indirect calorimetry and food intake measurements can lead to estimates of energy imbalance that are inconsistent with the weight gain measurements.
Our own attempt to use indirect calorimetry to validate the model predictions of energy expenditure and fuel selection highlighted two important issues. First, the mice that were consuming the high energy diets lost significant amounts of weight when moved to the indirect calorimetry cages indicating that their behavior was not representative of the mice not subjected to the procedure. Second, the measured energy expenditure rates were unrealistically high compared to the model predictions for all groups of mice. In fact, the measured energy expenditure rate was higher than the measured energy intake in the chow-fed mice that did not lose weight (an impossibility) and greatly exceeded the expenditure required to explain the weight loss in the mice fed the high energy diets. These discrepancies led us to diagnose a technical problem with the indirect calorimetry equipment. Thus, we were unable to validate the model estimates of energy expenditure and fuel selection.
The field of farm animal nutrition has a long and rich history of using mathematical modeling to analyze animal growth and identify nutritional factors that potentially limit growth rate [11]–[14]. The simplest models describe the efficiencies of various diets in their ability to deposit body energy, often specified in terms of body fat and protein [11],[13],[14]. Inputs to such models include energy intake, body weight, and the rates of body fat and protein deposition. The model outputs include the efficiencies of protein and fat deposition as well as the so-called maintenance energy requirement which is roughly defined as the energy intake required when the animal is not growing. An alternative representation uses energy intake, body weight, total energy expenditure (by calorimetry methods), and protein deposition rate (via nitrogen balance) as model inputs and predicts the maintenance energy requirement, fat deposition rate, and body protein and fat deposition efficiencies.
At the next level of complexity, animal growth models prescribe an energy partitioning rule that specifies how body protein will accumulate for a given food intake rate as a function of body weight, age, or body protein. Energy partitioning rules are often complex [12],[13], but can be thought of as similar to the Forbes function that specifies how energy imbalances are partitioned between body fat and fat-free mass. A significant difference is that our approach is applied to mature mice whose overall growth rate was minimal despite their ability to gain and lose fat-free mass in response to the various diets.
Once the partitioning rule is specified, the outputs of animal growth models include body fat mass, maintenance energy requirement, as well as body fat and protein deposition efficiencies given the food intake and body weight as model inputs. In contrast, our model outputs are body fat mass, fuel selection, and total energy expenditure which are more relevant for mouse obesity studies and avoids the known problem of arbitrarily distributing total energy expenditure between tissue deposition costs versus maintenance energy requirements [11], [14]–[16]. Animal growth models have often used power-law functions of body weight to model the maintenance energy requirements that were previously calculated using the above methods. Once specified, the model of maintenance energy requirements can be used along with the calculated efficiencies of protein and fat deposition and the energy partitioning rule to predict body weight and body fat change as a function of the food intake [11],[14]. We are presently developing a model of total energy expenditure in mice that will allow prediction of body weight and composition changes as well as fuel selection when food intake is the only input to the model.
A weakness of our methodology is that it does not distinguish the various components of energy output including resting metabolic rate, thermic effect of feeding, adaptive thermogenesis, physical activity, or any changes of energy excreted in urine and feces that are unaccounted for by the estimates of diet metabolizability. Furthermore, the method does not operate on a within-day time scale and therefore cannot address changes between day versus night or transitions between fed and fasted states. Indirect calorimetry is required to address these issues and would provide important information for the interpretation of our calculated longer-term estimates of energy output and fuel selection. We believe that the combination of our continuous-time methodology with indirect calorimetry measurements at judiciously chosen time points can be applied to various mouse models of obesity as a powerful tool for characterizing the metabolic dynamics underlying experimentally observed body weight changes.
We certify that all applicable institutional and governmental regulations concerning the ethical use of animals were followed during this research. All procedures were approved by the National Institute of Diabetes and Digestive and Kidney Diseases Animal Care and Use Committee.
Full details of the experiment were previously described [5]. Briefly, forty seven 3 month old male C57BL/6 mice weighing 25.9±1.2 g (The Jackson Laboratory, Maine) were housed individually and randomly assigned to five weight-matched groups: 1) C group (N = 12) continued on the chow diet; 2) HF group (N = 12) on a high fat diet (F3282; Bio-Serv Inc., NJ; 5.45 kcal/g with 14% energy derived from protein, 59% from fat, and 27% from carbohydrate); 3) EN group (N = 11) on the high fat diet plus liquid Ensure (Abbott Laboratories, Kent, UK), which had an energy density of 1.06 kcal/ml with 14% of energy derived from protein, 22% from fat, and 64% from carbohydrate; 4) HF-C group (N = 6) switched from high fat to chow after 7 weeks; 5) EN-C group (N = 6) switched from high fat plus Ensure to chow after 7 weeks. All animals received free access to water and food throughout the study. The high fat diet was provided using Rodent CAFÉTM feeders (OYC International, Inc., MA), and liquid Ensure was provided in a 30-ml bottle with a rodent sip tube (Unifab Co., MI) and liquid intake was measured every day. Solid food intake was corrected for any visible spillage and was measured every day for the high fat diet and every other day for the chow diet using a balance with a precision of 0.01 g (Ohaus model SP402). Body composition was measured once per week using 1H NMR spectroscopy (EchoMRI 3-in-1, Echo Medical Systems LTD, Houston, TX) after body weight was determined.
We begin with the law of energy conservation, also known as the energy balance equation:(1)where F is the body fat mass, FFM is the fat-free mass defined as the measured body weight, W, minus the fat mass, and and are the energy densities for changes in fat and fat-free masses, respectively [17]. IT is the total metabolizable energy intake rate corrected for spillage, and E is the energy output rate. We distinguish the energy output rate from the energy expenditure rate since we did not measure any changes of energy excreted in urine or feces. In other words, if the metabolizable energy content of each diet is constant then our calculation of the energy output is equivalent to energy expenditure.
Analogous to the Forbes theory of human body composition change [6], we hypothesized that there is a well-defined, time-invariant function, α, that describes the relationship between changes of FFM and F in male C57BL/6 mice:(2)Once the function α is specified, equation (1) can be solved for the energy output rate as a function of the measured energy intake rate and the rate of body weight change as follows:(3)The fat mass is given by solving the following differential equation:(4)Alternatively, if the Forbes assumption does not apply for a given mouse model (for example, during periods of significant growth), a curve could be directly fit to the measured fat mass time series data and used in place of equation 4. While this procedure would give equivalent results, it necessitates frequent body composition measurements for every experiment.
Note that very few assumptions were made in the development of our equations to estimate energy output. All of the above equations were derived from the law of energy conservation (1) and the only assumption was that there exits a well-defined Forbes relationship, α, relating changes of body fat and fat-free masses – an assumption that was directly confirmed by comparison to measured body composition data.
Since we are also interested in fuel selection, we must consider the fates of dietary macronutrients including their oxidation rates, storage in the body, as well as major inter-conversion fluxes where carbohydrate can be converted to fat (i.e., de novo lipogenesis) and amino acids can be converted to the carbohydrate glucose (i.e., gluconeogenesis). The following macronutrient balance equations represent these changes:(5)where P is body protein, G is glycogen, GNG is the gluconeogenic rate, DNL is the de novo lipogenic rate, and IF, IP and IC are the intake rates of dietary fat, protein and carbohydrate, respectively. The oxidation rates of fat, protein, and carbohydrate (FatOx, ProtOx, and CarbOx, respectively) sum to the total energy output, E.
To simplify the macronutrient balance equations, we note that glycogen stores are small, especially when compared with daily carbohydrate intake rates. For example, humans have a glycogen pool size of about 500 g which is equivalent to the typical amount of carbohydrate consumed over ∼2 days and equilibrates on a time scale of ∼1 day [9],[18]. The equilibration time is likely even more rapid in mice since they typically consume carbohydrate at a rate of ∼2 g/d and their glycogen stores are probably less than 0.6 g (assuming maximal glycogen pool sizes of 8% of liver weight and 0.6% of muscle weight as observed in rats [19] and assuming that mouse liver is less than 5 g and muscle is less than 30 g [5]). Thus, over the time-scale of interest the system is in a state of average carbohydrate balance:(6)Therefore,(7)If we define the net fat oxidation rate as follows:(8)then the equation for body protein change becomes:(9)Finally, we assume that FFM is proportional to body protein such that(10)Therefore, we have a simple a two-compartment macronutrient partitioning model which we have previously shown has an invariant manifold as its attractor [20]:(11)From equations 4 and 11, the net fat oxidation rate can be written as a function of the measured fat intake rate and the rate of body weight change:(12)Note that the carbohydrate balance assumption was only required to calculate the estimate of net fat oxidation, but was not required to calculate the energy output rate.
The shape of the Forbes curve has direct implications for how fat oxidation rate is related to changes of body fat. This can be seen by calculating the partial derivative of the net fat oxidation rate with respect to body fat:(13)Interestingly, this quantity has opposite sign in humans versus mice. Thus, great care must be taken when fuel selection measurements in mice are extrapolated to humans.
The respiratory quotient, RQ, is the carbon dioxide production rate divided by the oxygen consumption rate and was approximated by:(14)This approximation assumes a negligible contribution of de novo lipogenesis and gluconeogenesis which is reasonable since these fluxes act to offset each other with respect to CO2 production. Since the carbohydrate oxidation rate is approximately equal to the carbohydrate intake rate on long time scales, the calculated RQ may have slight inaccuracies during rapid transitions immediately after diet switches, but will be reasonably accurate thereafter.
To apply our mathematical model to data from our mouse experiment, food intake measurements were averaged over each diet period and we assumed stepwise transitions immediately after each diet switch followed by a smooth approach to the average intake of the final diet period. These curves are depicted as solid black lines in Figure 2 and represent the average of the individual intakes shown by the data points. Body weight measurements for the C, HF, and EN groups of mice were fit using third order polynomial functions of time, as depicted by the solid curves in Figure 1A. Following the diet switch in the HF-C and EN-C groups, the body weight curves were fit to exponential functions. The rates of change of body weight were then calculated by computing derivatives of the fitted curves. Other than their ability to adequately describe the model input data, the precise mathematical form of these curves is not important.
The Forbes body composition function, α, was fit to an exponential function of the body fat mass as shown in Figure 1B. Specifically, we assumed that the individual data points for fat-free mass versus body fat for each group of mice were described by the following equation:(15)The Forbes function, α, is then given by:(16)Since the intercept parameter, b, does not influence the Forbes function, we adjusted the FFM data for each group by subtracting the difference between the calculated intercept parameter for each group and its average value across groups. We then simultaneously fit the adjusted FFM data from all groups to arrive at our final Forbes function used for all of the groups.
The parameter values for the Forbes body composition function were determined via a Markov Chain Monte Carlo (MCMC) method [21] implemented in MATLAB (version R2008a; MathWorks Inc, Natick, MA). To approximate the posterior distribution of the parameters in the Forbes function (equation 16), we drew 100,000 MCMC samples of parameter values, of which the first 30000 were discarded as burn-in period; afterwards one fifth of the rounds were retained. Parameter sets were drawn from a proposal density that were normally distributed and centered on the previous value. The variance of the proposal density was tuned for an average acceptance rate of ∼0.25 during the burn-in period. The convergence of the chain was assessed both by visual inspection of the trace plots for all the parameters and through the Geweke test [22]. At each sampling, the probability of accepting the new parameter set given current parameter set was where r is the Metropolis ratio [21]. The posterior distribution of energy output (equation 3) was calculated from the joint distribution of the parameters in the function and the energy intake in each group of the animals assuming no correlation existed between the two distributions. The energy intake in each group of animals was normally distributed with a standard error of 0.39, 0.39, 0.41, 0.55, and 0.55 Kcal/d for the C, HF, EN, F-C, and EN-C groups, respectively. The 95% confidence intervals of the predicted energy output were obtained by calculating the 2.5th and 97.5th percentiles of the posterior distribution of energy output.
|
10.1371/journal.pntd.0000800 | The Onchocerca volvulus Cysteine Proteinase Inhibitor, Ov-CPI-2, Is a Target of Protective Antibody Response That Increases with Age | Despite considerable efforts, a suitable vaccine against Onchocerca volvulus infection has remained elusive. Herein, we report on the use of molecular tools to identify and characterize O. volvulus antigens that are possibly associated with the development of concomitant immunity in onchocerciasis.
Third-stage larvae (L3) and molting L3 (mL3) O. volvulus stage-specific cDNA libraries were screened with a pool of sera from chronically infected patients who had likely developed such immunity. The 87 immunoreactive clones isolated were grouped into 20 distinct proteins of which 12 had already been cloned and/or characterized before and 4 had been proven to be protective in a small O. volvulus animal model. One of these, onchocystatin (Ov-CPI-2), a previously characterized O. volvulus cysteine proteinase inhibitor was, overall, the most abundant clone recognized by the immune sera in both the L3 and mL3 cDNA libraries. To further characterize its association with protective immunity, we measured the IgG subclass and IgE class specific responses to the antigen in putatively immune (PI) and infected (INF) individuals living in a hyperendemic area in Cameroon. It appeared that both groups had similar IgG3 and IgE responses to the antigen, but the INF had significantly higher IgG1 and IgG4 responses than the PI individuals (p<0.05). In the INF group, the IgG3 levels increased significantly with the age of the infected individuals (r = 0.241; p<0.01). The IgG1 responses in the INF were high regardless of age. Notably, culturing L3 in vitro in the presence of anti-Ov-CPI-2 monospecific human antibodies and naïve neutrophils resulted in almost complete inhibition of molting of L3 to L4 and to cytotoxicity to the larvae.
These results add to the knowledge of protective immunity in onchocerciasis and support the possible involvement of anti-Ov-CPI-2 IgG1 and/or IgG3 cytophilic antibodies in the development of protective immunity in the PI and the INF. The results further support the consideration of Ov-CPI-2 as a leading target for an anti-L3 vaccine.
| Onchocerciasis is a chronic and highly debilitating disease of humans caused by a worm called Onchocerca volvulus. This worm can live in the human body for over 15 years. The disease affects mainly the skin and eyes and is the second leading infectious cause of blindness worldwide. There is currently no vaccine to prevent the infection. Available drugs can give short-term relief but cannot cure the infection. To prevent infection, a vaccine against the third-stage infective larva, L3, or the developing larva is required. These stages were shown to be the targets of protective immunity that develops in individuals who live in onchocerciasis endemic regions. One type of protective immunity has been shown to develop with age and is called concomitant immunity. In the present study, we have identified a number of larval antigens that may be associated with the development of such immunity. The most prominent of these antigens was Ov-CPI-2, also called onchocystatin, which had previously been shown to be a promising vaccine candidate. This antigen was further characterized and confirmed to be possibly also a target of immune protection that develops in the infected individuals with age and is referred to as concomitant immunity.
| Human onchocerciasis (river blindness) is a highly debilitating and blinding disease caused by the nematode, Onchocerca volvulus. Over 37 million people in endemic countries of tropical Africa, Latin America and the Arabian Peninsula are infected [1]. Currently, there is no cure for the disease and a vaccine is yet to be developed. Importantly, protective immunity against O. volvulus larvae has now been definitively demonstrated in humans, cattle and mice, thereby proving the conceptual underpinnings that a vaccine can be produced against this infection [2]. Both clinical and epidemiological data support also the concept that acquired immunity against O. volvulus infection occurs in infected humans and that this immunity increases with age. In chronically infected (INF) individuals in highly endemic settings, the skin microfilariae (mf) density tends to increase with age until the ages of 20 to 40 in most studies, suggesting that older individuals do develop a means of limiting further infections [3]. A similar trend was also observed for adult worms where in areas of high transmission, the number of palpable nodules reached an average of 3–5 over time and then leveled-off [4]. While it is conceivable that products from the existing infections may as well prevent super infections from getting established, the conglomeration of at times a few adult worms in one nodule is rather at odds with this perception. The leveling off of patent infections with age is consistent with the concept of concomitant immunity [5], [6], which is characterized by the tendency of the host to eliminate the newly introduced infective-stage larvae (L3), while the established adult worms and microfilariae are left almost unaffected.
Further evidence for the development of concomitant immunity in filarial infections comes from studies of lymphatic filariasis, where it was associated with parasite stage-specific immune responses. Levels of antibodies against Wuchereria bancrofti infective stage larvae (L3) were found to increase with duration of exposure [7], and there were differences in the class and subclass antibody responses to adult versus larval antigens of Brugia malayi [8]. More recently, the concept of concomitant immunity was further verified experimentally using the Acanthocheilonema vitae jird model [9]. The L3 and the developing or molting L3 (mL3) have been the focus of vaccination studies in filarial infections [10]–[15].
The roles of the different antibody isotype and subclass responses to O. volvulus crude extract of larval proteins and defined protective recombinant proteins in onchocerciasis have been studied in greater detail. Importantly, analysis of the immunoglobulin class and subclass responses to some of the protective antigens revealed that these proteins induced mainly cytophilic antibodies, IgG1, IgG3 and/or IgE in PI and INF. Boyer and colleagues [16] showed that in areas of high transmission, high IgG3 responses are associated with protective immunity, a finding which was later confirmed by other studies [6], [17]. In our previous studies we have shown that IgG3 and IgE responses to crude extracts of L3 as well as the IgG1 and IgG3 responses to a highly abundant larvae-specific protein, Ov-ALT-1, were also associated with concomitant immunity [6]. The filarial Ov-ALT-1 family members have been shown to be protective against infection with L3 of O. volvulus and in lymphatic filariasis in small animal models [13], [18]–[20]. Although the precise mechanisms that mediate killing of O. volvulus L3 in the PI and in individuals who developed protective immunity with age are still unknown, data from many studies support the view that antibodies are part of the effector mechanisms against incoming O. volvulus larval infection and could, together with the Th1 and/or Th2 cytokines produced, induce efficient anti-L3 antibody-dependent cell mediated cytotoxicity (ADCC) reactions [6], [16], [21]. However, no studies were performed to specifically identify and clone larval proteins that are possibly associated with the development of concomitant immunity. In order to identify such proteins, we have screened L3 and mL3 O. volvulus stage-specific cDNA libraries with a pool of sera from chronically infected patients who have likely developed concomitant immunity. One of the antigens identified, Ov-CPI-2 (onchocystatin), a previously characterized O. volvulus cysteine proteinase inhibitor was the most abundant clone recognized by the immune sera, and was selected for further analysis in two groups of individuals: the INF and the putatively immune (PI) individuals [22]. Additionally, we present results on in vitro cytotoxic effects of human neutrophils on third-stage larvae of Onchocerca volvulus in the presence of mono-specific antibodies to Ov-CPI-2.
The protocols used in this study were approved by the New York Blood Center's IRB and by an NIH accredited Institutional Review Board of the Medical Research Council Kumba, Cameroon. Each participant provided informed consent by signing or thumb printing a consent form after reading it or after the content of the form was read and/or explained to them. The serum samples were collected from residents of 5 villages around Kumba, Cameroon, a hyperendemic area for onchocerciasis [6]. These villages were Marumba I, Marumba II, Boa Bakundu, Bombanda, and Bombele. All participants were born or had resided for more than 10 years in the villages. The standard skin snip test for detection of microfilariae (mf) was performed on each subject and clinical symptoms of onchocerciasis were recorded. The averages of the mf counts of four skin snips taken from each individual were used in estimating the individual skin mf densities. None of the subjects had received ivermectin treatment prior to the collection of blood. In this study we used 176 serum samples collected from infected individuals; 115 males and 61 females, ranging in age from 3 to 75 years. In addition, serum samples of 21 putatively immune (PI) subjects were studied. These individuals had no signs or history of onchocerciasis and were parasitologically negative during at least a two-year follow-up survey employing the standard skin snip test and a polymerase chain reaction (PCR) assay [6], [22].
In order to identify O. volvulus L3 and mL3 larval proteins that are recognized by sera from infected individuals who have likely developed concomitant immunity, λ-Unizap XR cDNA expression libraries of these two life cycle stages (kindly obtained from Dr. Steven Williams via the NIAID/NIH Filariasis Research Reagent Repository Center) were immunoscreened as previously described [12] with only slight modifications. A pool of sera was prepared using equal volumes of 17 serum samples from infected individuals, 35 years of age and above, who had previously been shown to have high anti-L3 and mL3 antibody titres [6]. The serum pool was first pre-cleared with Escherichia coli lysate (SIGMA, St Louis, MO) and then used at a final dilution of 1∶400 for the immunoscreening of the cDNA libraries, which were plated at 25,000 plaque-forming units (pfu) per 150 mm diameter plates. The secondary antibody, goat anti-human IgG conjugated to horseradish peroxidase (KPL, Gaithersburg, MD) was used at 1∶8000 dilution. The positive plaques were identified by incubating the processed plaque lifts in a solution of phosphate buffered saline (PBS) containing 0.67 mg/ml diaminobenzidine tetrahydrochloride (DAB) (SIGMA, St Louis, MO) and 1 µl/ml of 30% H2O2. All positive plaques were recovered and re-purified using one or two additional screening cycles. Single phage clones were eventually recovered and stored in standard phage buffer (SM buffer) in the presence of 0.3% chloroform.
The inserts of the positive phages were amplified using PCR with the M13 forward and M13 reverse primers. Briefly, 45 µl of PCR supermix (Invitrogen, Carlsbad, USA), 1 µl of each primer (100 ng/µl) and 1–10 µl of the eluted recombinant phage were mixed and incubated at 94°C for 5 minutes. Amplification was carried out for 35 cycles of: denaturation at 94°C for 45 seconds, annealing at 55°C for 45 seconds, and extension at 72°C for 75 seconds. The final amplification cycle included an additional extension step at 72°C for 10 minutes (Eppendorf MasterClycler Gradient, Eppendorf, USA). The PCR amplicons were analyzed by standard agarose gel electrophoresis on a 1.2% gel. Only clones that produced a single band on electrophoresis were selected for sequencing. Those with two or more bands were further plaque purified before sequencing. The PCR products were purified using the Rapid PCR Purification System kit (Marligen Biosciences, Ijamsville, MD) and quantified. Then 10–20 ng of each PCR product was sequenced using the M13 reverse universal sequencing primer (GENEWIZ DNA Sequencing Service, New Jersey, NJ). Gene specific primers were prepared to complete the sequencing reactions in both directions.
BLASTx and BLASTn searches were carried out using the cDNA sequences in frame with the EcoR1 cloning site at the 5′ terminus against the GenBank non-redundant protein and nucleotide databases, respectively, as well as against the Nembase2 and Wormbase databases. Open reading frames were computed using the DNASIS software package (Hitachi, CA) or the ORFing facility at the NCBI homepage (www.ncbi.nlm.nih.gov).
Recombinant Ov-CPI-2 fused to Schistosoma japonicum glutathione S-transferase (GST) polypeptide was expressed in the pGEX-1N vector and purified as previously described [13]. The IgG1, IgG3, IgG4 and IgE responses to recombinant GST-Ov-CPI-2 were determined by indirect ELISA essentially as described by MacDonald et al [6]. Briefly, GST-Ov-CPI-2 or GST were diluted in 0.05M carbonate-bicarbonate buffer, pH 9.6 at a concentration 1 µg/ml and used to coat the wells of Immunolon 2 plates (Dynex, VA), overnight at 4°C. After blocking excess binding sites with 3% (w/v) casein, individual serum samples which had previously been pre-cleared with E. coli lysate were diluted to 1∶200 and reacted with the bound antigen for 90 minutes at room temperature. For the analysis of IgE levels, plates were coated with 10 µg/ml of rOv-CPI-2 or the control (GST) and serum samples were pre-absorbed with protein G-Sepharose (Pharmacia) before being used at a 1∶20 dilution. For IgG subclass responses, the bound antibodies were detected by using a 1∶1000 dilution of monoclonal antibodies against different human IgG subclasses (Hybridoma Reagent Laboratory, Kingsville, Md.). This step was followed by incubation with a 1∶8000 dilution of horseradish peroxidase-conjugated goat anti-mouse IgG (H+L) (KPL, Gaithersburg, MD) for another 60 minutes at room temperature. IgE antibodies in the sera were detected by using a horseradish peroxidase-conjugated, ε-chain-specific anti-human IgE monoclonal antibody (Zymed, San Francisco, CA) at a dilution of 1∶750. Tetramethylbenzidine (SIGMA, St Louis, MO) was prepared according to the supplier's recommendations and used as the enzyme substrate. Absorbance (OD) was read after stopping the reaction with an equal volume of 1M H2SO4 at 450 nm using an Emax ELISA reader (Molecular Devices, CA). The OD values for rOv-CPI-2 are presented as the net values after subtracting the OD values of the GST control from those of the corresponding GST-Ov-CPI-2 readings for each serum sample.
Monospecific human antibodies to recombinant Ov-CPI-2 were purified as described by Lustigman et al. [23]. Briefly, cells from induced lysogenic cultures of λgt11 expressing Ov-CPI-2 were resuspended in 0.1 M phosphate buffer containing 10 mM dithiothreitol (DTT) and sonicated. Sodium Dodecyl Sulphate (SDS) was added to a final concentration of 1% and then centrifuged at 10000× g for 10 min at 4°C. The supernatant was desalted on a G-25 column using 0.1 M phosphate buffer, 1 mM DTT, 0.1% SDS, pH 6.8. The supernatants containing β-galactosidase or β-galactosidase Ov-CPI-2 fusion polypetide were then coupled to CNBr-Sepharose 4B using the protocol provided by the manufacturer (Pharmacia). Antibodies from 100 ml of plasma of an infected individual, who had high titers of anti-Ov-CPI-2 antibodies, were affinity purified on the immobilized polypeptides. The mono-specificity of the eluted antibodies was confirmed by Western blot analysis and immunogold electron microcopy [23]. Negative antibodies were purified from the same donor sera from an affinity column containing crude extracts of induced lysogenic culture of λgt11 coupled to the CNBr beads. The IgG titer of the anti-Ov-CPI-2 purified antibodies was 1∶3125 as determined by ELISA, while the negative antibodies had an anti-Ov-CPI-2 cross reacting antibody response at 1∶5 dilution. The purified antibodies were passed through a 0.22 µM filter for sterilization.
The L3 killing or inhibition of molting assays were done as described by Johnson et al. [21]. Briefly, L3 were washed in RPMI-1640 with 1× GPS (glutamine, penicillin, streptomycin) and the worms were diluted to 5 worms per 50 µL in the medium containing 20% fetal calf serum (complete medium). Worms were distributed to 10 wells of a 96-well plate per treatment group and 2×105 normal neutrophils isolated by dextran sedimentation were added to each well in 50 µL of complete medium. Then, 100 µL of the anti-Ov-CPI-2 purified antibodies or the negative control antibodies were added to each of the 10 wells. Complete medium without antibodies was also included as a control. The 96-well plates were then incubated at 37°C in a 5% CO2 incubator until day 6 when the cultures were observed for molting; the presence or absence of the fourth-stage larvae (L4) and the empty cast of the L3 under an inverted microscope. Viability of the larvae was then determined by MTT (3-(4,5 dimethylthiazol-2yl)-2,5 diphenyl tetrazolium bromide) staining as previously described [21]. The experiments were repeated thrice on separate days and the results presented are the averages and the ranges of the three experiments.
The Mann-Whitney test was used to compare the median of IgG subclass and IgE class responses of the INF and PI groups to the antigen. Spearman's rank correlation test was used to test the significance of the correlation between the antibody responses and the age of the patients (expressed as the correlation coefficient, r). Unpaired t test with Welch's correction was used to compare mean ± SEM of % L3 molting and % viability in the presence or absence of anti-Ov-CPI-2 antibodies and neutrophils. A p value of <0.05 was considered significant.
A total of 62 L3 and 25 mL3 immunoreactive cDNA clones out of a total of 50,000 and 25,000 PFUs, respectively, were isolated after screening the libraries with a pool of sera from older (>35 years) individuals who were likely to have developed concomitant immunity. The 87 isolated and sequenced clones were grouped into 20 distinctive proteins: 7 in L3, 10 in mL3 and 3 that were present in both L3 and mL3 (Tables 1 and 2). Interestingly, 12 of these larval proteins had already been cloned and characterized using other screening sera and other or similar cDNA libraries. Many of them (Ov-CPI-2, Ov-RAL-2, Ov-FBA-1, Ov-103 and Ov-73k) had been shown previously to be potentially associated with anti-O. volvulus protective immunity [14], [24], or to be potential antigens for the serological diagnosis (Ov-16, OV1CF and Ov-33) of onchocerciasis (Tables 1 and 2). Moreover, 4 of these (Ov-CPI-2 [13], Ov-FBA-1 [25], Ov-103 [26], [27] and Ov-RAL-2 [14], [15], [2] have been proven to be protective in O. volvulus mouse model. One of these, onchocystatin (Ov-CPI-2), a previously characterized cysteine proteinase inhibitor of O. volvulus was the most abundant and second most abundant clone recognized by the immune sera in the L3 (59.7%) and mL3 (16%) cDNA libraries respectively (Tables 1 and 2). The 4 L3s and 4 mL3s novel proteins identified accounted for a total of 10.5% and 16% of the isolated L3 and mL3 clones, respectively. The majority of these O. volvulus novel proteins have homologues in Brugia malayi. One of the L3 novel proteins is a putative new member of the fatty acid retinoid binding protein family (Accession number GQ202199). This clone, GQ202199, although had only 25% identity to the fatty acid and retinol binding protein-1 of O. volvulus (ACL98477.1; Ov-FAR-1), it had 60–62% identity to a B. malayi putative FAR-1 protein [XP001900470. 1]. We have therefore named this protein Ov-FAR-2. Interestingly, Ov-FAR-1 (also known as Ov-RBP-1 or Ov20) was shown to be also protective in the O. volvulus mouse model [14]. Only one encoded protein (Accession number GQ202201) had no homology to the annotated B. malayi genome but it had a low level of homology (E value of 3.8; E not significant!) with a hypothetical C. elegans protein, B0432.14 (NP_001033326.1).
Since the recombinant Ov-CPI-2 (rOv-CPI-2) protein was identified by immunoscreening with sera from infected individuals 35 years of age or more who are likely to have developed concomitant immunity, it was of interest to investigate the development of the IgG1, IgG3 and IgE antibody responses to the antigen in infected individuals in relation to their age (N = 176 for IgG1 and IgG3 analyses; N = 68 for IgE analysis). Although the IgG1 responses to rOv-CPI-2 antigen were not correlated with age (Fig. 1a), they were relatively elevated in all ages; with an overall 82.4% of IgG1 responders (mean 0.61±0.72). For the IgG1 analysis, there was no significant difference also in the proportion of responders vs. the non-responders in individuals of ≤20 years of age or those >20 years of age. The IgG3 response to the rOv-CPI-2 antigen was, however, positively correlated with age (r = 0.241; p<0.01) (Fig. 1b). For the IgE responses there was no significant correlation with age (Fig. 1c).
Since the rOv-CPI-2 protein was identified by immunoscreening with sera from infected individuals who are likely to have developed concomitant immunity, it was of interest to also investigate whether the native protein also induced the production of IgG and IgE antibodies in humans categorized as putatively immune (PI). These individuals have been exposed to the parasite over several years but have not developed clinical or parasitological signs of infection. The IgG subclass responses to rOv-CPI-2 in the putatively immune individuals (N = 21) were analyzed in comparison to the responses in age- and sex-matched infected individuals (N = 21). As shown in Figure 2, although the PI and INF individuals had similar anti-rOv-CPI-2 specific IgG3 responses (median of 0.3 and 0.32, respectively), the PI individual serum samples were generally more reactive with the antigen than the INF individuals (Upper 95% CI value of 0.730 for PI vs. 0.417 for INF). The median IgG1 responses of the INF, however, were significantly higher (p = 0.03) than those of PI (Fig. 2). The IgG4 responses against rOv-CPI-2 were also significantly higher (p = 0.04) in the INF than in PI as expected since IgG4 antibody responses are generally associated with an active and patent infection. On the other hand, the IgE responses between the two groups were not significantly different.
In vitro cytotoxicity assays demonstrated that purified monospecific human antibodies against rOv-CPI-2 in the presence of normal human neutrophils were able to inhibit 91% of molting from L3 to L4 as determined on day 6 in culture (Table 3), and resulted in almost a total loss of larval viability (72%). In normal conditions, about 43–68% of L3s molt to L4 in 6 days. The differences in % molting between the complete medium control and the negative non-specific antibodies control were not statistically different, while they are highly significant when the effect of monospecific anti-Ov-CPI-2 antibodies on molting was compared with the two controls (P = 0.0007 and P = 0.028 respectively). However, the viability of the L3 that did not molt in the presence of anti-Ov-CPI-2 antibodies was not statistically different from that of the control cultures. This was mostly due to the fact that in one of the three experiments the % viability was 82% vs. 0% and 2% in the other two experiments (data not shown). Longer cultures up to 10 days might be needed for consistent killing data.
The objective of the present study was to identify O. volvulus larval antigens that may be important in the development of concomitant immunity to the L3 that develops in the INF with increasing age, and which is independent of the immune responses that are directed against the adult worms and microfilariae associated with patent infection [6]. Concomitant immunity is a concept initially developed in cancer immunology and later also in lymphatic filariasis [5], [9]. It was only recently reported in onchocerciasis, where it is believed to contribute in preventing most of the newly acquired L3 infections from developing, and results in a stabilization of the adult worm and microfilarial burdens with age in the INF [3], [6], [4]. Thus, a pool of sera from individuals who have likely developed such immunity, based on their high IgG1 and IgG3 responses to crude extracts of L3 and mL3, was used to screen the corresponding cDNA libraries. The 20 distinct immunoreactive antigens identified included 4 proteins (Ov-CPI-2, Ov-FBA-1, Ov-103 and Ov-RAL-2) that had previously been shown to be protective in a small animal model [13], [25], [26], [27], [14], [15], [2].
Ov-CPI-2 (O. volvulus cystatin, also known as onchocystatin) appeared to be, overall, the most immunodominant cDNA clone reacting with the pool of sera. The high frequency of detection of the Ov-cpi-2 clone in the present study may be a reflection of its abundance in the L3 EST datasets as reported previously [12] and/or of the particular mix of antibodies against the larval proteins in the pooled sera we used for screening. Due to its immunodominance, Ov-CPI-2 was selected for further characterization with a focus on its increased recognition with age, a phenomenon that has been previously associated with the development of concomitant immunity [6] and thus strengthening its selection as a promising subunit vaccine candidate against O. volvulus infection. Ov-CPI-2 was previously cloned using antisera from chimpanzees experimentally immunized with attenuated L3s [28]. It was shown to be also expressed in all the other stages of the O. volvulus parasite except in mature microfilariae [28]. It was localized in the hypodermis, basal layer of the cuticle and in the eggshell of developing microfilariae [28]. Further studies established that this 17 kDa protein could be essential for development of the infective stage larvae as it probably regulates endogenous cysteine proteases that are essential for molting [29], [30]. Subsequent vaccination-challenge experiments in O. volvulus mouse model using recombinant GST-Ov-CPI-2 and alum as the adjuvant showed that it could induce significant protection levels of 43–49% [13]. Further support for its protective role has come from protective studies done with its homologues in other nematodes and including lymphatic filariae [31], [32], [33].
Analysis of the individual levels of the cytophilic antibody IgG3 to Ov-CPI-2 as a function of age showed an age-dependent association of the antibody responses, which were significantly increased with age and therefore consistent with the development of concomitant immunity to larval antigens in this population [6]. The INF and the PI individuals had similar anti-rOv-CPI-2 specific IgG3 responses (median of 0.3 and 0.32, respectively), while the INF had clearly higher IgG1 and IgE responses. The maintenance or up-regulation of cytophilic (IgG1, IgG3, and IgE) and complement-fixing (IgG3) antibody responses against larval antigens, some of which are not necessarily larvae specific, as was found in this study was shown before to be also associated with immune protection in onchocerciasis [6], [16], [17].
Our results provide further support to the possibility that the anti-Ov-CPI-2 antibodies may have a role in the ADCC effector mechanisms. We observed that human neutrophils inhibited molting of L3 by 79–100% in the presence of purified human mono-specific antibodies against Ov-CPI-2. Johnson et al [21] have shown that sera from PI and INF are able to inhibit molting of L3 and kill them when cultured in the presence of neutrophils from normal humans. Antibodies and ADCC were also found to be an important part of protective effector mechanisms in O. volvulus mouse model [15], [2].
In one preliminary study, when rOv-CPI-2 was used to stimulate human peripheral blood mononuclear cells (PBMCs) from INF individuals (n = 6) and PI (n = 4), it appeared that it induced a mixed IFN-γ and IL-5 response (cytokine levels were determined by ELISA) in approximately 50% of the samples (data not shown). When the cytokine responses were determined using the ELISPOT assay and PBMCs from additional individuals, 6 out of 6 PI individuals had higher (∼10×) frequencies of IFN-γ than IL-5-producing cells, while in the INF group 8 out of 10 individuals had IL-5 dominated responses, and the remaining 2 were only IFN-γ responders. Although the numbers tested so far are small, it points to the possibility that in the PI, the Th1 cytokine response dominates, while a more mixed Th1/Th2 response against Ov-CPI-2 is observed in the INF individuals. Overall, the data presented in this study clearly support the notion that Ov-CPI-2 is a promising leading target of protective immunity in onchocerciasis and the results discussed justify further vaccination studies on this antigen for human use.
|
10.1371/journal.pcbi.1003787 | Dynamic Mechanisms of Neocortical Focal Seizure Onset | Recent experimental and clinical studies have provided diverse insight into the mechanisms of human focal seizure initiation and propagation. Often these findings exist at different scales of observation, and are not reconciled into a common understanding. Here we develop a new, multiscale mathematical model of cortical electric activity with realistic mesoscopic connectivity. Relating the model dynamics to experimental and clinical findings leads us to propose three classes of dynamical mechanisms for the onset of focal seizures in a unified framework. These three classes are: (i) globally induced focal seizures; (ii) globally supported focal seizures; (iii) locally induced focal seizures. Using model simulations we illustrate these onset mechanisms and show how the three classes can be distinguished. Specifically, we find that although all focal seizures typically appear to arise from localised tissue, the mechanisms of onset could be due to either localised processes or processes on a larger spatial scale. We conclude that although focal seizures might have different patient-specific aetiologies and electrographic signatures, our model suggests that dynamically they can still be classified in a clinically useful way. Additionally, this novel classification according to the dynamical mechanisms is able to resolve some of the previously conflicting experimental and clinical findings.
| According to the WHO fact sheet, epilepsy is a neurological disorder affecting about 50 million people worldwide. Even today 30% of epilepsy patients do not respond well to drug therapies. Neocortical focal epilepsy is a particular type of epilepsy in which drug treatments fail and surgical success rate is low. Hence, research is essential to improve the treatment of this type of epilepsy. Recent advances in brain recording methods have led to new observations regarding the nature of neocortical focal epilepsy. However, some of the observations appear to be contradictory. Here, we develop a computational modelling framework that can explain the different observations as different aspects of possible mechanisms that can all lead to seizure onset. Specifically, we classify three main conditions under which focal seizure onset can happen. This classification is clinically important, as our model predicts different treatment strategies for each class. We conclude that focal seizures are diverse, not only in their electrographic appearance and aetiology, but also in their onset mechanism. Combined multiscale recordings as well as stimulation studies are required to elucidate the onset mechanism in each patient. Our work provides the first classification of possible onset mechanism.
| Neocortical focal seizures are episodes of pathological brain activity that appear to originate from spatially localised regions of the neocortex. The classical understanding of such seizures is that localised pathological tissue generates epileptic discharges (epileptogenic zone [1]), which subsequently recruit connected tissue, resulting in an epileptic seizure. Hence, the removal of the epileptogenic zone would result in seizure freedom [1]. Such a view is particularly applicable to focal epilepsy patients with e.g. cortical dysplasia, where a clearly localised anatomical abnormality of the cortex is present.
However, the classical understanding of neocortical focal seizures has not remained unchallenged, especially when treating patients without any clearly localised anatomical abnormalities. For instance, it is proposed that instead of a localised region of pathological tissue, an epileptogenic network [2]–[4] could underlie the generation of focal seizures. The spatial extent, and the participating regions of such a network are not yet clearly identified. Some indication is provided by the work of Stead et al. (2010) and Schevon et al. (2008), who report the recording of highly spatially localised epileptiform activity on the scale of cortical columns [4], [5]. Such electrographic activities, termed “microseizures” [4], [5], were recorded more frequently and for longer durations near the seizure onset zone [4]. Interestingly microseizures were also observed in non-epileptic control subjects, albeit in fewer locations and occurring less frequently than in epilepsy patients [4]. The authors hence proposed the hypothesis that “pathological microdomains” (i.e. microdomains that are able to generate and sustain isolated epileptiform hyperactivity states) might be found in healthy brains without leading to seizure onset. However, when occurring with sufficient density in one spatial area, they can form an epileptogenic network causing focal seizure onset from that area.
An alternative mechanism underlying (focal) seizure onset is proposed on the macroscopic scale. Badawy et al. (2009) demonstrated that the motor threshold of drug naive focal epilepsy patients decreased up to 24 h before a seizure on the ipsilateral side to the seizure focus [6]. A similar study using patients with mesial temporal lobe epilepsy also hints at an elevated motor cortex excitability preceding the seizure onset [7]. Hence, a change in overall cortical excitability has been suggested as a mechanism for focal seizure onset [8], [9]. This hypothesis is in line with the long-standing concept that seizures are a consequence of changing excitability of the brain [10]. However, the mechanism by which this general increased excitability over large parts of the cortex leads to focal onset dynamics is not specified.
An essential point of recent debate that is not explicit in either of the above suggestions of focal seizure onset mechanisms concerns the mechanisms of seizure recruitment and propagation. Based on the observation of single unit activity in human focal onset seizures, Truccolo et al. (2011) proposed that the recruitment process is “highly heterogeneous, not hypersynchronous, suggesting complex interactions among different neuronal groups even at the spatial scale of small cortical patches” [11]. In contrast, Schevon et al. (2012) suggests that seizure propagation is a well-structured process, where the recruitment progresses as a smooth wavefront. Recruited tissues show a synchronous firing activity that is phase locked to the local field potential.
It becomes clear that focal seizure onset and recruitment is still far from understood, and that prevailing hypotheses and observations lack a unifying framework in which they can be tested and analysed. In order to achieve this, we turn to mathematical models of cortical spatio-temporal dynamics. Traditionally, two types of models have been used: (i) Continuum models (e.g. [12], [13]) or neural field models (see [14], [15] for reviews) treat cortical tissue as a homogeneous continuous medium. The spatial extent often ranges from a few millimetres to a few centimetres [16], [17]. Pattern formation and travelling waves of activity have been studied extensively in these systems (see [18], [19] for reviews). Such spatio-temporal patterns have been related to epileptic activity. For example [17], [20] model the recruitment and propagation of a focal onset seizure as a propagating wave over a continuous medium. (ii) Network models treat the cortex as a connected network of cortical units (nodes), where often nearest neighbour, random, small-world or hierarchical connectivities are used. Depending on the definition of the network nodes, these models are used across all scales from local populations of neurons [21] to the whole brain level [22], [23]. Network based models have investigated how network structures impact seizure synchronisation dynamics [23]–[25], seizure frequency [26], or the spread of seizure activity from an epileptic focus [27],[28]. To specifically model the mesoscopic epileptic dynamics of extended cortical tissue, [29] suggests arranging coupled units of neural mass models (see [30] for review) as a sheet. Similarly, [31], [32] arrange neural mass models according to the tessellated surface of the brain and coupled neighbours to simulate scalp and intracerebral dynamics of focal seizures. Such an approach, although technically a network approach, can approximate the behaviour of continuum models (compare [29] and [20]). Recently, [33] also relates a network of mass models to an equivalent field model directly.
However, the connectivity in realistic cortical tissue appears to require a combination of both continuum and network approaches. Connections to nearby neighbours are very dense [34], such that it approaches the continuum case. Nevertheless structured long-range connections can form a complex network that is best described by a networks approach [35]. Hence, to describe the mesoscopic scale of the cortex, combinations of both network and continuum approaches have also been suggested, e.g. including heterogeneous long-range connections in neural field models (see for example [36]–[39]). Starting from a network perspective, Voges et al. (2010) propose to use a network model that includes dense local connections, approximating the continuum case, and incorporate remote excitatory connections that bridge distances of several millimetres [35]. The remote connections are furthermore structured and tend to target remote clusters or patches.
In this work, we advance upon previous spatio-temporal network models of cortical tissue on the mesoscopic scale and use a dense array of cortical units (cortical columns) that reflect the activity of local neuronal populations. For connectivity between the units we use the suggestion in [35] and incorporate dense local connections as well as patchy remote connections. This model has the advantage of combining both types of modelling approaches and thereby we create a spatially hierarchical model to study multi-scale dynamics.
Using this model, we investigate the dynamical mechanisms leading to the observation of focal onset seizure activity. We find that three different classes of dynamical mechanisms are compatible with a focal onset of an abnormal rhythm. Each of these classes show particular distinguishing features in terms of their dynamics and response to stimulation. Furthermore, they suggest alternative treatment strategies that could provide the basis to improve treatment options for patients in the future.
We commence by defining the smallest unit in our model: the cortical minicolumn. This choice is based on the highest spatial resolution of the clinical observations with which we compare the model output. To reflect the electric activity of a minicolumn, we use an established model of excitatory and inhibitory neural population activity: the Wilson-Cowan model [40].
This model expresses the percentage firing activity in an excitatory () and an inhibitory () neural population over time. It assumes that the and populations are coupled to each other and that the inputs to a target population sum together and influence the firing activity of this population. We use such a coupled unit to represent a single cortical minicolumn (see Fig. 1).
The equations for our model are: (1)where is the fractional firing activity in the excitatory population; is the fractional firing activity in the inhibitory population; and denote the basal activity level of the excitatory and inhibitory populations, respectively; is the noise input to the excitatory population (e.g. subcortical input) with as the coupling strength of the noise input; and the connectivity constants (with or ) regulate the coupling strength between the populations.
is a sigmoid function, which derives from a distribution of firing thresholds in the underlying neural population [40]. It is defined as , where is the steepness of the sigmoid and is the offset (in ) of the sigmoid. We fix the sigmoid parameters () following previous work [41], as variations in the other parameters effectively result in a change of the sigmoid shape.
The Wilson-Cowan model has been subject to extensive studies in the last decades [42]–[44]. The slightly simplified version in Eqn. 1 (see also [45]) was shown to maintain the same bifurcation structures as the original model [41]. The simplification removed the bracket of (or ) that the sigmoid was multiplied by in the original equations. Mathematically, the term has little impact on the dynamics. It is essentially rescaling the phase space and parameter space.
In order to model cortical tissue, we connect an array of minicolumn units to form a cortical sheet (also referred to simply as a sheet). We also refer to each of these minicolumn units simply as units. This formulation assumes an effectively two dimensional structure for the cortex. In reality, there is interplay between the three dimensional cortex, subcortical structures and other brain regions. However by making the simplifying assumption above, these influences in brain dynamics are absorbed into the intrinsic parameters of a minicolumn. Similar approaches of modelling the cortex as a 2D sheet can be seen in [12], [15], [29], [46], [47].
As an approximation we assume all minicolumns to be in size [48]. A macrocolumn is then formed by minicolumns, which agrees with the size suggested in [49]. Furthermore we investigate cortical sheets with minicolumn units (i.e. , or macrocolumns). Thus and in Eqn. 1 become vectors of the length ; and the connectivities become matrices of the dimension . We limited the size of the sheet to minicolumns in length, as we assume the mean activity of such a sheet reflects the signal recorded on a single ECoG electrode.
Each excitatory population is additionally driven by noise () representing input from other unmodelled regions, e.g. subcortical input. The noise is the same within each macrocolumn in agreement with experimental findings and the definition of macrocolumns [48], [50]. We used noise values drawn from a standard normal distribution as input. The effective noise coupling strength is set to . In this setting the system is not entirely dominated by the noise input but the noise influences the deterministic dynamics. Simulations of the system used a fixed step solver, with a stepsize of 2 ms. Qualitatively equivalent results are found for smaller stepsizes. Fig. 1 schematically summarises the model.
In the model we use three types of connections between minicolumns, based on the cortical connectivity suggested by Voges et al. (2010) [35]. All choices for parameters of the connectivity are also based on [35], where they are derived from tract tracing experiments in human cortical tissue.
(I) The first type consists of local excitatory connections, where each excitatory population of a minicolumn unit connects to the excitatory populations of neighbouring units in its immediate proximity (Fig. 2 (a), top). Here, each unit has a probability to connect to its neighbours that follows a Gaussian fall-off with distance. The standard deviation of the Gaussian is set to , as within radius most local connections are found [35]. We furthermore do not allow for local excitatory connections beyond a radius as these are incorporated into a specific longer range connectivity scheme, as described below. Fig. 2 (a, bottom) shows an example of one unit (red) and the neighbouring units (black) is sends local excitatory connections to. The connectivity matrix for the local excitatory connections is denoted , where each connection has the weight (subscript denoting local connections).
(II) The second type of connections is from the excitatory population of each unit to the inhibitory populations of close neighbours (Fig. 2 (b), top). We use the same algorithm and parameters as in (I) to generate these connections. Fig. 2 (b, bottom) shows an exemplary realisation of local inhibitory connections from one unit. We refer to the connectivity matrix for the local inhibitory connections as , where each connection has the weight .
(III) The third type are remote patchy overlapping connections from each excitatory population to excitatory populations at some distance (Fig. 2 (c)) [35]. All the parameters are following the suggestions in [35]. We generate random patches for each macrocolumn and all minicolumns within the macrocolumn can connect to these patches with outgoing connections. This fulfils the suggested average ratio of of local connections to remote connections [35]. Each patch consists of minicolumns (the patch radius is , i.e. 5 minicolumns radius, i.e. minicolumns) and is located within distance. Macrocolumns share patches with one direct neighbouring macrocolumn, which can increase the distance between macrocolumn and target patch to more than . These parameters are in line with the suggested and experimental values listed in [35]. The algorithm that generates the remote connectivity matrix is described in Text S1. We call the connectivity matrix for the remote excitatory connections , where each connection has the weight , (subscript denoting remote connections).
The connectivity matrix therefore consists of , and the self-excitation value of each excitatory population on the diagonals (, subscript denoting self connection). Similarly consists of and the connection value of the connection within the minicolumn unit on the diagonals (). The other matrices are diagonal matrices only, as they are exclusively connections within a minicolumn. Long-range inhibition is not included, following [35].
In order to aid the understanding of the resulting connectivity being created by the aforementioned rules, Fig. S1 additionally show the in/out degree and the distance distributions of the local, as well as remote connections. Text S1 further explains the details of the connectivity.
A cortical sheet with toroidal boundaries was used in the construction of the connectivity matrices, following [35], to avoid boundary cut-off effects caused by lack of basal input due to lack of neighbours. Text S2 discusses in detail how different boundary conditions affect the model dynamics and we will show that all our presented results are not affected qualitatively by the choice of boundary conditions.
The choice of model parameters for the isolated Wilson-Cowan unit was based on dynamical reasoning. The dynamics of a single minicolumn unit (in the following referred to as unit) has been subject to extensive studies. The invariant dynamic behaviour in an unit is limited to either a stable fixed point (node or focus) or a stable limit cycle, and two stable fixed points (see [41], [42] for details). As we are interested in the transition between fixed point and limit cycle we select model parameters in the vicinity of the transition to oscillations. Depending on the combined parameter variation, either a homoclinic or a Hopf bifurcation occurs. However the single unit in model is incapable of oscillations, even with increased constant input . Text S3 shows details of the current parameter setting for a single node.
Based on the dynamics of a single unit, the dynamics of the fully coupled sheet is classified as: (i) fixed point (corresponding to the lower fixed point in the unit; the spatial average of the whole system does not show prominent regular oscillations over time); (ii) oscillation (corresponding to the limit cycle in the Wilson-Cowan unit; the spatial average of the whole system shows high amplitude oscillations over time); or (iii) bistability between fixed point and oscillatory state. Although the coupled Wilson-Cowan systems are known to show a complicated repertoire of oscillatory states (in term of regularity and phase relationships [43]), we do not sub-classify the oscillatory states further. The epileptic EEG or ECoG has a considerable noise component and is non-stationary such that a reliable classification from clinical data is challenging. Also, a theory of spatio-temporal patterns in large heterogenous networks of nonlinearly coupled nonlinear oscillators is lacking. However, it was shown previously that a combination of mathematical understanding of a single network unit and computational studies of the network can lead to improved understanding of clinically relevant phenomena (e.g. the generation of oscillatory afterdischarges in epileptogenic cortical tissue [29].
In the clinical setting diverse waveforms can be observed in electrographic recordings of neocortical focal seizures. However, we seek a simplification of this situation in our model, which captures some essence of abnormal dynamics during seizures. We therefore focus on the existence of high-amplitude oscillations in the model output as representative of seizure activity, in contrast to a low-firing state, which is representative of “background” or inter-ictal activity. This idea follows previous modelling studies (for example [51]–[53] and references therein). The approach is further supported by the suggested clinical definition [54] of a seizure state as oscillations in unit firing, which are phase locked to high amplitude local field potential oscillations. The background state is characterised by irregular firing patterns, which do not correlate with any oscillations in the local field potential.
In a single unit, we shall hence identify the background with the fixed point. As the Wilson Cowan oscillator only has one limit cycle representing synchronous rhythmic firing activity on the local population level, we shall identify this limit cycle with the local seizure state. Our model is additionally capable of a third state: the permanently firing state (referred to as “upper fixed point” in the Wilson-Cowan model). This state is not identified with any clinically observable state, and we hypothesise that the parameter settings required to reach this third state do not play a functional role during focal-seizure onset.
In the simulated coupled sheet, we understand high-amplitude synchronous (plus minus phase shift) oscillations in firing and LFP over several connected units as the seizure core [54]. Hence, full recruitment will be understood if the whole sheet is in such a state, where all units are in a synchronous oscillatory state. Text S5 describes how we detect these full recruitments or localised non-recruiting seizure cores for each figure. Other types of oscillations (e.g. non-synchronous low amplitude oscillations, which could represent non-pathological oscillations) on the full-sheet level were not specifically identified or analysed.
The matlab code for the model is published online (ModelDB Accession number: 155565).
In a first step we focus on the mean-field dynamics of the model and how they vary due to changes in parameters. In subsequent sections we use this insight to investigate the spatio-temporal mechanisms by which focal onset seizures can occur. For each mechanism we summarise how it can be distinguished from other mechanisms, how they relate to clinical and experimental observations, and which treatment strategies could be effective.
To chart the dynamics of the model cortical sheet with respect to parameter changes, we focus initially on spatially homogeneous variations in the four parameters highlighted with red frames in the schematic in Fig. 3, i.e. , , and . Fig. 3 demonstrates that there are large regions of parameters for which the system resides in the background state (black regions in Fig. 3), or oscillatory state (dark blue regions in Fig. 3). Additionally, in some parameter regions the oscillatory state can be found to be bistable to the background state (light blue regions in Fig. 3). A consequence of this is that a system in this parameter region can exhibit either background or oscillatory dynamics under the same parameter conditions. The transition from background to oscillatory activity is dependent upon all four scanned parameters. Pairwise scans of additional model parameters can be found in Text S4, which demonstrates that combinations of other parameters also give rise to background, oscillatory or bistable dynamics.
From the dynamical systems perspective it is often assumed that the epileptic brain resides in a parameter setting close to the onset of oscillations [51]. Hence, we selected one standard parameter set for our model in line with this idea, as indicated in Fig. 3 (a) and (b) by a red dot. This standard parameter set serves as our model interictal state, or monostable background state. Dynamically, the interictal state is a node and excitability can be detect for a range of stimuli in this state (data not shown). For an exemplary parameter change (red arrow in Fig. 3 (a)) we have also analysed the transitions in detail in a noise-free system. The monostable background state is the only stable fixed point in our system at . The onset of bistable high-amplitude oscillations occurs suddenly at . At the transition to monostable oscillations (at about ), the background node ceases to exist and the oscillatory state becomes the only stable state. When changing from the standard interictal parameter setting, similar transitions occur, only the background node remains stable and does not cease to exist.
Having demonstrated the effect of global parameter changes on the mean-field dynamics of the model, we proceed to examine the different ways in which transitions to seizure activity can occur spatio-temporally.
The parameter scans in the previous section imply that a slow parameter change that crosses from the background to the oscillatory region can induce a transition from background to seizure dynamics on the mean-field. An example of such a parameter ramp over time is indicated by the arrow in Fig. 3 (a) and in Fig. 4 (a). This suggestion follows a traditional modelling approach of seizures induced by bifurcations (see for example [55], [56], also c.f. [57]). In simulations of this scenario the onset of the abnormal rhythm is approximately simultaneous in all spatial locations, as the corresponding parameter is modified simultaneously in all units across the sheet. In our case, the transition occurs at about as a bifurcation from a node to an oscillatory state, where the onset of oscillation frequency and amplitude is sudden and discontinuous, and the node ceases to exist.
Using such a transition, we sought to establish whether the model can produce focal onset dynamics. Typically cortical tissue is not globally homogeneous. We therefore consider the impact of a locally altered region of model cortex, which is realised by assuming a local parameter heterogeneity in the model. We specify a patch in the middle of the model cortical sheet that receives increased feed-forward excitation. This heterogeneity is not visually detectable in the interictal state (see Fig. 4 (c), first panel). However, in a simulation with a parameter ramp as shown in Fig. 4 (a), the heterogeneous region displays an earlier response (Fig. 4 (c)). Dynamically, the earlier ignition of activity in the heterogeneity is due to the introduced difference in feed-forward excitation, which lowers the threshold beyond which oscillatory dynamics can ensue.
Next we explore the mechanisms leading to focal onset rhythmic activity when the whole model cortical sheet is in the bistable background state. A bistable state has been proposed to underlie situations in which the transition to abnormal brain activity is not caused directly by global parameter changes (see e.g. [59]). It was postulated specifically as underlying the transition to epileptic seizures in the context of generalised [53], [60] and focal seizures [17], [28], [61].
In order to explore this scenario we prepare the cortical sheet in a global parameter setting of bistability. If the sheet is initiated in the background state, it will remain in the background state in the absence of strong perturbations. To initiate oscillatory activity, the background state can be disturbed in two possible ways: either by a short, temporary stimulus or by a persistent stimulus. We shall explore both perturbations in the following.
We prepared the cortical sheet in the bistable background state by decreasing the feed-forward inhibition compared to the interictal parameter setting. Our choice to decrease feed-forward inhibition is inspired by the suggestion that a failure of inhibitory restraint [54] contributes to seizure onset. Equally, a change in , or other parameters could have been used.
In our model, the bistable background state (dynamically also a node) does not show any obviously different dynamics compared to the monostable background state. However, when perturbed locally by a pulse-stimulus, the whole cortical sheet can transit to the co-existing oscillatory state. In Fig. 5 (c) and (d) we demonstrate how this transition unfolds in terms of spatial-temporal dynamics. After the stimulation, a subset of the stimulated units transits to the bistable oscillatory regime, which subsequently recruits neighbouring units into the oscillatory state. The recruitment in this connectivity parameter setting progresses as a wave, similar to the observation by Schevon et al. (2012) [54]. The comparison between clinical data and the simulation is invited in Fig. 5. In both cases the continuous progression of a wavefront of increased firing activity is observed.
In the current connectivity setting, heterogeneities in the propagation dynamics can also be observed. This is due to the heterogeneously created remote projections, which can support the activation of tissue at some distance from the primary recruitment site. A purely local connectivity creates an even propagation front (Text S7A, and Fig. S8 (a)), and a purely remote connectivity gives rise to stochastic patchy propagation (Text S7A, and Fig. S8 (b)). A mixture of propagation behaviours between these two extremes can be observed when a connectivity scheme that combines both features is used (as in Fig. 5 (d)).
To demonstrate that these findings are repeatable and reliable despite the noise input to the system, we scanned the recruitment speed after a (fixed) pulse stimulus for different values of and in and around the bistable parameter setting. Averaged over 5 trials using different noise input, little variation in recruitment speed due to noise was found for a fixed parameter set. However, recruitment speed did vary with the parameter settings (see Text S7B for details).
The stimulus size was also found to influence the recruitment, and a minimal stimulus size was found to exist depending on the parameter setting (Text S8A and Fig. S11). This means that a critical number of units have to be stimulated to induce the transition of the sheet to the seizure state, when it is bistable. This finding is potentially important for the clinical determination of the spatial extent of pathological stimuli in a patient-specific context.
A perturbation to the model sheet need not be externally generated, but can arise due to local, abnormal activity generated within the model. In Fig. 6 we demonstrate that the existence of an oscillating patch in the sheet can also trigger a transition into seizure dynamics.
Fig. 6 (d, e) demonstrate that if the parameter setting of the surrounding sheet is monostable in the background state, the hyperactive microdomain remains isolated in its epileptic activity (red trace in Fig. 6 (d)). This agrees with the clinical observation of spontaneous microseizures that remain spatially localised and do not recruit surrounding tissue. If, however, the rest of the system is in a bistable setting, a continued local perturbation by an oscillatory microdomain can start to recruit the surrounding units into the seizure state (Fig. 6 (f, g)).
The propagation pattern of recruitment is similar to the case of recruitment following a pulse stimulation to the bistable sheet. Depending on the connectivity settings, smooth propagating waves, patchy propagations, or a mixture of both can be observed. Text S7B demonstrates that an oscillatory microdomain can produce recruitment speeds of between and , using some example parameter changes in . This is within a range of propagation speeds reported experimentally (0.1–100 m/s [63]).
In order to check the robustness of this onset mechanism, we tested the dependency of recruitment on both the parameter setting of the surrounding and the size of the pathological microdomain (see Fig. S12 and Text S8A). We find that for a fixed bistable parameter setting, a minimum threshold exists for the number of units that are required to induce recruitment. When the parameter setting lies closer to the monostable oscillatory setting, fewer units are required for recruitment. This behaviour is stable with different noise inputs and microdomain positions in the model sheet. This finding is important for the clinical determination of pathological vs. neutral microdomains in a patient-specific context.
After demonstrating focal seizure onset in a globally oscillatory and a globally bistable scenario, we now turn to the case of a globally monostable background. We shall investigate a system in the monostable background state except for one or multiple localised hyperactive microdomains. We examine the spatial conditions under which these hyperactive patches can recruit their surrounding, even though globally the oscillatory state neither exists exclusively (class I) nor coexists (class II) in the absence of these patches.
We prepare the system in the monostable background state (the standard interictal parameter set), except for some oscillatory microdomains. We begin by systematically assessing how the recruitment from these microdomains depends upon the total number of hyperactive units and the number of subclusters that microdomains are organised into. Here a subcluster is a contiguous patch of units, positioned randomly on the sheet. Fig. 7 (a) shows that despite the surrounding being in the monostable background state, partial or full recruitment can be registered in some configurations. E.g. Fig. 7 (a) demonstrates that when 2250 units (10% of the whole sheet) are hyperactive, no recruitment is registered if all the units are organised into one compact patch (red dot). However if this same number of hyperactive units are organised into 17 subclusters of equal size, randomly distributed over the model sheet, noticeable recruitment can be observed (purple triangle). The recruitment behaviour additionally depends on the exact parameter of the surrounding. For example if the exogenous input parameter, , is set to a value closer to the global bistability (, Fig. 7 (a)) recruitment starts at a lower total number of hyperactive units and with a lower number of subclusters than when using , further away from the bistability (Fig. 7 (b)). This recruitment behaviour is stable with regards to the noise input in our simulation. However, the exact values of the total number of hyperactive units and the number of clusters vary slightly with the position of the (sub)clusters (see Fig. S13, Text S8A).
Fig. 7 (d,f) show example time series from two simulations using the same number of hyperactive units (2250 units, 10% of the whole sheet), but different numbers of sub-clusters. In the case of a single cluster, only a few units are recruited (Fig. 7 (f)). In the case of many sub-clusters, the whole sheet is recruited (Fig. 7 (d)). Recruitment can be observed to begin in areas of increased subcluster density (for example right side of T = 1.6 s in Fig. 7 (d)). The “normal” monostable tissue between nearby subclusters is recruited first. In this way, the subclusters that lie in close proximity recruit the healthy tissue between them to form a bigger contiguous cluster of oscillatory activity (T = 2 s in Fig. 7 (d)), and eventually recruit the whole cortical sheet (T = 4 s, T = 10 s in Fig. 7 (d)).
The recruitment of monostable surrounding tissue in the background state is not as intuitively understandable as for instance the case of recruitment of a bistable surround. The scans in Fig. 7 (a,b) show that the spatial arrangement of “recruiters” is important. We propose that the basic mechanism is based on the coherent oscillatory input to units in the background state, which can incite them to oscillate despite their configuration being monostable. The parameter change in the microdomains induces the microdomains to become intrinsically oscillatory. Hence, the recruitment from microdomains induced by this local parameter change is a bifurcation from a node to an oscillatory state. The onset of oscillations, while ramping , occurs with a sudden change in frequency and amplitude. We additionally address the effect of boundary conditions on this mechanism in Text S2.
We have shown that clusters of autonomous oscillations can induce recruitment of the whole system to the seizure state. In this section we investigate additionally whether a system-wide bistability can be induced by localised, bistable clusters of tissue (i.e. a set of bistable microdomains). The reasoning is that if the network of microdomains is bistable, specific localised stimuli will be able to induce localised oscillatory behaviour in the patches, which in turn would lead to recruitment of the monostable surrounding as in class IIIa.
For such a scenario, it is required to determine the conditions under which a local cluster of tissue is bistable. Hence we scan the size of a microdomain embedded in a monostable background surrounding versus an exemplary local parameter change () and determine whether a microdomain patch is bistable by applying a single-pulse stimulus (Fig. 8 (a)). An elevation in leads to bistability of the microdomain. Upon further increase of , the microdomain becomes monostable oscillatory. This bifurcation also occurs with a sudden change in amplitude and frequency. As the patches become smaller, has to be higher to reach bistability (or the monostable oscillations) in the microdomain. The dependency of the dynamic behaviour on the size of the microdomain can be understood if we consider that the oscillatory state in the system emerges from the coupling of individual units.
Using information from the previous parameter scan, we set up a monostable sheet and distribute bistable microdomains within it (Fig. 8 (b)). Such a system remains in the monostable background state in the absence of perturbations. Multiple single-pulse stimuli applied randomly at different locations can be used to activate some bistable patches (Fig. 8 (c)). Some degree of coactivation (i.e. an active patch subsequently activating a connected silent patch) can also be observed. Once activated and in high enough density, the patches can cause recruitment of their non-oscillatory environment as shown in the previous section. Fig. 8 (c) shows a time course of multiple stimuli activating silent bistable patches, which ultimately results in full recruitment of the sheet.
In this study we used a novel spatio-temporal model of the dynamics of cortical minicolumns, coupled by multi-scale cortical connectivity, to categorise possible mechanisms of focal seizure onset. We showed that in this framework, apparently conflicting clinical observations regarding focal seizure onset can be understood and unified. We furthermore suggested how to test for the different onset categories, and made predictions regarding possible treatment methods for each category.
The three mechanisms we identified by which a focal seizure onset can occur are: (I) A global parameter change which induces a global bifurcation of a piece of cortical tissue to the seizure state. (II) A global bistability combined with a local trigger leading to transition to the seizure state. (III) A globally monostable state with local parameter changes causing recruitment of the whole system. We expect that either mechanism may dominate the onset of focal seizures in different patients.
The model employed herein uses the approach of discretised, coupled spatial units to reflect the activity of a piece of contiguous cortical tissue. Each unit in the current model is described by Wilson-Cowan equations, which embody the collective activity of local excitatory and inhibitory neural populations [40]. Compared to detailed neuronal models of cortical activity (e.g. [67]), the Wilson-Cowan model is computationally less demanding and the number of parameters to analyse is manageable. However the parameters of the Wilson-Cowan model are more abstract in nature. Thus, if for example cellular mechanisms of focal seizure onset are to be investigated (e.g. [68]), a detailed neuronal model is required. Similarly, if the detailed laminar and horizontal interaction between different types of excitatory and inhibitory populations is of interest, the populations in our model can be extended. However, in our current study, describing the dynamics of cortical minicolumns in terms of the lumped activity of generic excitatory and inhibitory neural populations allowed us to model a hierarchy of clinically relevant spatial scales by reducing the level of detail for the analysis.
The classical Wilson-Cowan model has been used to reflect EEG/ECoG dynamics in the delta to beta range [41], [69]. Similarly, we used it here to model seizure oscillations in this frequency range. Faster or slower dynamics are therefore not considered in our current approach, although it will be interesting in future studies to investigate the influence of these aspects, for example the addition of slower time scales. The incorporation of additional intrinsic long-term dynamics (e.g. adaptation or learning) can lead to the creation of additional types of dynamics, which could also be relevant for clinical question. If the time scale separation is sufficient (i.e. intrinsic long-term dynamics are on the order of seconds or longer) Fenichels theorem [70] indicates that our presented attractors would remain as manifolds in the full system with a slower time scale. Hence the slower time scale dynamics would modulate and orchestrate the transitions between the stable dynamics presented here. Indeed, the global parameter configuration (monostable, bistable, and oscillatory) used in our current model could be fluctuating over time according to some slow dynamics. It might be that the parameters of the cortex of patients as well as healthy subjects are constantly changing [8], putting cortical tissue in different global configurations at different times. However, in an epileptic patient, either these global fluctuation are either too extreme leading to a global bifurcation into the seizure state (class I), or would remain silent if not co-occuring with a local trigger (class II), or do not affect seizure onset directly (class III). In patients with stereotypical seizure onset (i.e. the seizure onset is repeatedly from the same region with a similar electrographic pattern), the underlying long time-scale dynamics are either similar from seizure to seizure, or at least giving similar dynamical conditions. Hence the categories would apply to all seizures of the same stereotype (in the same patient). Our classification is hence crucial to determine (patient-specifically) the exact role of the parameter fluctuation dynamics in seizure onset. Practically, a constant multi-scale monitoring of the cortical activity, as well as regular stimulation tests should be carried out to determine the global and local parameter configuration.
In our model, we equated high amplitude oscillations with a pathological state in each mini column. This is mainly inspired by the observation that the seizure core contains highly active neurons with firing patterns phase locked to the oscillatory LFP [54]. We believe that in our case, firing activity might provide a better benchmark for comparison of clinical and simulation data than LFP, as the generators of the different components of focal seizure field potentials are largely unknown. Hence, following [54], we identified high amplitude oscillations in firing as the seizure state and low level firing as the background state. Additionally, the approach of identifying oscillations with seizures and fixed points with background activity is well established in the modelling literature (see for example [51], [55], [56], [71], [72]). It is in line with the long-standing suggestion of dynamic diseases [73], [74], where the disease state is identified as an oscillatory attractor and the background state as an non-oscillatory, primarily noise dominated state. Only very little clinical or theoretical understanding exists regarding the different waveform morphologies in focal seizures [75], [76] and how seizure onset mechanisms influence them. Weiss et al. (2013) [77] point out that high frequency oscillations phase locked to low frequency oscillations at seizure onset could be an indicator for increased, structured firing in the underlying tissue and hence an indicator for the seizure core. Future studies should specifically investigate how focal seizure onset field potential morphologies arise, as well as how they relate to firing patterns. Potentially, the knowledge gained by studies of waveform morphology in purely temporal framework such as [41], [71], [78] could be of use.
Each of the onset mechanisms we describe relies on a certain configuration of global parameters, where global is in reference to the scale of the model of about one square centimetre of cortex. However, in reality global parameter changes in the brain will vary from the whole-brain level to the scale of our current model, all of which can influence the global parameter configuration in our model. A range of physiological and pathological conditions could cause such variations. For example different phases of the sleep-wake cycle or hormonal variations [8] can change the excitability of brain. Pathological conditions include misregulation of excitation and inhibition [10]. If pathological parameters changes exist in a limited part of the cortex, then the focal seizure could be limited in its spatial extent. However, if abnormal dynamics entrain a large region, they could activate other whole-brain networks (including subcortical networks) leading to generalised abnormal activity (secondary generalisation).
In this context the model can also resolve the apparent contradictions in the experimental literature on the mechanisms of focal seizure onset. The contradiction of focal seizure onset being a result of global (whole brain network changes) or local (abnormally behaving cortical columns) mechanisms is no longer a contradiction in our model. We have shown that global as well as local mechanisms can interact and we have classified the interaction in three major categories. Hence, global changes can cause (class I), or support (class II), or modify (class III) seizure onset. Equally, local changes can trigger (class II), cause (class IIIa), or support (class IIIb) focal seizures. It is hence no surprise that clinical and experimental observations supporting both global as well as local mechanisms are found. Similarly, the contrasting observations from [11] and [54] can also be united: it might be that very near to the “focus” recruitment propagates as a wave over the local network. However, further away regions are probably recruited via remote or long-distance connections first and activation is primarily patchy. Hence, the conflicting recruitment dynamics described by Truccolo et al. (2011) and Schevon et al. (2012) is explained in our model by the propagation of activity via different networks. Interestingly, Schevon et al. (2012) [54] hypothesised that the ictal penumbra could restrain the propagation of epileptic activity due to an “inhibitory veto”. In our model of non-recruiting microdomains, we find that the restraint is not explicitly excessive inhibitory firing activity in the penumbra. Rather, the net synaptic input into each unit in the penumbra is not strong enough to entrain them to become oscillatory.
A question that arises from our study is whether the categories we established can be generalised to any spatio-temporal system showing bifurcations or bistabilities between a non-oscillatory (fixed point) and an oscillatory state. We propose that the detailed transition dynamics will depend on the specific system. However, we postulate that the three categories are general features of spatio-temporal systems showing either a bifurcation or a bistability between fixed point and oscillation. This is mainly due to the observation of the three categories in other spatio-temporal models using different model formalisms as well as underlying connectivity. For instance [17] essentially show class IIa in their partial differential equation model. Class I has been shown in a coupled Amari-type model representing a whole-brain network [24]. [64] show a class IIIa transition in their rule-based model of microseizures and recruitment. To our knowledge, class IIIb has not been demonstrated explicitly so far. We emphasise that the dynamical classification only becomes useful in the context of a relevant model, and the interpretation becomes useful when it is applied to the clinical context, e.g. to search for the cause of the seizure and to devise potential treatment strategies.
We have outlined major features and the expected observations of each class of onset mechanism in the Results section. A question that remains is how one would practically tell the classes and subclasses apart in a clinical setting. This question is crucial, as treatment will depend on the individual mechanism of seizure onset in a patient. We suggest that high resolution spatio-temporal recordings, similar to [79], combined with local perturbation studies (similar to [80], but on different spatial scales) might be the key to answer this question. In the context of local perturbations, we point out that although we only demonstrated the impact of pulse stimulation in our current study (to essentially reset the activity of the excitatory population), practically the effect of different types of stimulation has to be assessed prior to its usage for the classification of the seizure onset.
In this context, we recommend the development of patient-specific models to classify the dynamic seizure onset mechanism. This would involve incorporating the patient-specific connectivity of the affected cortical area (e.g elucidated from high resolution track density imaging [81]–[83]), as well as online parameter fitting according to passive and active high-resolution spatio-temporal recordings. This could enable the use of closed loop counter-stimulation devices (as demonstrated in Fig. S15). Additionally, such patient-specific models can be employed to predict optimal treatment protocols, for example minimal cortical micro-incisions to stop the recruitment of tissue into full seizures (see Fig. S16 and [84]).
In the case of a global shift of parameters (affecting larger brain regions) causing or facilitating seizure initiation and recruitment, it is probably desirable to target the reason for the global shift directly rather than trying to suppress seizure onset locally. In fact class I onset demonstrates that although one particular cortical location appears to be the source of seizure initiation (epileptogenic zone), the mechanism causing the seizure can be a global parameter shift in an extended tissue. The “epileptogenic zone” only reacts first due to its increased local threshold. Then, despite reducing or removing the local activity in the seizure onset zone, the seizure still starts, albeit from a different “most active” site. This concept of the existence of alternative foci has been proposed from clinical reasoning [1], [3] to explain why some surgical resections of epileptogenic zones have little effect.
Conceptually, we hence propose to distinguish between global or generalised causes of focal seizures, which induce the seizure by a global parameter shift - and local or focal causes of seizure, which can be facilitated by global bistability settings. The spatial extent of the cause of the seizure, however, can differ greatly from the spatial extent of the observed seizure onset. The traditional concepts of the epileptogenic zone and the seizure onset zone do not fully account for this. The understanding and treatment, of focal-onset seizures might benefit from further clinical and computational studies of seizure onset mechanisms on multiple spatial scales.
|
10.1371/journal.pbio.2000936 | CACNA1C gene regulates behavioral strategies in operant rule learning | Behavioral experiments are usually designed to tap into a specific cognitive function, but animals may solve a given task through a variety of different and individual behavioral strategies, some of them not foreseen by the experimenter. Animal learning may therefore be seen more as the process of selecting among, and adapting, potential behavioral policies, rather than mere strengthening of associative links. Calcium influx through high-voltage-gated Ca2+ channels is central to synaptic plasticity, and altered expression of Cav1.2 channels and the CACNA1C gene have been associated with severe learning deficits and psychiatric disorders. Given this, we were interested in how specifically a selective functional ablation of the Cacna1c gene would modulate the learning process. Using a detailed, individual-level analysis of learning on an operant cue discrimination task in terms of behavioral strategies, combined with Bayesian selection among computational models estimated from the empirical data, we show that a Cacna1c knockout does not impair learning in general but has a much more specific effect: the majority of Cacna1c knockout mice still managed to increase reward feedback across trials but did so by adapting an outcome-based strategy, while the majority of matched controls adopted the experimentally intended cue-association rule. Our results thus point to a quite specific role of a single gene in learning and highlight that much more mechanistic insight could be gained by examining response patterns in terms of a larger repertoire of potential behavioral strategies. The results may also have clinical implications for treating psychiatric disorders.
| To deal with an uncertain and complex world, animals have developed a large repertoire of behavioral heuristics and default strategies that spring into action in unknown situations. Building on this a priori repertoire, animals may find various ways to succeed on a given behavioral task. Therefore, determining the exact behavioral strategy followed during a task may be essential for understanding the cognitive processes involved. Using computational models to analyze behavior, we examined how a genetic variation in a gene that encodes a calcium channel and has been associated with learning deficits influences the way in which animals acted on a task in which a reward was associated with a specific behavior. We found that a knockout of the relevant gene does not lead to a general learning impairment but rather led animals to adopt a behavioral strategy different from the one employed by the control animals. Specifically, knockout animals managed to increase their reward returns by basing their responses more on the previous reward location rather than on reward-indicating stimuli, like the controls did. These findings may prove useful for behavioral therapy in the context of psychiatric disorders associated with this specific gene variation.
| The ability to learn, to adapt one’s own behavior in order to optimize positive and avoid negative feedback, is central to all living beings. Animals like humans are constantly seeking to infer the (causal) structure of their environment and to predict the outcomes of their actions [1,2], sometimes to a degree that “superstitial beliefs” about environmental contingencies may form [3]. However, since the environment is only partially observable, with an often infinite or exponentially exploding space of possibilities, inferring the optimal course of actions is, in general, a computationally intractable problem [4]. Hence, animals like humans rely on heuristics, strategies, and ecological biases that favor certain types of environmental contingencies over others and narrow down the search space [5,6]. For instance, rodents are predisposed to employ ecologically sensible strategies like win-stay, lose-shift, or alternate [7–9], and these behaviors are also more readily acquired [10,11]. Moreover, recent evidence suggests that animal learning, even on apparently simple conditioning tasks, may often engage active, evidence-driven choices among behavioral strategies, rather than mere passive strengthening of stimulus-response (SR) associations [12,13]. The upshot here is that animal learning may rely on a large variety of different previously acquired or predisposed strategies, rather than on a uniform mechanism, like SR strengthening, that could either be enabled or disabled.
Given this larger repertoire of potential a priori strategies and heuristics with which animals may enter any given experimental task context, there may be more than one way for increasing reward [14,15], even though this may not have been experimentally intended. In such situations, more conventional analysis in terms of error counts or reaction times compared among experimental groups may perhaps (wrongly) infer that one group is simply diminished in its learning abilities compared to another, while the observed differences may actually be rooted in the different behavioral strategies applied. Individual differences in learning strategies have recently been approached within the framework of computational reinforcement learning (RL) and decision-making theories [e.g., 16,17], which have been applied, for instance, to disentangle model-free and model-based forms of learning and their distinct neural substrates [18–20] or to reveal strategic interactions during (model-based) sequential planning [21]. Building on ideas like these, we show here how a genetic variation may determine an animal’s learning strategy and associated performance patterns. Specifically, we studied learning in terms of behavioral strategies on a 2-choice operant cue discrimination task in 2 groups of mice, a group with a genetic modification associated with severe learning deficits [22], namely, a selective knockout (KO) of the Cacna1c gene coding for the α subunit of Cav1.2 L-type high-voltage-dependent calcium channels (Cav1.2NesCre), and matched controls with intact gene expression (Cav1.2fl/fl). Calcium influx through voltage-dependent calcium channels is essential for synaptic plasticity and other cellular adaptations [23] thought to underlie instrumental learning [24]. The CACNA1C gene has also been implicated in the pathophysiology of psychiatric disease [22,25], rendering it an important target for understanding learning mechanisms not only in normal but also psychiatric conditions. By studying the detailed pattern of behavioral responses and rewards received across the learning process, combined with analyses based on behavioral (reinforcement) learning models, we found that Cav1.2NesCre mice did not simply exhibit a general learning deficit but rather relied on different behavioral strategies than Cav1.2fl/fl mice.
To examine potential learning deficits caused by selective ablation of the Cav1.2 L-type calcium channel subunit, a cue discrimination task consisting of 3 different task phases was set up (Fig 1). The basic task required animals to perform a nose poke response on one side of a touchscreen box if a light square came up in the upper central position of the screen and to the other side if the stimulus was shown in the lower position (Fig 1I). Hence, the animals had to associate cue location (top, bottom) with response side (left, right) for a maximum of reward returns (correct responses according to this “cue rule” were always rewarded). In task phase I, trials with the cue in top or bottom position occurred in a pseudorandom order (see Materials and methods for details), but correction trials consisting of repetitions of the unsuccessful trial were run after each wrong (nonrewarded) response. In task phase II, “slightly ambiguous” trials with reduced stimulus contrast were introduced, i.e., with cues appearing in both positions but the one indicating the correct response side brighter than the other (Fig 1II). In phase III, “fully ambiguous” trials were added with cues in both positions with the same brightness (responses to both sides were rewarded in this case [Fig 1III]). Also, in phases II and III, trials which required animals to perform responses to the side opposite to the one rewarded in the previous trial (termed “shift trials” here) were always followed by 1–2 trials in which the cue indicated the side rewarded in the previous trial (“stay trials”) (see Materials and methods and S1 Fig for details). As detailed below, these manipulations in phases II and III served as additional probes for the behavioral strategies or biases exhibited by the animals (except for the analyses targeting specifically these trials, however, different trial types were pooled for analysis whenever possible). They were introduced in sequential task stages and not all at once so as to not overload the early learning phase for the animals. No significant side bias was found in either group of animals on the first day of testing, nor was there a difference in side preference between groups (see Materials and methods).
First, we assessed the number of correct responses according to the experimenter-defined “cue rule” (Fig 2). Both groups of animals, Cav1.2NesCre and Cav1.2fl/fl, displayed levels of accuracy as defined by this rule that ranged significantly above chance in all 3 task phases, as confirmed by both subject-level binomial and group-level t tests (see Fig 2 legend for statistical details). This verifies that performance levels in both groups significantly increased over the course of the task, even if assessed purely in terms of the experimenter-defined cue rule. Nevertheless, a 2 × 3-factorial analysis of variance (ANOVA) with “group” (Cav1.2NesCre versus Cav1.2fl/fl) as between- and “task phase” as within-subjects factor revealed a main effect of group (F(1,21) = 20.19, p < .001), a main effect of task phase (F(2,42) = 40.62, p < .001), and an interaction effect (F(2,42) = 18.7, p < .001), indicating that performance was significantly worse for Cav1.2NesCre compared to Cav1.2fl/fl animals, in a task-phase-dependent manner, as further confirmed by post hoc tests (Fig 2B; note that for these across-phase comparisons low- and high-contrast trials were combined in phases II & III, as these were not present in phase I, see Materials and methods and below).
Next, we examined whether performance depended on whether the previous trial required a response to the same side for obtaining reward as the current one (“stay trial”) or a response to the opposite side (“shift trial”). For both shift and stay trials, we again observed a main effect of group (shift trials: F(1,21) = 6.55, p = .018, stay trials: F(1,21) = 22.59, p < .001), a main effect of task phase (shift trials: F(1.35,28.33) = 11.12, p = .001, stay trials: F(2,42) = 47.97, p < .001), and a significant phase-by-group interaction (shift trials: F(1.35,28.33) = 11.01, p = .001, stay trials: F(2,42) = 3.54, p = .038). While Cav1.2fl/fl mice essentially showed the same performance pattern across task phases in both types of trials (i.e., an increase from phase I to phase II, both post hoc tests p < .001, and no further increase in phase III, both p > .1), as they should if they acted according to the cue rule, for Cav1.2NesCre, the response pattern strongly differed on shift and stay trials: while, in fact, Cav1.2NesCre animals performed around chance level in stay trials throughout the whole first task phase before jumping to higher performance levels in phases II and III (post hoc tests phase I versus II: p = .033, phase II versus III: p = .003, see Fig 3A), they demonstrated high performance on shift trials in phase I, which then, however, dramatically declined across task phases (post hoc tests phases II versus III: p = .001, see Fig 3B). Indeed, Cav1.2NesCre even outperformed Cav1.2fl/fl animals on shift trials during phase I (post hoc tests Cav1.2fl/fl versus Cav1.2NesCre phase I: p = .044), while Cav1.2fl/fl outperformed Cav1.2NesCre in all other conditions (all p < .05).
The fact that Cav1.2NesCre animals did not exceed chance level in terms of the cue rule throughout stay trials in phase I and shift trials in phase III suggests that they may not have gathered the experimentally intended cue-response association. This interpretation is further supported by a separate analysis of trials with reduced stimulus contrast (Fig 3C): while Cav1.2fl/fl animals were significantly influenced by this manipulation, as one would expect if their behavior were controlled by the cue (low- versus high-contrast trials, phase II: t(11) = 1.82, p = .096; phase III: t(11) = 3.34, p = .007), reducing cue contrast did not affect the behavior of the Cav1.2NesCre animals (low- versus high-contrast trials, phase II: t(10) = .33, p = .75; phase III: t(10) = –.21, p = .836).
The results above raise the question of how the Cav1.2NesCre mice did manage to improve task performance, although they apparently did not issue their responses in accordance with the presented cue. In novel environments, rodents quickly adopt and learn to adapt ecologically prepared or previously acquired strategies like win-stay or win-shift [e.g. 7,10,11], which we will denote as “outcome rules” in the following, rather than extracting the (cue-based) rules experimentally imposed. To assess this, we first evaluated how consistent their pattern of responses is with any of the 4 elementary outcome-based strategies: win-stay, win-shift, lose-stay, and lose-shift. Fig 4A gives the relative frequency of responses across trials on each day that are consistent with a win-shift and a lose-shift rule for Cav1.2NesCre and Cav1.2fl/fl animals (note that by symmetry, p(win-shift) = 1–p(win-stay), and p(lose-shift) = 1–p(lose-stay)).
In general, there was a main effect of (outcome-based) strategy (F(1,21) = 78.03, p < .001) and task phase (F(1,21) = 78.03, p < .001), but the strategy depended on phase (strategy x phase: F(2,42) = 47.06, p < .001) and was further modulated by group (strategy x group: F(1,21) = 32.16, p < .001 and strategy x group x phase: F(2,42) = 3.306, p = .046): on the one hand, outcome-strategy-conforming responses appeared to change in both groups with task phase and trial history, but on the other, this adaptation was clearly different for Cav1.2NesCre and Cav1.2fl/fl animals. Specifically, both groups favored lose-shift over lose-stay, that is, tended to shift to the other side after an unsuccessful response (binomial tests on lose-shift against chance across all task phases: p≤.001 for 23/24 animals). After successful trials, however, Cav1.2NesCre mice considerably differed in their behavior from Cav1.2fl/fl: During phase I, they shifted significantly more often after wins than Cav1.2fl/fl (post hoc test: p = .001) and even marginally more so than after losses (post hoc test win-shift versus lose-shift for Cav1.2NesCre: p = .079), thus basically increasing their overall probability for shifting (Fig 4A left). Cav1.2fl/fl mice, in contrast, exhibited significantly more shifts after incorrect trials than after correct ones (post hoc test win-shift versus lose-shift for Cav1.2fl/fl: p < .001), with no preference for win-shift over win-stay (see Fig 4A right, p(shift | win) ≈ .5, t(11) = –.02, p = .98). In phases II and III, Cav1.2NesCre animals then reduced shifting after correct but not incorrect responses (post hoc tests win-shift in phase I versus II: p = .012, phase II versus III: p < .001). Although this general trend is similar to that shown by Cav1.2fl/fl (phase I versus II and phase II versus III: p < .001), Cav1.2NesCre animals still win-shifted significantly more often than Cav1.2fl/fl in phase II (p = .001) and, by trend, in phase III (p = .069). Thus, in essence, Cav1.2NesCre animals increased the overall probability for shifting during phase I and then selectively and progressively decreased their win-shift tendency in the next 2 phases (see also S2 Fig).
Note that the larger proportion of stay-correct trials in task phases II and III by task design (see above and Materials and methods) makes downregulating win-shift responses a sensible strategy. This raises the question of whether the below-chance decrease in win-shift responses observed also in Cav1.2fl/fl animals reflects (partial) adoption of a win-stay strategy or whether this decrease could partly be related to other, confounding factors, like correlations among cue and outcome rules (overlap between cue and win-stay (outcome) rule: phase I: ~47%, phase II: ~74%, and phase III: ~82%). The ambiguous trials introduced in phase III (with equal cue intensities in the top and bottom positions) might help to dissociate cue- versus outcome-based rules, since, on these trials, the cue is completely noninformative with respect to response side. However, both Cav1.2NesCre and Cav1.2fl/fl animals kept on applying a win-stay strategy in these trials (Cav1.2NesCre: 60% win-stay, t(10) = 2.80, p = .019 compared to chance; Cav1.2fl/fl: 59%, t(11) = 2.84, p = .016), although Cav1.2fl/fl animals showed this behavior significantly less often than in other phase III trials (t(11) = –6.16, p < .001), while Cav1.2NesCre animals did not (t(10) = –1.39, p = .19). Cav1.2NesCre (but not Cav1.2fl/fl) animals also exhibited a highly significant correlation between win-stay behavior on ambiguous and nonambiguous trials across days (Cav1.2NesCre: t(10) = 6.59, p < .001, Cav1.2fl/fl: t(11) = .82, p = .43, and Cav1.2NesCre versus Cav1.2fl/fl: t(21) = 2.12, p = .046), further supporting the idea that Cav1.2NesCre (in contrast to Cav1.2fl/fl) animals more generally followed an outcome rule. These observations thus suggest that, in truly ambiguous situations, Cav1.2fl/fl animals may also partly revert to a win-stay strategy. In situations like phase I, however, where win-staying bears no advantage over win-shifting, Cav1.2fl/fl (unlike Cav1.2NesCre) animals did not show any such preference but followed the more rewarding cue rule (see above).
While decreasing win-shifting is sensible in phases II and III given the task design, it does not explain why Cav1.2NesCre animals clearly adapted their outcome-based behavior over time also in phase I or why, more generally, this might be a worthwhile approach in the present task. Ultimately, one would expect animals to increase the likelihood for a certain strategy if it turned out to be more rewarding than alternatives and, in particular, compared to random responding. We therefore next examined the frequency of rewards the Cav1.2NesCre animals had actually received on responses consistent with each of the 4 outcome-based response options (Fig 4B). Different outcome strategies were indeed associated with significantly different reward probabilities (main effect strategy: F(1,10) = 457.7, p < .001), and these were further modulated by task phase (strategy x task phase interaction: F(1.29,12.86) = 545.06, p < .001). While shifting after an incorrect trial was the most rewarding strategy (post hoc comparison of lose-shift to all others in all phases: p < .001), reward probability for shifting versus staying after a correct trial really depended on the experimental phase: a win-shift strategy was more rewarding in phase I (win-shift versus win-stay: p < .001) and less rewarding in phases II and III (post hoc tests both p < .001). Moreover, employing binomial tests, win-shift-consistent responses in phase I were significantly more often associated with reward than would have been expected by chance (for all Cav1.2NesCre animals: p < .001). Thus, based on the actual outcomes the animals had experienced, a win-shift/lose-shift strategy should have been perceived as more rewarding than either chance responding or any alternative combination of outcome-based responses in phase I, while win-stay/lose-shift would have been the most effective combination in phases II and III, consistent with the pattern of responses Cav1.2NesCre animals actually displayed (Fig 4A).
The analyses in the previous sections showed that the actual response pattern in Cav1.2NesCre mice is more consistent with outcome-based rules, while that of Cav1.2fl/fl mice is more consistent with the cue rule, and that these strategies could have been perceived as rewarding by the animals based on the actual experiences they have had. To conclusively statistically demonstrate that outcome-based versus cue-based behavior indeed sufficiently explained the animal choices both across and within task phases, we next examined the animals’ behavior in terms of bootstrap distributions and formal reinforcement learning (RL) models generated from the empirical data.
An outcome rule bootstrap distribution was generated by having an “artificial agent” making choices purely based on the 4 outcome rules but with probabilities for the 4 response options (win-stay, win-shift, lose-stay, and lose-shift) derived from the animals’ actual choice frequencies on each day. Hence, this “agent” would choose to win-stay, win-shift, and so on with the exact same probabilities on each day as the actual animals but would only act upon these and ignore other behavioral options or task features, like the cue. This was contrasted with an “agent” that acted purely based on the 4 cue-related response options (“top-cue/left-response,” “top-cue/right-response,” “bottom-cue/left-response,” and “bottom-cue/right-response”), again with the probabilities for showing this behavior on each day instantiated by the animals’ actually displayed response frequencies. By running these agents for 1,000 iterations of the same basic task design and phases as used for the animals, bootstrap distributions were generated consistent with a pure outcome-based or pure cue-based strategy, applied with the same probabilities actually exhibited by the animals.
While 11/12 Cav1.2fl/fl animals escaped the outcome rule bootstrap distributions (i.e., performed significantly better) in both shift (Fig 5) and stay (S3 Fig) trials, this was only the case for 4/12 Cav1.2NesCre animals. Indeed, Cav1.2fl/fl and Cav1.2NesCre animals significantly differed with regards to their agreement with the outcome bootstraps (χ2 = 8.71, p = .003), indicating that performance in the Cav1.2NesCre group is much better explained in terms of adaptation of outcome-based strategies than is the case for Cav1.2fl/fl animals. Vice versa, when animal behavior was evaluated in terms of the cue rule bootstrap distribution, only 3/12 Cav1.2fl/fl, but 10/12 Cav1.2NesCre, animals escaped the cue rule distributions on either shift or stay trials, with, again, significant differences among the groups (χ2 = 8.22, p = .004, see S4 and S5 Figs). Thus, Cav1.2fl/fl animal behavior tended to be consistent with the cue rule but inconsistent with the outcome rules, while the reverse tended to be true for Cav1.2NesCre animals.
As a final step to prove that the observed performance progress across trials could indeed entirely be explained by adapting outcome- versus cue-based strategies, RL models were estimated from the behavioral data (see Materials and methods). In brief, these models consist of linear update rules for the values of (situation, action)-pairs (Eq 1 in Materials and methods) and a sigmoid-type choice probability function (Eq 3 in Materials and methods), which renders situation-dependent choices among response options, with probabilities based on the learned relative values of the different actions in that situation. The learning process itself was modeled by a Rescorla–Wagner (RW) rule [26], augmented by a Pearce–Hall (PH)-type “associability” mechanism [27], which regulates the learning rate in accordance with the recent history of reward prediction errors (see Materials and methods for details; a pure RW model and a simplified hybrid RW+PH model with 1 rate parameter fixed were checked as well and yielded almost identical results). Hence, these models would progress from trial to trial, sequentially gathering evidence for the different reward returns to be expected from the different response options in each environmental situation and issuing choices on each trial based on these learned values.
Two types of models were constructed, one equipped purely with the 4 elementary outcome rules (see above; termed “outcome model” in the following), ignoring the cue or any other potential strategy, and one with the 4 cue-based response options as given above (“cue model”), basing its choices solely on these. Thus, both models are described by exactly the same number of response options and parameters and differed solely in terms of what these response options would refer to. The set of 5 parameters that characterizes each of these models is comprised of a global learning rate κ, an additional time-dependent rate component regulated by a parameter η in accordance with the recent history of reward prediction errors, and 3 individual exploration parameters βk for each separate task phase kϵ{I,II,III}. While the 2 learning rate parameters κ and η determine how values for (situation, action)-pairs are updated as the animals progress through their individual trial sequences (see Eq 2 in Materials and methods), the 3 parameters βk control the steepness of the sigmoid choice function (Eq 3 in Materials and methods) and, thus, how deterministically (β→∞) or probabilistically (β→0) an animal would behave given the learned values. All parameters were estimated from the animal data by maximum likelihood, and models were compared via hierarchical Bayesian inference, yielding expected posterior probabilities for the 2 types of models (cue-based versus outcome-based) given the observed behavioral data (see Materials and methods for details).
Cav1.2NesCre animals were much better described by the RL model, which updates the 4 elementary outcome strategies, than by the cue RL model, with the expected posterior model probabilities being p¯(outcome model|data)=.71 versus p¯(cue model|data)=.29 (S6 Fig left) and an exceedance probability (i.e., the likelihood with which the posterior model probability of the outcome model exceeds that of the cue model; see Materials and methods) of p = .95 (S6 Fig right). By contrast, the behavior of the Cav1.2fl/fl animals was much better captured by the RL model updating cue-association strategies, with expected posteriors p¯(cue model|data)=.79 versus p¯(outcome model|data)=.21 and an exceedance probability for the cue over the outcome model of p = .99. This analysis confirms that an outcome model much more plausibly describes the behavior of the Cav1.2NesCre animals than a cue model, while, vice versa, the cue model agrees much better with the behavior of the Cav1.2fl/fl animals than the outcome model.
In our study, the specific Cacna1c knockout was achieved via the Cre-lox system using a Nestin promoter-driven Cre recombinase in mice double-transgenic for both Cav1.2fl/fl and Nestin-Cre (i.e., Cav1.2NesCre). However, mice carrying the Nestin-Cre transgene alone (NesCre) have previously been associated with metabolic issues [28] as well as with potential deficits in learning [29], although that same study found that the deficit was rather specific to fear conditioning (freezing responses) and did not carry over to other kinds of learning. Nevertheless, given these previous indications, we conducted a further control study with 12 NesCre and 12 wild-type (WT) mice using the exact same experimental protocol as employed before. These 2 additional controls were run on experimental phase I only, since only the first task stage is required to assess whether the animals would pick up the cue rule in principle. As done above, we then compared all groups based on percentage of correct responses, bootstrap distributions designed to assess the deviation of actual performance from the one expected under a given response rule, and RL models. All of these 3 analysis approaches consistently indicated that the additional control groups are better described in terms of the cue-based response strategy: both NesCre and WT mice show increased accuracy (according to the cue rule) at the end of phase I compared to Cav1.2NesCre mice (ANOVA main effect on groups: F(3,44) = 15.04, p < .001; both post hoc tests p < .001, see also S7 Fig) but were not different from the previous Cav1.2fl/fl control group (both p > .99). Furthermore, only 1/12 NesCre and 0/12 WT mice left the 90% confidence bands of the cue rule bootstrap distribution, i.e., most animals showed performance levels as expected under the cue rule. Chi-square-based tests confirmed that significantly more of the NesCre and WT animals adhered to the cue rule than Cav1.2NesCre knockout animals (NesCre versus Cav1.2NesCre: χ2 = 16.67, p < .001, WT versus Cav1.2NesCre: χ2 = 20.31, p < .001), with numbers comparable to those within the original control group (NesCre versus Cav1.2fl/fl: χ2 = 1.2, p = .273, WT versus Cav1.2fl/fl: χ2 = 3.43, p = .064). Conversely, in terms of the outcome-based response rules, 11/12 NesCre and 10/12 WT mice went beyond the confidence bands of the outcome rule bootstrap distribution, i.e., significantly deviated from the performance expected under the outcome rule. Again, these numbers are significantly different from those in the Cav1.2 knockout strain (NesCre versus Cav1.2NesCre: χ2 = 8.71, p = .003, WT versus Cav1.2NesCre: χ2 = 6.17, p = .013) but not from those in the Cav1.2fl/fl group (NesCre versus Cav1.2fl/fl: χ2 = 0, p = 1, WT versus Cav1.2fl/fl: χ2 = .38, p = .537; results were similar when only task phase I was considered for all animals, see Materials and methods). Lastly, our RL-model-based analysis revealed that, similar to the Cav1.2fl/fl controls, both NesCre and WT mice have a much higher posterior probability for the cue rule model (NesCre: p¯=.929, WT: p¯=.929) than for the outcome rule model (NesCre: p¯=.071, WT: p¯=.071). Thus, we conclude that the change in response strategy observed for the knockout strain is indeed specific to this group, that is, can be attributed to the ablation of Cav1.2 calcium channels and is not due to other confounding factors.
In solving a specific cognitive task, animals may follow a variety of different behavioral strategies or ecological heuristics. Here, we demonstrate, through a detailed analysis of response patterns, reward returns, and computational models statistically estimated from the animals’ actual behavior [30], that both Cav1.2NesCre and Cav1.2fl/fl mice improve performance across different phases of a cue discrimination task but do so by fundamentally different means. While Cav1.2fl/fl mice seemed to infer the correct cue-response association, Cav1.2NesCre animals learned to increase their reward returns by conditioning their responses on previous outcomes in an optimal manner. It is thus important to note that while “standard” behavioral analysis would have inferred that both strains of animals, Cav1.2fl/fl and Cav1.2NesCre, could basically learn the cue-discrimination rule, only with the Cav1.2NesCre animals being significantly worse than the Cav1.2fl/fl, here we arrived at a quite different conclusion: in fact, our results suggest that Cav1.2NesCre animals did not learn the cue rule at all but rather based their behavior on adapting another reward-increasing strategy. We could also demonstrate that this alteration in behavioral strategy is specific to the Cav1.2 ablation and cannot be attributed to the expression of Nestin-promoter driven Cre per se. That Cav1.2NesCre mice were still able to adapt their responses based on observed outcomes also demonstrates that Cav1.2NesCre mice do not suffer from a general learning deficit but perhaps one that is domain-specific or more closely tied to other cognitive capacities, like working memory [31].
In a real ecological context, there is a multitude of different factors and contextual conditions that will or could influence whether a behavioral plan will ultimately lead to success. When an animal hits on a favorable outcome it had not expected, it is often difficult to discern which of the many foregoing sensory inputs and actions were indeed crucial for predicting and achieving this success and which are irrelevant and would better be ignored; also termed the “credit assignment problem” [32]. To parse out the crucial and predictive factors from limited experience, the small sample of reality an animal has access to in its lifetime, would be difficult enough if the world were rather deterministic. But on top of that, the world is highly uncertain and probabilistic, full of unforeseen events that complicate inference on environmental structure. To deal with this, animals have evolved or acquired a larger repertoire of general strategies, heuristics, shortcuts, and response preferences [5,6] that could guide them through novel situations by biasing probability distributions toward certain subsets of situation/action/outcome triples and, potentially, are also designed to minimize an animal’s cognitive effort [33].
All these different response strategies and biases almost certainly come into play in any novel laboratory situation an animal faces. Animals are unlikely to behave completely randomly but, according to recent evidence, might actually actively probe and test out different such strategies and behavioral “hypotheses” whilst learning [12,13]. In rodents, common a priori strategies are win-stay, win-shift, lose-shift, or alternate [7,9,10], which all could make sense in one environmental setting or another [8,11,34]. For instance, rats hoard food in their underground burrows—once the food is gone from one of the chambers or, more generally, a food source is depleted, it is certainly reasonable to win-shift. In other circumstances, food sources may be expected to refill (perhaps in certain temporal intervals), such that a win-stay strategy would be more appropriate. Hence, it seems reasonable that even the Cav1.2fl/fl animals in our task reverted to one of these more basic strategies, namely win-stay, when the cue information became ambiguous on certain trials. In fact, some of the Cav1.2fl/fl animals that did not show as high performance levels as their group mates appeared to have adopted an outcome-based strategy as well, further illustrating that this is not a behavioral pattern “outside the normal scope” but a sensible, although suboptimal, way to approach this task.
Often, from an experimenter’s point of view, it might in fact not be that easy to construct experimental situations that clearly dissociate what the experimenter intended the animals to do from one of these basic, a priori strategies animals bring into the task. For instance, to probe working memory in rodents, delayed response or alternation tasks on a T-maze are often used. However, some studies have suggested that, rather than using working or short-term memory to encode correct responses as intended by the experimental design, rodents may rely on subtle external bodily cues like head orientation maintained throughout the delay phase of the task [14,15]. Such behavioral alternatives or confounds with other previously acquired or prepared strategies may easily be overlooked or, indeed, sometimes hard to avoid. There could also be other hidden dependencies or predictable structure in a task design (especially in learning tasks that evolve across trials) that are not immediately obvious and that animals may attempt to exploit. Even if overall (given unlimited experience) everything is well balanced and controlled by design, the individual history of trial drawings, responses, and outcomes may be locally highly nonrandom and bias animals differentially toward certain strategies rather than others [35, see also 36,37]. Thus, consideration of different possibly rewarding behavioral strategies may not only provide a lot more mechanistic insight by revealing the details of how an animal solves the task posed but may, in fact, prove important for drawing the right conclusions with respect to the involved cognitive and behavioral capacities.
For these reasons, it is also important that the behavioral analysis takes the animals’ perspective by taking into account the exact same set or sequence of trial types as encountered by the animals in order to accurately infer what they could have possibly learned about different behavioral policies given their individual choice and reward histories [35]. In fact, it is a particular advantage of the bootstrap- and model-based analyses reported here that they adopt the animal’s point of view, incorporating the exact same sequence of stimulus events, behavioral responses, and reward returns as experienced by the animals, and thus yield day-by-day (or even trial-to-trial) predictions on an individual animal basis (e.g., Fig 5). Further note that the next trial type encountered by an animal is often a consequence of the animal’s own behavior, as in the case of correction trials, hence a reflection of the animal’s own choices that should be acknowledged in an analysis of behavioral strategies. With regards to the specific experiments conducted here, we therefore would like to emphasize that none of the analyses made any a priori assumptions about the distribution or occurrences of trial types but went exactly with those trials as empirically observed and that, of course, the experimental conditions for all animal strains were exactly the same (and hence all differences in experienced trial sequences a consequence of differences in the animals’ behavior). It may also be worth noting that the animals showed strategy-consistent responses across all 3 task stages, although these considerably differed in the composition of trials and trial transitions they harbored.
Our RL model analysis was based on a Rescorla–Wagner rule augmented by a Pearce–Hall-type mechanism [27,38], in which the learning rate is regulated by the recent history of successful outcome predictions, increasing as the mismatches between predicted and experienced rewards become more numerous (and thus indicating to the animal that its current response policies need to be altered). Although the results using this type of learning rule were completely consistent with those obtained with a pure Rescorla–Wagner rule (as well as with a simplified hybrid model; see Materials and methods), we adopted this learning mechanism, as it turned out to yield the best description of the behavioral data based on Bayesian model comparisons (see S8 Fig). Animal learning has also variously been described in terms of stimulus-response versus action-outcome associations [39,40]. In light of this discussion, it may therefore be important to emphasize again that all animal strains studied here did indeed use the outcomes to regulate and advance their behavioral policies, i.e., were all able to utilize the reward feedback in adjusting their behavior in a profitable manner. The specific difference between the animal strains was that the Cav1.2NesCre mice did not base their responses on the presented cue but on where the reward occurred on the previous trial (i.e., on the previous outcome), and the bootstrap- and model-based analyses, which specify the assumptions about the behavioral process in formal detail, confirmed that this provided a sufficient explanation for the animals’ behavior.
Calcium (Ca2+) influx through N-methyl-D-aspartate (NMDA) and voltage-gated calcium channels is crucial for synaptic plasticity [24] and nuclear gene expression associated with plasticity [41] and cell survival [42]. Although there are several other sources of cellular Ca2+ influx, like NMDA or T- or N-type Ca2+ channels [43,44], a cortex-wide, functional ablation of L-type Ca2+ channels through a knockout of the Cacna1c gene is therefore supposed to impair plasticity, and thus learning, in a variety of settings. Indeed, this selective knockout has been associated with deficits in spatial memory [45,46] and observational fear learning [47]. Although the cortex-wide lack of Cav1.2 in the present preparation prevents a more precise localization of the effects reported here, a recent study employing a similar touchscreen-based, two-choice stimulus discrimination task in mice revealed a specific dependence on dorsal striatum [48], in line with earlier studies linking this region to the mediation of SR associations [39,49,50]. Interestingly, this study also involved correction trials, which may facilitate outcome-based strategies, and found that performance scores in the lesioned animals settled around similar levels as observed for the Cav1.2NesCre mice here, about 60%–65%. Although outcome-based strategies were not investigated by Delotterie and colleagues [48], these findings could indicate that the present effects may be more specifically rooted in L-type Ca2+ channel dysfunction in dorsal striatum.
Findings like these may also open new perspectives on treatment options in psychiatry. Specifically, the CACNA1C gene has been associated with severe psychiatric illnesses such as depression [51], bipolar disorder [52], autism [53], and schizophrenia [51,54], for which it has been listed as a risk gene [51,55]. The present results suggest that for modifying a patient’s behavior, certain avenues, potentially resting on biologically more elementary behavioral options, may be fruitful to explore where other forms of behavioral therapy have failed. They at least suggest that a detailed scrutiny of how a patient solves a specific set of cognitive or emotional tasks, on top of standard clinical assessment, may provide valuable insights into how to best address the “behavioral malfunctioning” of a patient in terms of the specific behavioral interventions and rules the therapy should focus on (see also [56]). Hence, in future applications, the combination of standard cognitive test batteries with model-estimation techniques, as used here, could lead into the design of more specifically targeted, individualized behavioral therapies, on top of its potential use for diagnosis and prediction.
All experiments complied with regulations covering animal experimentation within the EU (European Communities Council Directive 2010/63/EU), and were approved by German animal welfare authorities (Regierungspraesidium Karlsruhe: ethical approval no. 35-9185-81-G-227-12).
All mice were bred in the animal facility of the Central Institute of Mental Health, Mannheim, Germany, and maintained on a C57BL/6N background. CNS-specific ablation of the L-type voltage gated calcium channel Cav1.2 was achieved by inactivation of the Cacna1c gene using the cre-loxP system. More specifically, Cav1.2NesCre mice (genotype: Cav1.2 L2/L2, Nestin-Cre -/+) were generated by crossbreeding mice carrying 2 loxP-flanked (“floxed”) Cav1.2 alleles [45] and mice with an additional heterozygous transgene expressing Cre recombinase under the control of a Nestin promoter [57]. In these animals, the loss of Cav1.2 functionality is accomplished by the excision of Cacna1c exons coding for the IIS5 and IIS6 transmembrane segments of the pore-forming subunit α1C of the Cacna1c gene via Cre recombinase that is expressed in all cell types of the CNS through embryonic development. Crossbred Cav1.2fl/fl mice homozygous for the floxed Cav1.2 allele but lacking Cre recombinase expression (genotype: Cav1.2 L2/L2, Nestin-Cre -/-), and thus not suffering calcium channel loss, served as primary controls within the experiment. To rule out that observed learning differences between these 2 groups were not attributable to metabolic [28] or other potential factors [29] associated with the expression of Cre recombinase in the Cav1.2NesCre group, we assessed performance of 2 additional control groups: animals with a heterozygous transgene expressing Cre recombinase under the control of a Nestin promoter without loxP-flanked Cav1.2 alleles (NesCre; genotype: Cav1.2 +/+, Nestin-Cre +/-) and pure wild-type animals (WT; genotype: Cav1.2 +/+, Nestin-Cre -/-).
In total, 12 male Cav1.2NesCre, 12 Cav1.2fl/fl, 12 NesCre, and 12 WT mice were single-housed in conventional macrolon cages (Type II, 26 × 20 × 14 cm) with sawdust (RehofixMK-2000; Rettenmaier & Söhne, Rosenberg, Germany), nesting material, and free access to water. Single housing was chosen as it has been shown to be less stressful for male mice than group housing under standard maintenance conditions (i.e., no enrichment) [58]. All animals were approximately 12 weeks old at experimental onset. Colony room settings included a temperature of 23±2°C, relative humidity 50%± 5%, and a reversed 12 hour light–dark schedule with the lights off at 7:00 AM [58,59]. Experiments were conducted during the dark phase, which constitutes the active phase of mice.
Mice were food restricted [60] to 85%–90% of their initial free-feeding body weight in order to maintain a high degree of motivation during operant training. The mean initial body weight was assessed on 5 consecutive days when animals had free access to food. Body weight and health status were monitored daily prior to testing. At the start of food restriction, mice were food deprived for 1 light phase and received 2.0 g to 3.3 g food in the subsequent dark phase, depending on individual loss. Henceforth, the amount of food was adjusted in accordance with the deviation from the intended 87.5% of initial body weight. Touchscreen-trained mice additionally received 7 μl sweet condensed milk (SCM; Milchmaedchen, Nestle, diluted in 1:4 tap water) as a reward for correct responses during training. They were previously acquainted to SCM in their home cages to avoid later refusal as a reward. In order to minimize handling as a source of anxiety and to reduce stress, all mice were handled without physical restraint following the cup handling protocol described in [61]. Mice were scooped up and allowed to walk freely over the handler’s open hands.
Mouse touchscreen chambers (Model 80614–20, Campden Instruments Ltd., Loughborough, Leics., United Kingdom) included several infrared (IR) light beams for movement detection, a 3 W house light for controllable illumination, and a tone generator for auditory signaling. The inner chamber consisted of black Perspex walls arranged in a trapezoid shape (height 19 cm, width 24 respectively 6 cm, depth 17 cm) and a metal grid floor. The longer end of the chamber was equipped with a touch-sensitive screen (IR-detector-based) partly covered by a 3-hole Perspex mask in order to separate the display into three equal response windows (7 x 7 cm). The lateral fields were used to detect touches, while the central field was used to display the reward-side-indicating cue. Reward cues consisted of bright gray 2.5 cm squares (brightness = 70%, 85%, or 96%, hue and saturation = 0 in HSB colors). Correct responses triggered the display of a tone, illumination in the reward tray, and delivery of 7 μl sweet condensed milk. The reward tray (2 x 2 x 2 cm) of an externally placed feeder for liquid suspensions positioned at the touchscreen-opposed end of the chamber contained another light beam detector used to count entries, start and stop latencies, and trigger initiation of the next trial. Hence, the subject was required to initiate a trial at the narrow end of the chamber, subsequently traverse to the opposed touchscreen side for response, and, finally, traverse back in order to retrieve the reward. This ensured optimal prospect to the screen and constant start conditions on each trial.
|
10.1371/journal.pcbi.1000213 | Transmembrane Topology and Signal Peptide Prediction Using Dynamic Bayesian Networks | Hidden Markov models (HMMs) have been successfully applied to the tasks of transmembrane protein topology prediction and signal peptide prediction. In this paper we expand upon this work by making use of the more powerful class of dynamic Bayesian networks (DBNs). Our model, Philius, is inspired by a previously published HMM, Phobius, and combines a signal peptide submodel with a transmembrane submodel. We introduce a two-stage DBN decoder that combines the power of posterior decoding with the grammar constraints of Viterbi-style decoding. Philius also provides protein type, segment, and topology confidence metrics to aid in the interpretation of the predictions. We report a relative improvement of 13% over Phobius in full-topology prediction accuracy on transmembrane proteins, and a sensitivity and specificity of 0.96 in detecting signal peptides. We also show that our confidence metrics correlate well with the observed precision. In addition, we have made predictions on all 6.3 million proteins in the Yeast Resource Center (YRC) database. This large-scale study provides an overall picture of the relative numbers of proteins that include a signal-peptide and/or one or more transmembrane segments as well as a valuable resource for the scientific community. All DBNs are implemented using the Graphical Models Toolkit. Source code for the models described here is available at http://noble.gs.washington.edu/proj/philius. A Philius Web server is available at http://www.yeastrc.org/philius, and the predictions on the YRC database are available at http://www.yeastrc.org/pdr.
| Transmembrane proteins control the flow of information and substances into and out of the cell and are involved in a broad range of biological processes. Their interfacing role makes them rewarding drug targets, and it is estimated that more than 50% of recently launched drugs target membrane proteins. However, experimentally determining the three-dimensional structure of a transmembrane protein is still a difficult task, and few of the currently known tertiary structures are of transmembrane proteins despite the fact that as many as one quarter of the proteins in a given organism are transmembrane proteins. Computational methods for predicting the basic topology of a transmembrane protein are therefore of great interest, and these methods must be able to distinguish between mature, membrane-spanning proteins and proteins that, when first synthesized, contain an N-terminal membrane-spanning signal peptide. In this work, we present Philius, a new computational approach that outperforms previous methods in simultaneously detecting signal peptides and correctly predicting the topology of transmembrane proteins. Philius also supplies a set of confidence scores with each prediction. A Philius Web server is available to the public as well as precomputed predictions for over six million proteins in the Yeast Resource Center database.
| The structure of a protein determines its function. Knowledge of the structure can therefore be used to guide the design of drugs, to improve the interpretation of other information such as the locations of mutations, and to identify remote protein homologs.
Indirect methods such as X-ray crystallography and nuclear magnetic resonance spectroscopy are required to determine the tertiary structure of a protein. Membrane proteins are essential to a variety of processes including small-molecule transport and signaling, and are of significant biological interest. However, they are not easily amenable to existing crystallization methods, and even though some of the most difficult problems in this area have been overcome in recent years, the number of known tertiary structures of membrane structures remains very low. Computational methods that can accurately predict the basic topology of transmembrane proteins from easily available information therefore continue to be of great interest. To be most valuable, a predicted topology include not only the locations of the membrane-spanning segments, but should also correctly localize the N- and C-termini relative to the membrane.
Many proteins include a short N-terminal signal peptide that initially directs the post-translational transport of the protein across the membrane and is subsequently cleaved off after transport. A signal peptide includes a strongly hydrophobic segment which is not a part of the mature protein but is often misclassified as a membrane-spanning portion of a transmembrane protein. Conversely, a transmembrane protein with a membrane-spanning segment near the N-terminus is often misclassified as having a signal peptide. Therefore, signal peptide prediction and transmembrane topology prediction should be performed simultaneously, rather than being treated as two separate tasks.
Membrane proteins are classically divided into two structural classes: those which traverse the membrane using an α-helical bundle, such as bacteriorhodopsin, and those which use a β-barrel, such as porin. The β-barrel motif is found only in a small fraction of all membrane proteins (e.g., in the outer membrane of Gram negative bacteria and in the mitochondrial membrane). Lately, some attention has been given to some irregular structures such as re-entrant loops and random coil regions. In this work, however, we focus on the α-helical class, both because most membrane proteins fall into this class, and because they constitute most of the known 3D structures.
The two most common machine learning approaches applied to the prediction of both signal peptides and the topology of transmembrane proteins are hidden Markov models (HMM) and artificial neural networks (ANN), while some predictors use a combination of these two approaches. HMMs are particularly well suited to sequence labeling tasks, and task-specific prior knowledge can be encoded into the structure of the HMM, while ANNs can learn to make classification decisions based on hundreds of inputs.
The first HMM-based transmembrane protein topology predictors were introduced ten years ago: TMHMM [1] and HMMTOP [2]. Both of these predictors define a set of structural classes which capture the variation in amino acid composition of different portions of the membrane protein. For example, the membrane-spanning helix is known to be highly hydrophobic, and cytoplasmic loops generally contain more positively charged amino acids than non-cytoplasmic loops (the so-called positive-inside rule). During training the HMM learns a set of emission distributions, one for each of the structural classes. TMHMM is trained using a two-pass discriminative training approach followed by decoding using the one-best algorithm [3]. HMMTOP introduced the hypothesis that the difference between the amino acid distributions in the various structural classes is the main driving force in determining the final protein topology, and that therefore the most likely topology is the one that maximizes this difference for a given protein. HMMTOP [4] was also the first to allow constrained decoding to incorporate additional evidence regarding the localization of one or more positions within the protein sequence. The presence of a signal peptide within a given protein has also been successfully predicted using both HMMs [5] and ANNs [6].
As mentioned above, the confusion between signal peptides and transmembrane segments is one of the largest sources of error both for conventional transmembrane topology predictors and signal peptide predictors [7],[8]. Motivated by this difficulty, the HMM Phobius [9] was designed to combine the signal peptide model of SignalP-HMM [5] with the transmembrane topology model of TMHMM [1]. The authors showed that including a signal peptide sub-model improves overall accuracy in detecting and differentiating proteins with signal peptides and proteins with transmembrane segments.
In this work, we introduce Philius, a combined transmembrane topology and signal peptide predictor that extends Phobius by exploiting the power of dynamic Bayesian networks (DBN). The application of DBNs to this task provides several advantages, specifically: (a) a new two-stage decoding procedure, (b) a new way of expressing non-geometric duration distributions, and (c) a new approach to expressing label uncertainty during training. Philius is inspired by Phobius and tackles the problem of discriminating among four basic types of proteins: globular (G), globular with a signal peptide (SP+G), transmembrane (TM), and transmembrane with a signal peptide (SP+TM). Philius also predicts the location of the signal peptide cleavage site and the complete topology for membrane proteins.
We report state-of-the-art results on the discrimination task and improvements over Phobius on the topology prediction task. We also introduce a set of confidence measures at three different levels: at the level of protein type, at the level of the individual topology segment (e.g., inside, membrane, outside), and at the level of the full topology. Confidence measures for topology predictions were introduced by Melén et al. [10], and we expand upon this work with these three types of scores that correlate well with the observed precision.
Finally, based on the Philius predictions on the entire Yeast Resource Center [11] protein database, we provide an overview of the relative percentages of different types of proteins in different organisms as well as the composition of the class of membrane proteins.
Transmembrane protein topology prediction can be stated as a supervised learning problem over amino acid sequences. The training set consists of pairs of sequences of the form (o,s) where o = o1,…,on is the sequence of amino acids for a protein of known topology, and s = s1,…,sn is the corresponding sequence of labels. The oi are drawn from the alphabet of 20 amino acids , and the si are drawn from the alphabet of topology labels, , corresponding respectively to cytoplasmic (“inside”) loops, membrane-spanning segments, non-cytoplasmic (“outside”) loops, and signal peptides. After training, a learned model with parameters Θ takes as input a single amino acid test sequence o and seeks to predict the ‘best’ corresponding label sequence s* (with no unknowns).
We solve this problem using a DBN, which we call Philius. Before describing the details of our model, we first review HMMs and explain how they are a simple form of DBN. The generality of the DBN framework provides significantly expanded flexibility relative to HMMs, as described in [12]. A recently published primer [13] provides an introduction to probabilistic inference using Bayesian networks for a variety of applications in computational biology.
HMMs are conceptually simple and yet also almost unlimited in their flexibility [14]. An HMM is a generative model in which an observed sequence is generated according to an underlying but unknown sequence of states. More precisely, an HMM is a joint probability distribution over a set of 2N variables: the N observations o, and the N hidden states, s. The HMM assumes that the joint distribution over these 2N variables can be factorized as follows:(1)where s = {s1,…,sN}, o = {o1,…,oN}, , and where i represents position along the observed sequence. An HMM is often used to compute the probability distribution over the observations Pr[o] by summing (or marginalizing) over all possible hidden state sequences s in the above joint distribution. An HMM might also be used as a means to infer a most probable sequence of states from an input sequence of observations. The factorization property of an HMM makes these sorts of computations (collectively called statistical inference) based on an HMM tractable, and has been one of the keys to their widespread success.
The two conditional relationships that define an HMM are generally constant with respect to the position i. An HMM such as this is referred to as a time-homogeneous model (since the parameters are homogeneous with respect to time). This time-homogeneity allows the HMM to represent sequences of states and observations of arbitrary length N with a fixed and finite number of parameters. Most HMMs and dynamic Bayesian networks are time-homogeneous.
It is perhaps most common in the literature to represent an HMM using a state transition graph in which each node is a state in the model, and directed edges between pairs of nodes show the allowed (non-zero probability) transitions between states. Such a graph shows only the allowable state transitions–nothing in this graph describes the observation distributions Pr[oi|si] nor is anything stated about the HMM joint distribution and the factorization properties mentioned in Equation 1.
In many applications and publications using HMMs, the HMM state transition diagram may be the only descriptive graphic provided. In our research, we often use in addition a quite different graphical description of an HMM, one that depicts a very different set of HMM properties. As mentioned above, Equation 1 makes explicit the factorization properties of an HMM, and these properties allow for efficient inference on the HMM. We can use a type of graph known as a Bayesian network (BN) to visually and precisely convey this set of properties, as is done in Figure 1. Figure 1a shows the “static” relationship between a state variable and the associated observation at a single point i corresponding to the factor Pr[oi|si] in Equation 1. Figure 1b shows the graph for the expanded HMM corresponding to Equation 1, which includes a node for each state and observation variable for all time-points i = 1,…,N. This figure makes clear the dynamic aspect of the model, i.e., Pr[si|si−1] and Pr[oi|si] for all i. A Bayesian network (BN) is one type of graphical model in which edges are directed, and in which directed cycles are not allowed [15].
A frame (often also referred to as a slice or time-slice) in an HMM corresponds to one vertical section, corresponding to a single time point i. For example, in order to model a protein of length N, we could use an HMM that consists of N frames, where each amino acid has its own local copy of the basic HMM template. In an HMM, this slice contains only two random variables. We refer to the first and last frames as the prologue and epilogue of the model respectively, and to each frame in between as a chunk. In order to create an HMM of length N, the chunk is replicated N−2 times, a process sometimes referred to as unrolling. The prologue and epilogue often differ slightly from the chunk, allowing for distinct modeling at the extreme ends of the sequence. In the BN representation, we follow the convention that shaded nodes represent observations (also collectively referred to as the evidence), while unshaded nodes represent hidden variables. The chain of hidden variables is where the HMM gets its name–there is a presumed underlying set of hidden variables that form a (first order) Markov chain.
The BN representation of an HMM illustrates the minimum factorization properties required of a joint probability distributions that fits the model. More generally, the use of the term graphical model [16], implies that there is a graph (a set of nodes and edges) in which nodes correspond to random variables, and edges encode in a mathematically precise way the set of conditional independence (or factorization) properties of any probability distribution over those random variables which can be represented by the graph.
Dynamic Bayesian networks (DBNs) are BNs that extend over time (or some other dimension such as genomic or protein sequence position). DBNs are strict generalizations of both HMMs and BNs and are constructed in much the same way: by concatenating identical (except possibly the first and last) copies of a “static” BN and linking the adjacent BN copies together in some consistent way. The same advantage of being able to model sequences of essentially unbounded length using a finite number of parameters that gives the HMM much of its power carries over naturally to the DBN. In fact, any HMM is an instance of a DBN—Figure 1a shows the static BN which when repeated over and over gives us the DBN description of an HMM in Figure 1b. The converse, that any DBN is an instance of an HMM, is however not true.
Philius's state transition diagram is shown in Figure 3. The model includes three basic regions–cytoplasmic, membrane, and non-cytoplasmic–each containing multiple states and representing one or more different topology labels. At this level of description, Philius exactly mimics Phobius. In the Phobius HMM, the states shown in Figure 3 are implemented as collections of HMM states, with transitions defined to produce the desired segment duration distributions. In Philius, by contrast, the duration modeling is explicit.
For the typical HMM as in Figure 1b, a state transition diagram along with the transition probabilities and emission distributions is sufficient to completely describe the model. The same DBN is used in training and decoding, the only difference being that the states are observed during (supervised) training and hidden during testing. With DBNs, it is common to use different graph topologies for training and decoding. Philius uses three different graphs, shown in Figure 2.
The training DBN shown in Figure 2a addresses the duration and labeling issues described earlier. The Markov chain backbone over the state nodes is the same as in a typical HMM, and the relationship between statei and statei−1 is defined by the usual state transition matrix, Pr[si|si−1], represented in the state transition diagram shown in Figure 3. Beyond the backbone, this DBN differs significantly from the standard HMM. Within each frame, the state node is related to three other random variables: the durationClass , the emissionClass, and the topoLabel. The first two are hidden variables, but in both cases the relationship to the state is a deterministic mapping that does not impact the computational complexity. The mapping from state to durationClass reflects which states share similar duration properties. Similarly, the mapping from state to emissionClass reflects which states share similar emission distributions. The emissionClass node is the one that ‘emits’ the amino acid according to the appropriate distribution. The amino acid is observed during training and during the first decoding stage.
The relationship between state and topoLabel is enforced using an observed child mechanism [19], i.e., the value of state is constrained by the observed value of topoLabel. There can be a many-to-one relationship between the state and the topoLabel: one value of topoLabel, such as inside, allows the state variable to take on several different values, while another label, such as cleavage site constrains the state variable to a single value. This approach is more flexible than the class-HMM described by Krogh in [23] in which each state emits a (class, observation) pair.
As previously described, the wildcard label places no restrictions on the current state, while the sequence of states remains constrained by the allowed state transitions and state durations, thereby preserving the grammar. Even with fully labeled training data, there is some uncertainty in the locations of the boundaries between adjacent segments. To account for this uncertainty and to allow the model more flexibility during training, we remove up to five labels on either side of every boundary (while keeping at least one label per segment), and replace these labels with the wildcard label. During training the model will adjust the location of the boundary in order to maximize the probability of each training example given the model parameters. Other researchers have addressed this issue with a two-stage training procedure in which an initial model is trained and then used to relabel the training data, before the final model is trained. This type of two-stage training approach may result in a final model that is overly dependent on the decisions made by the initial model. Our wildcard label approach allows us to train the model in a single pass, maintaining the expression of uncertainty regarding the labels, and can also be used in a semi-supervised setting, combining partially-labeled data with fully-labeled data.
The duration modeling for each duration class is handled by the stateCountDown and changeState nodes. Three basic types of duration models are allowed: (i) fixed and finite durations; (ii) random and finite durations; and (iii) geometric (possibly infinite) durations. The first two types are defined using a CPT Pr[D = d|Cv], representing the probability that the duration of the current segment D is equal to d, conditioned on the duration class Cv. The dimensions of this table are Dmax by |Cv|, where Dmax is the maximum finite duration and |Cv| is the number of different duration classes to be learned. When a transition to a new (different) state occurs, a randomly chosen duration is used to initialize the stateCountDown node. This value is decremented in each successive frame until it reaches a value of 1 whereupon the changeState node is set to true and a state transition is triggered in the next frame. The states with a geometric duration distribution are handled using a slightly different mechanism. For these states, the stateCountDown node is assigned the value of 0, which is not decremented in the subsequent frame. Instead, the binary changeState node is set randomly to true or false based on the self-looping probability p for the appropriate duration class.
The model is trained on labeled data (with wildcards as described above) using the EM algorithm. The free parameters learned during training consist of the start state probabilities, the transition probabilities for the few states that have more than one allowed next-state, the emission distributions for each emission class, the duration distributions for the finite duration classes, and the self-looping probabilities for the geometric duration classes, for a total of 388 free parameters. (There are 6 possible start-states, 4 states with more than one possible next state, 15 different emission classes, 87 finite-duration model parameters and 6 geometric-duration model parameters.) The emission class probabilities were smoothed by adding a single pseudo-count to each of the accumulated counts during training. Although the EM algorithm is only guaranteed to converge to a local maximum, in this case the uncertainties during training are only related to the exact placement of the segment boundaries and we found that repeated EM training runs did not result in significantly different parameters (data not shown).
The Viterbi algorithm is commonly used to find the most likely sequence of hidden states in an HMM given the observations and the model parameters. For a DBN, a generalized version of the Viterbi algorithm similarly finds the single most likely assignment to the set of all hidden variables h = [h1,…,hH] given the evidence variables e = [e1,…,eE] and the model parameters Θ:
In this application, however, we are interested in finding the most likely sequence of labels λ*, where the variables in λ form a subset of h, but the best partial assignment λ* is not necessarily contained in the best overall assignment h*. Computing λ* is intractable in general [24], because it requires first that we compute the probabilities of all possible assignments and then sum over all assignments that correspond to each possible sequence of labels. In order to estimate the most likely sequence of labels, we have developed a novel two-stage approach. In the first stage, we compute the posterior probabilities for each λ by marginalizing out all other hidden variables. Defining a sequence of labels λ directly based on these posterior probabilities may produce a sequence that does not obey the grammar of the underlying model. Instead, we use the posterior probabilities on the labels in a second stage to influence the choice of the ‘best’ assignment h*, while enforcing the same grammar defined by the state transition matrix. Each of the two decoding stages uses a different graph than the one used in training, and these graphs are shown in Figure 3b and 3c.
This two-stage decoding is similar to the posterior Viterbi algorithm described in [25] and applied to predicting the topology of β-barrel membrane proteins, and is also similar to the optimal accuracy decoding used in [26] to combine information from homologous proteins. Both of these approaches use Viterbi-like algorithms to find the permissible sequence of states that maximizes some function of the posterior state probabilities. Here, we are effectively finding the permissible sequence of states that maximizes the product of the posterior label probabilities, subject to the topology grammar. By using DBNs combined with virtual evidence, there is no need to construct special-purpose inference algorithms; the only changes are in the definition of the topology of the graphical model and in the incorporation of the virtual evidence.
In the first stage decoding DBN, shown in Figure 2b, the observed topoLabel in the training graph is removed and replaced with a hidden topoState which is dependent on the current state and the previous topoState, and combines both the current topology label () and whether or not the label has just changed (i.e., a new segment has been started). Incorporating this change-of-label information was found to significantly improve the precise localization of the signal peptide cleavage site. In addition, a new summary variable, pType , has been added which takes on one of four values in {G, SP+G, TM, SP+TM}, representing the four basic protein types. The pType node keeps track of whether or not a particular state assignment includes a signal peptide, and whether or not it includes a (non-SP) transmembrane segment. This is done by initializing pType = G and then or-ing together the pType from the previous frame with information from the current topoState to determine the pType up to and including the current frame. Full inference is performed on this graph to compute the posterior probabilities of all nodes given the evidence (the amino acid sequence) and the model parameters. Specifically, this first stage of the decoding produces as output the posterior probabilities for the topoState variable in each frame as well as the posterior probabilities for pType in the final (right-most) frame. Note that these posterior probabilities on the final protein type node should not be confused with a posterior probability on the location of the C-terminus of the protein; for each type in {G, SP+G, TM, SP+TM}, it represents the total probability, after all other hidden variables have been marginalized out, that the test protein is of that type.
The second stage decoding DBN, shown in Figure 2c, is significantly simpler than the other two graphs: the amino acid evidence has been removed along with the emissionClass node, as has the entire segment duration portion of the graph. In order to incorporate the information from the first stage, a new observed child node topoVE has been added in each frame. The parent of this new node is the topoState node, and the conditional relationship is defined, in a position- inhomogeneous manner, based on the posterior label probability computed in the first stage:Because the posterior probabilities already include the effects of the transition, emission and duration probabilities, these no longer need to be included in the second stage. The output of the second stage of the decoder is the topology resulting from the Viterbi assignment to the hidden variables in Figure 2c. The Viterbi topology λv is now much closer to the optimal solution λ* because of the inclusion of the posterior probabilities from the first stage.
Experimental information can also be easily incorporated into this decoding process. For example, if the protein type is known, then the final pType node can be constrained to match. If other information is known, such as the location of the C-terminus or details regarding particular membrane-spanning segments, this too can be easily incorporated as additional evidence constraining the topoState nodes in those frames where the evidence exists.
In the Results section, we describe three types of confidence scores: protein type, per-segment, and topology. The first score reflects Philius's confidence in the assignment of the protein type–G, SP+G, TM or SP+TM. The protein type score is computed using the posterior probabilities for the pType variable in the final frame of the first stage decoding DBN. This computation produces a single set of probabilities Pr[y] for each evaluated protein. The second stage of the decoder produces the topology prediction and the predicted protein type yˆ. The confidence score associated with the protein type prediction is the posterior probability Pr[yˆ]. The second type of score is the per-segment score, which represents an estimate of the accuracy of the label and boundaries of a particular segment. For this score, we use the Viterbi segmentation from the second stage and compute the arithmetic mean of the first stage posterior probabilities within that segment for the Viterbi-assigned topology label. The third score applies only to transmembrane proteins and reflects Philius's confidence in the overall predicted topology. We define this score as the minimum segment score over all predicted membrane segments and the N-terminal and C-terminal segments.
We used the Phobius dataset [9] during model development. This dataset consists of four non-overlapping subsets of 1087 globular (G) proteins, 1275 globular proteins with signal peptides (SP+G), 247 transmembrane (TM) proteins and 45 transmembrane proteins with signal peptides (SP+TM). The maximum homology among the 247 TM proteins is 80%, and the maximum homology among the 45 SP+TM proteins is 35%. The same cross-validation folds and the same labels that were used to train and test Phobius were also used in this work.
Two additional datasets were obtained and used in the final testing and evaluation of the model: the SCAMPI dataset [27] of 124 transmembrane proteins (http://octopus.cbr.su.se/index.php?about=download) and the SignalP 3.0 [28] training dataset. The labels in the SCAMPI dataset include re-entrant regions which do not completely span the membrane. These were removed and relabeled as inside or outside because Philius does not currently model those types of segments. The maximum homology among these 124 proteins is 40%. Based on homology between these and the original Phobius TM proteins, this set was divided into one set of 77 proteins that does not overlap the Phobius dataset (maximum homology 80%), and one set of 47 proteins that does. For the purposes of training and testing Philius we only used the signal peptide portion of the SignalP dataset, combining the eukaryotic and bacterial proteins into a single set of 1728 proteins. Truncated versions of these proteins were used in training because the labels covered only the signal peptide and cleavage-site of each protein.
We evaluated the performance of Philius on the development dataset using ten-fold cross-validation. We measured the performance of the model as well as the accuracy of all three types of confidence scores. For proteins containing a signal peptide, we also considered the accuracy with which the cleavage site is localized.
We chose to compare our method to Phobius because it is the only method that we know of that simultaneously predicts signal peptides and complete transmembrane topologies. Several methods, such as MemBrain [29] and Proteus [30], predict transmembrane helices and signal peptides, but without any topological (inside/outside) information. The web server PONGO [31] gives predictions from individual transmembrane topology and signal peptide predictors without combining the individual predictors.
Initially, we evaluate how accurately Philius identifies a given protein's class as G, SP+G, TM or SP+TM. Table 1 shows the performance of Phobius and Philius at this task using accuracy, precision, sensitivity, specificity and Matthews correlation coefficient as metrics. Note that, because the SP+TM subset consists of only 45 examples, fewer than 2% of the 2654 proteins in the development set, we will sometimes group them together with the other TM proteins to provide more meaningful statistics. The largest difference between Philius and Phobius at this level is in the precision for the TM and SP+TM category, for which Philius calls 29% fewer false positives than Phobius. (Phobius finds 265 of the 292 true positives, and miscalls 82 of the 2362 true negatives; on the same data, Philius finds 268 TPs and miscalls 58 TNs.) Overall, the performance on the G and SP+G subsets has decreased slightly in exchange for an improvement on the TM subset which is of greatest interest. Note that the class sizes in this dataset are skewed (48% SP+G, 41% G, and 11% TM and SP+TM), and that compared to a complete proteome, the transmembrane proteins are underrepresented in this dataset by a factor of 2 to 3.
For each prediction, Philius reports a protein type confidence score, and Figure 4 shows that this score correlates extremely well with the precision of the classification decision. Furthermore, on this dataset, more than 70% of the confidence scores are greater than 0.95. For the TM and SP+TM proteins (the smallest class), the confidence score tends to be somewhat optimistic, as indicated by the points below y = x. We attribute this skew to the fact that the model was tuned to maximize the balanced accuracy across the three major classes.
Next, we evaluated the performance of Philius at the segment level. Philius predicts four basic segment types: signal peptide, transmembrane segment, and inside and outside loops. For a transmembrane segment, the predicted segment must overlap the annotated segment by at least five amino acids to be deemed correctly identified. In order to correctly identify a signal peptide, the model must only predict its existence at the N-terminus of the protein. Because many of the inside and outside loops are very short, the overlap required for these segments is only one amino acid. The sensitivity and precision of the model in predicting each of these segment types is shown in Table 2. Accuracy and specificity cannot be calculated at the segment level, because there is no sensible way to define the number of true negatives. Results for outside segments are reported for all segments as well as for the subset of outside loops within transmembrane proteins (i.e., those with at least one non-SP TM segment). All of the inside segments reported are loops within TM proteins. Predicting whether a loop between two adjacent TM segments is on the ‘inside’ or on the ‘outside’ of the membrane is clearly the most challenging aspect of this task.
As shown in Figure 5, the segment-level scores correlate well with precision. The membrane segment and inner and outer loop scores tend to be conservative, as indicated by the points above y = x. The segment score should be interpreted conditioned on the assumption that the protein type has been correctly predicted.
Although the precise boundaries of the membrane segments of a transmembrane protein are somewhat difficult to define, the cleavage site of a signal peptide can be very precisely defined if the first amino acid of the mature protein is known. We therefore also evaluated Philius' ability to correctly localize the signal peptide cleavage site.
Combining the SP+G and the SP+TM proteins into one group and the G and TM proteins into another, the development dataset contains 1320 proteins with signal peptides and 1334 without. In the cross-validation experiment, Philius predicts 1271 true positives, 1278 true negatives, 49 false negatives, and 56 false positives (accuracy = 0.96, precision = 0.96, sensitivity = 0.96, and specificity = 0.96).
Of the 1271 predicted true positives, in 948 cases (75% of the predicted positives, and 72% of all positives), the annotated cleavage site is found exactly. Among the errors, there is very little skew in the localization error: in 51% of the cases, the cleavage site is predicted “early” (median offset is 3 amino acids), and in 49% of the cases the cleavage site is predicted “late” (median offset is 2 amino acids).
For proteins with transmembrane segments (with or without a signal peptide), it is important to be able to correctly predict the entire protein topology. Getting this prediction right requires not only that all of the transmembrane segments be correctly identified, but that the loop regions between the membrane segments be correctly localized. Grouping the TM and SP+TM proteins together, Philius predicts the correct topology for a total of 212 out of 292 proteins (72.6%). For comparison, Phobius predicts 198 correct topologies (67.8%) on this same dataset. Table 3 shows the confusion matrices for Philius and Phobius. Within each half of the table, values on the diagonal represent correct protein-type predictions, while off-diagonal values represent errors. For G and SP+G proteins, a correct protein-type prediction implies a correct topology, whereas for TM and SP+TM proteins this is not necessarily the case. For these proteins, the first number represents the number of correct complete topologies while the second number represents the number of incorrect topologies. (Incorrect protein-type calls necessarily imply incorrect topologies.)
Figure 6 shows that the full-topology confidence score correlates reasonably well with the observed precision for the transmembrane proteins in the dataset. As with the segment scores, the full-topology confidence score should be interpreted conditioned on the assumption that the protein type has been correctly inferred.
Following the model-development phase, we evaluated Philius on an enhanced dataset that includes the SCAMPI dataset [27] and the SignalP 3.0 dataset of signal peptide proteins [27]. These new datasets partially overlap the original Phobius datasets that were used during model development as shown in Figure 7. We incorporated this new data to create a new set which we used for a final round of ten-fold cross-validated training and testing. This new dataset was made up of the original Phobius G and SP+TM subsets, the SignalP signal peptide set (combining eukaryotic and bacterial proteins), and a merged TM set created by combining the 124 TM proteins from the SCAMPI set with the 200 non-homologous TM proteins from the Phobius TM subset, for a total of 324 TM proteins.
Results were evaluated in two areas: full-topology accuracy on the transmembrane proteins, and signal peptide prediction accuracy on the SignalP dataset. The full-topology accuracy on the TM proteins after performing ten-fold cross-validation on this new dataset is summarized in Table 4. The accuracies reported in the first 2 rows of the table are consistent with one another and with the accuracy of 72.6% reported on the development set. Comparing the last two rows in the table it is clear that the novel portion of the SCAMPI dataset contains membrane proteins that are more difficult to predict. This is likely due to the presence in the SCAMPI set of 20 proteins known to have one or more re-entrant segments. Of these 20 proteins, all but one are in the SCAMPI \ Phobius set, and the full-topology accuracy on these 19 proteins is only 53% (10/19).
Training and testing Phobius in the same way on this new merged dataset yielded an overall TM topology accuracy of 62.7% (203 out of 324). Compared to Phobius, on this new dataset, Philius achieves a relative increase of 13% in the number of correct topologies (230 correct topologies vs 203).
The signal peptide performance is improved over that reported for the development dataset. We attribute this improvement to the higher quality SignalP dataset. On 1728 signal peptides, Philius predicted 1679 true positives and 49 false negatives (30 were classified as transmembrane proteins, while 19 were classified as globular proteins) for a sensitivity of 0.97 (compared to 0.96 on the Phobius SP set). Furthermore, 1292 cleavage sites are predicted exactly, representing 75% of all signal peptides in the test set, compared to 72% when trained and evaluated on the Phobius SP set.
Although we combined the eukaryotic and bacterial signal peptides during training, we also report in Table 5 the results broken down by taxon. For these results, the positive set is the SignalP dataset of signal peptides (with the counts for each subset as shown in the table), and the negative set is the Phobius globular protein set (1087 proteins). The results represent the summary from a ten-fold cross-validation experiment. Although we are not using the same set of negative (non-SP) proteins and thus cannot exactly replicate the experiments leading to the SignalP 3.0 performance figures reported by Bendtsen et al. in [28], Philius' detection and discrimination of signal peptides is comparable to that reported for SignalP-HMM for eukaryotes and Gram negative bacteria. The cleavage site accuracy reported here for Philius is slightly worse than SignalP-HMM for the eukaryote and the Gram negative sets (down 4% and less than 3% respectively), but is significantly worse for the Gram positive set (down 24%). This decline in performance is to be expected, considering that we trained a single model for all three categories, and the Gram positive signal peptides are significantly different from the other two types.
The key difference between Philius and SignalP, however, is that SignalP is trained to discriminate between proteins with and without signal peptides, excluding transmembrane proteins, whereas Philius has been trained to discriminate between proteins with and without signal peptides and those with and without other (non-SP) membrane-spanning segments. It has previously been reported that SignalP 3.0 falsely predicts 21% (52 of 247) of the Phobius TM dataset as containing signal peptides and that 30–65% of all predictions from SignalP 3.0 on whole proteomes overlap with TMHMM 2.0 predictions [8]. Philius, in contrast, predicts only 5% (13 of 247) of the Phobius TM dataset as containing a signal peptide.
Kim et al. [32] described the experimental localization of the C-terminus for 617 Saccharomyces cerevisiae proteins predicted by TMHMM to be multi-spanning membrane proteins using a reporter construct. Based on consistent experimental results as well as BLAST homology searches, the C-terminal location could be confidently assigned for a total of 546 proteins. For 69% of the 546 proteins, the initial TMHMM prediction of the C-terminal location agreed with the experimental result. New topology predictions were made using both TMHMM and prodiv-TMHMM [33] constrained by the experimentally determined C-terminal location.
The Philius predictions for the 546 proteins described above match the experimentally assigned C-terminal location 78% of the time (428 out of 546). For those C-terminal segments that were correctly predicted by Philius, the median confidence score was 0.90. For those incorrectly predicted, the median score was 0.72. Figure 8 shows the total counts and fraction of correctly localized C-terminals as a function of the C-terminal segment confidence score.
Constrained Philius topology predictions were then made and compared to those given in [32]. The Philius-predicted topology matched both TMHMM and prodiv-TMHMM for 41% of the 536. (For 10 out of the original 546 proteins, the length of the protein given in the supplementary data of [32] did not match the length of the ORF of the same name in the YRC database, so these proteins were disregarded for all other comparisons.) proteins, only prodiv-TMHMM for 21%, only TMHMM for 16%, and neither for 22%. The constrained predictions from TMHMM and prodiv-TMHMM match each other for 48% of the 546 proteins.
The constrained Philius predictions included 40 topologies containing a predicted N-terminal signal peptide. Of these, 31 signal peptides had high confidence scores (greater than 0.9), and all but one of these were also classified as containing a signal peptide by SignalP 3.0. Of these 30 putative signal peptides identified both by Philius and by SignalP, TMHMM annotates 18 (60%) as transmembrane segments. Four of these proteins are classified as SP+G by Philius, indicating that the mature protein is likely a globular protein and not a membrane protein. Of these proteins, three (YFL051C, YNL019C, and YNL033W) are putative proteins, and the fourth (YFL067W) is of uncharacterized function.
A final version of Philius, trained on all of the training data, was used to predict and score the protein type and topology for all 6.3 million proteins in the YRC public data repository [11] as of March 24, 2008. This database contains Uniprot/SwissProt, the NCBI non-redundant database, the MIPS protein sequence database, and a variety of organism-specific databases, including the Saccharomyces Genome Database, Sanger's S. pombe database (pompep), Wormbase, Flybase and The Arabidopsis Information Resource. Running Philius on this set required approximately 7.2 s per protein, for a total of approximately 1.5 years of CPU time.
A summary of the predictions can be found in Table 6. The median protein type confidence scores are very high for all protein types. The median topology confidence score for TM proteins is 0.69, which agrees with the typical topology accuracy of 70%. Table 7 shows the relative representation of the four basic protein types, for four species. The total fraction of predicted membrane proteins, between 22% and 29% is consistent with previous estimates. Table 8 shows the fraction of predicted TM and SP+TM proteins that have a single membrane-spanning segment in the mature protein. Single-spanning membrane proteins represent approximately 20% to 35% of all membrane proteins, and an even larger fraction of membrane proteins with signal peptides. For putative multi-spanning transmembrane proteins, proteins predicted to contain an even number of membrane segments outnumber those predicted to have an odd number of membrane segments nearly 2 to 1 (data not shown). This enrichment of membrane proteins with an even number of TM segments may be due to internal duplication events resulting in an even number of TM segments, or the process of membrane insertion may be optimized for pairs of segments. Although the N-terminus of a membrane protein is in general more likely to be on the cytoplasmic side of the membrane, this bias is strongest for proteins with an even number of membrane segments. Two extreme examples illustrate this phenomenon: less than 41% of the putative seven-transmembrane segment proteins are predicted to have the N-terminal on the inside (the large family of GPCR proteins have the N-terminal on the outside), whereas 96% of the proteins predicted to have twelve transmembrane segments are predicted to have the N-terminal on the inside. This same phenomenon was seen in our training data and in other genome-wide prediction studies [32],[34].
Figure 9 shows the Philius topology prediction for the human presenilin protein. This topology matches the nine-transmembrane topology which has been recently described [35],[36] and is supported by experimental evidence. The nine membrane-spanning regions are shown as vertical cylinders and the cytoplasmic and non-cytoplasmic segments as horizontal bars. Each segment is colored according to type and shaded according to the confidence score. The seventh membrane-helix is missed by many topology predictors and is assigned a relatively low confidence score by Philius and as such is shaded gray. The protein type score for this protein is 0.99, and the full-topology score is 0.56.
We have described Philius, a DBN-based approach to transmembrane protein topology prediction. Philius incorporates a two-stage decoding procedure that approximates the most likely label sequence given the protein sequence, a flexible way of handling uncertainty in training labels or partially labeled data, three different types of duration models, and a simple mechanism for tying parameters in order to limit model complexity. We have shown improvements in topology prediction accuracy over Phobius and comparable signal-peptide discrimination to SignalP-HMM. Furthermore, Philius uses a probabilistic framework to derive three informative confidence measures which have been shown to correlate well with observed precision. Finally, we have made available through the YRC web page a prediction server and 6.3 million predicted protein topologies. The predictions provide a global view of membrane protein topology and are a significant resource for scientists interested in understanding protein structure and function.
With respect to the transmembrane protein topology prediction task, we plan to improve Philius in several respects. First, it has previously been shown that the performance of Phobius could be increased from 67.8% to 74.7% correctly predicted TM topologies by including homologs during the decoding stage [26]. Philius currently achieves 72.6% accuracy on the same dataset. We believe that Philius's performance could be similarly improved by exploiting homologs. Other directions for future work include learning periodic motifs (such as the hydrophobic moment [37]) in transmembrane helices, and including parallel tracks of information, such as hydrophobicity measures, in addition to the amino acid sequence. A model that differentiates between single-spanning and multi-spanning membrane proteins may also better capture some of the diversity among these proteins, at the risk of data-sparsity problems. However, including additional features such as hydrophobicity or otherwise clustering the amino acids may help to limit over-fitting to the training data. Furthermore, most existing membrane protein models, including Philius, are guilty of over-simplifying the problem, ignoring, for example, re-entrant segments which penetrate but do not completely span the membrane, or interfacial helices which run roughly parallel to the membrane surface [38]. Modeling and predicting these types of features without reducing the accuracy on more “conventional” membrane proteins remains an open problem.
Recently, some insight has been gained into which properties of a protein govern the insertion of its membrane segments. Specifically, it has been shown that for a potential transmembrane helix of a given protein, the apparent free energy of insertion ΔGapp of a TM helix can be expressed as a function of the concentration ratio Kapp between the membrane integrated and the non-integrated forms: ΔGapp = −RT ln Kapp [39],[40]. Furthermore, this ΔGapp can be approximated as a sum of position- and residue-dependent contributions from each amino acid in the helix, plus a hydrophobic moment contribution and a length correction [27],[40]. The additive nature of ΔGapp, neglecting the hydrophobic moment term, supports the conclusion that probabilistic models in which the probabilities of individual amino acids are multiplied together, or equivalently the log-probabilities are summed, provide an accurate representation of the underlying membrane integration process. The length correction term can be compared to log Pr[Dh], where Dh is the length of the core membrane helix and Pr[Dh] is learned. Within the DBN framework, it is also possible to incorporate additional dependencies between nearby amino acids in order to capture effects such as the hydrophobic moment.
Since their introduction to biological sequence analysis [41], hidden Markov models have been considered one of the best ways to model amino acid and DNA sequences. DBNs generalize HMMs and offer a number of significant advantages. While adding complexity to an HMM requires an ever-expanding state space, a DBN can be used to more precisely describe the relationships desired among the random variables, thereby limiting the complexity only to what is actually needed. Because DBNs expose additional factorizations that might not be apparent in an HMM, DBNs may require fewer parameters and allow computationally more efficient probabilistic inference procedures than the corresponding HMM. Recently, Yao et al. [42] have applied DBNs to the task of secondary structure prediction and it seems like a logical step to similarly extend other applications such as gene prediction [43], protein homology detection [23], and coiled-coil prediction [44] from HMMs to DBNs. The DBN used here for protein topology prediction can easily serve as the basis for any similar segmentation and labeling task simply by specifying a different set of states and a different grammar.
|
10.1371/journal.ppat.1002116 | The SV40 Late Protein VP4 Is a Viroporin that Forms Pores to Disrupt Membranes for Viral Release | Nonenveloped viruses are generally released by the timely lysis of the host cell by a poorly understood process. For the nonenveloped virus SV40, virions assemble in the nucleus and then must be released from the host cell without being encapsulated by cellular membranes. This process appears to involve the well-controlled insertion of viral proteins into host cellular membranes rendering them permeable to large molecules. VP4 is a newly identified SV40 gene product that is expressed at late times during the viral life cycle that corresponds to the time of cell lysis. To investigate the role of this late expressed protein in viral release, water-soluble VP4 was expressed and purified as a GST fusion protein from bacteria. Purified VP4 was found to efficiently bind biological membranes and support their disruption. VP4 perforated membranes by directly interacting with the membrane bilayer as demonstrated by flotation assays and the release of fluorescent markers encapsulated into large unilamellar vesicles or liposomes. The central hydrophobic domain of VP4 was essential for membrane binding and disruption. VP4 displayed a preference for membranes comprised of lipids that replicated the composition of the plasma membranes over that of nuclear membranes. Phosphatidylethanolamine, a lipid found at high levels in bacterial membranes, was inhibitory against the membrane perforation activity of VP4. The disruption of membranes by VP4 involved the formation of pores of ∼3 nm inner diameter in mammalian cells including permissive SV40 host cells. Altogether, these results support a central role of VP4 acting as a viroporin in the perforation of cellular membranes to trigger SV40 viral release.
| Viruses exploit host cells for their propagation. Once an adequate number of viral particles have been assembled within the cell through the aid of cellular machinery of the host cell, the virus must be released from the cell for the virus to spread. For nonenveloped viruses or viruses that are solely encapsulated by a protein shell, this step most commonly involves the perforation of cellular membranes resulting in the lysis or death of the host cell. The mechanism for how this key terminal step in the viral life cycle is performed is poorly understood. We demonstrated that for the model nonenveloped virus SV40, the newly discovered virally encoded protein, termed VP4, perforates membranes by forming pores with a diameter of ∼3 nm in host cell membranes. While these pores are not of a sufficient size to provide a conduit that permits the movement of the virus through the membrane, they support membrane destabilization that leads to the disintegration of the membrane of the host cell and viral release.
| The virus life cycle is comprised of a number of consecutive, discrete, and tightly regulated steps. These steps include binding, internalization, penetration, replication, assembly, and release. The final step of the viral life cycle involves the exit of viral progeny from the cell to support virus dissemination for infection. Most enveloped virions are released from the cell by a budding or a membrane fission reaction with the virus acquiring a membrane coat during this process [1], [2]. Nonenveloped viral release generally requires cell lysis after the viral multiplication cycle has been completed [3], [4]. While this is a critical fundamental event in the viral life cycle, the mechanism of nonenveloped viral release is weakly defined.
The timing of the nonenveloped virus exit step is critical for an optimal life cycle, as lysis should occur immediately after an adequate number of virus particles have been assembled. This lytic event is associated with the disruption of cellular membranes, which leads to cell death. However, a cytolytic pathway that avoids apoptosis would also be advantageous, as this would ensure that the virus is not encapsulated by apoptotic membrane blebs, which would inhibit binding to host cell membranes.
Simian Vacuolating virus 40 (SV40) is a member of the polyomavirus family. SV40 was the first eukaryotic virus sequenced over thirty years ago [5], [6] and studies of SV40 have advanced our understanding of nuclear transport, transcriptional regulation, and cell transformation [7]. Therefore, it serves as an excellent paradigmatic virus to help expand our knowledge of the undefined stages of the nonenveloped viral life cycle including viral release. Three structural late proteins are found in the viral particle: VP1, VP2 and VP3. VP1 forms 72-pentameric capsomeres to create the viral capsid. Each capsomere contains a single copy of either minor structural protein VP2 or VP3 in its central cavity [8]. The same transcript encodes VP2 and VP3 with translation initiation occurring from sequential Met residues to create N-terminal truncations of each other. We recently discovered that another downstream Met in the VP2/VP3 transcript also acted as a translation initiation site to encode an additional protein termed VP4 [9]. VP4 is a 125 amino acid protein with a central hydrophobic domain (Figure 1A). As the synthesis of VP4 coincided with the time of viral-mediated cell lysis, VP4 was proposed to play a role in viral release.
In this study, we have investigated the ability of VP4 to bind and disrupt lipid membranes. For this purpose, a tagged version of VP4 was constructed and its properties of membrane binding and disruption were analyzed. These studies showed that VP4 efficiently bound red blood cell (RBC) membranes and supported membrane perforation as probed with a hemolysis assay. In addition, VP4 also disrupted liposomes mimicking the composition of mammalian plasma and nuclear membranes. Deleting or substituting the hydrophobic domain of VP4 drastically abolished its membrane disruptive activity, indicative of a role for the hydrophobic domain in membrane lysis. VP4 formed small ∼3 nm diameter pores in mammalian membranes. Altogether, these results support a central role for VP4 in the direct disruption of cellular membranes to trigger SV40 viral release.
To obtain recombinant SV40 VP4 for analysis of its activity, initially VP4 containing a C-terminal His tag was expressed in Escherichia coli. His tagged VP4 was undetected after 12 hr of induction (data not shown). The SV40 late protein VP4 contains a 19 amino acid central hydrophobic domain as suggested by hydrophobicity analysis (Figure 1A). The hydrophobic nature of VP4 combined with its potential lytic activity pose challenges for its stable bacterial expression, and its purification as a water-soluble and active protein [9].
The fusion of water-soluble proteins to the N-terminus of aggregation-prone polypeptides is an efficient and effective approach to maintain the solubility of a protein suitable for bacterial expression [10]. The addition of a bulky soluble protein such as glutathione S-transferase (GST) might diminish potential lethal properties of VP4 that contribute to its inability to be expressed in E. coli. To test if this strategy would support bacterial VP4 expression, a construct of VP4 with an N-terminal GST tag and a C-terminal His tag (GST-VP4, Figure 1A) was generated and its bacterial expression was analyzed.
The GST-VP4 construct was effectively expressed in E. coli after induction, however, the protein accumulated in the insoluble protein fraction indicative of its appearance in intracellular inclusion bodies (data not shown). Bacteria sequester insoluble proteins into intracellular inclusion bodies to maintain cellular homeostasis [11]. Since proteins found in inclusion bodies are commonly misfolded and therefore likely inactive, strategies for optimizing the production of soluble GST-VP4 were explored.
Osmolytes such as proline can be used to stabilize proteins in cells [12], [13]. The addition of proline (20 mM) and high concentrations of salt (NaCl 300 mM) to the medium followed by growth at the reduced temperature of 30°C significantly increased the solubility of GST-VP4 (Figure 1B, lanes 3 and 4). High concentrations of salt in the culture media induce E. coli to concentrate proline in vivo and decrease the tendency of some proteins to aggregate [14]. GST-VP4 from the soluble fraction was purified to homogeneity utilizing sequential interactions with its dual affinity tags (Figure 1C, lane 6). Size-exclusion chromatography demonstrated that GST-VP4 was monomeric (Figure S1). The cleavage of GST from VP4 with TEV protease produced the two corresponding proteins GST and VP4 (Figure 1C, lane 7). However, the release of GST from VP4 led to the aggregation of VP4. Therefore, purified uncleaved GST-VP4 was employed for all subsequent studies.
To determine whether VP4 displays membrane disruptive activity, a hemolysis assay was employed using bovine RBCs. RBCs constitute a simple and efficient model system to study protein-membrane interactions and lysis. RBCs were treated with various concentrations of purified GST-VP4 at 37°C for 30 min. The level of hemolysis was measured by determining the fraction of hemoglobin released into the supernatant after centrifugation. GST-VP4 mediated hemolysis was found to be concentration dependent. GST-VP4 efficiently lysed bovine RBCs in a concentration dependent manner inducing 50% lysis at a concentration of 5 µg/ml (Figure 2A). The equilibrium value of percent hemolysis showed sigmoidal dependence on GST-VP4 concentration, which is a characteristic feature of a process dependent on self-oligomerization [15]. Maximum hemolysis was observed with concentrations of 10 µg/ml of GST-VP4 or higher.
To determine the temperature dependency of hemolysis, time courses for hemolysis were studied at different temperatures using 10 µg/ml of GST-VP4. Aliquots were removed at the indicated times and percent hemolysis was determined. GST-VP4 rapidly and efficiently lysed the RBCs at 37°C (Figure 2B). The maximal level of hemoglobin was released after 10 min of treatment at 37°C. At the lower temperatures, hemolysis occurred after a lag phase of 5 to 15 min. This lag phase period increased as the temperature was reduced. In addition, the maximal level of hemolysis diminished as the temperature was lowered. This suggests that the disruption of RBCs by GST-VP4 involved a temperature dependent rate-limiting step.
The presence of metal ions and pH has been reported to influence the hemolytic activity of viroporins and bacterial pore-forming toxins [16], [17]. To further analyze the hemolytic activity of VP4, the metal ionic and pH dependency of the GST-VP4 mediated hemolytic reaction was characterized. Hemolysis was unaffected by the presence of Cs+, K+ or Mg2+ ions (Figure 2D). However, hemolytic activity increased by 25% in the presence of Ca2+. As biological levels of Ca2+ are regulated and vary over a large range from nM to mM [18], a broader range of calcium concentrations were tested for their effect on VP4 mediated hemolysis. Moderate increases in hemolysis were observed in the presence of 5 mM or higher concentrations of Ca2+ (Figure 2E). In addition, maximal hemolytic activity for GST-VP4 was observed at pH 7.2 (Figure 2C). The hemolytic activity of GST-VP4 was dramatically reduced in acidic conditions and mildly decreased in alkaline conditions. Altogether, GST-VP4 showed optimal membrane disruption activity at 37°C in the presence of mM calcium levels at neutral pH. All studies that follow were performed at the optimal temperature of 37°C and neutral pH.
Hydrophobic stretches of ∼20 amino acids in length can interact with membranes to form transmembrane segments [19]. To establish the role of the hydrophobic domain of VP4 in its ability to perturb membranes, the hydrophobic domain was deleted from GST-VP4 (GST-VP4ΔHD) and the membrane disruptive activity of the deletion mutant was analyzed. GST-VP4ΔHD was expressed and purified in the same way as GST-VP4 (Figure 1B and C). Deletion of the hydrophobic domain significantly reduced the membrane disruptive activity of VP4, indicative of the hydrophobic domain playing a critical role in plasma membrane lysis (Figure 3A). GST alone did not display membrane disruptive activity, verifying that the late viral protein VP4 possessed hemolytic or membrane disruption activity (Figure 3A). Furthermore, hemolytic activity of VP4 was also abolished when the hydrophobic domain was substituted with either of two well-characterized transmembrane segments from bacterial leader peptidase [20], [21], demonstrating that the hydrophobic domain of VP4 is specifically required for its lytic activity (Figure S2A and Text S1).
A cell-binding assay was employed to determine if VP4 interacts with cell membranes. GST-VP4, GST- VP4ΔHD and GST were separately incubated without and with RBCs for 30 min at 37°C, and bound and unbound fractions were separated by centrifugation. Cell binding was determined by the amount of protein that sedimented with the cells. In the absence of RBCs, all of the proteins tested remained soluble and were therefore found in the supernatant (Figure 3B, lanes 2 and 3). However, in the presence of RBCs, only GST-VP4 was localized to the RBC pellet, indicative of the efficient binding of VP4 to RBCs (Figure 3B, lane 6). Both GST- VP4ΔHD and GST remained in the supernatant after centrifugation demonstrating the necessity for the hydrophobic domain of VP4 for cell binding.
To determine if the bound protein was integrated into the lipid bilayer of the cells, the bound fractions were alkaline extracted with membrane and soluble fractions separated by ultracentrifugation [22]. After alkaline extraction, the abundant RBC membrane protein, anion exchanger 1 (AE1), was found in the membrane pellet (Figure 3C, lane 5) [23]. Interestingly, the vast majority of the membrane associated GST-VP4 (88%) was found in the supernatant after alkaline extraction and centrifugation (Figure 3B, lane 7 compared to 8). This suggested that GST-VP4 was not fully integrated into the lipid bilayer.
We next investigated if GST-VP4 required RBC membrane proteins for its hemolytic activity. To remove extracellular-exposed RBC proteins that could potentially serve as a platform for VP4 binding, RBCs were treated with both proteinase K and trypsin to generate protease-treated RBCs (pRBC). Membrane fractions from treated and untreated RBCs after hypotonic lysis to remove intracellular proteins were isolated by centrifugation. The protein content of the membrane fractions was monitored by SDS-PAGE. Numerous proteins corresponding to a wide range of sizes were observed in the untreated sample, but proteins were absent from the protease treated sample demonstrating the effectiveness of the protease treatment (Figure 4A). Protease treatment of RBCs or the removal of the surface proteins did not affect the GST-VP4 mediated hemolysis reaction (Figure 4B). Surface proteins were not required for the membrane disruptive activity of GST-VP4. This suggested that GST-VP4 interacted with the plasma membranes of the RBCs directly via the lipids.
To verify that VP4 acts on lipids directly, a membrane disruption assay was employed that utilized liposomes with lipid compositions representative of various biological membrane sources. This fluorescence-based spectroscopic assay detected the release of liposome-encapsulated fluorophores following the addition of GST-VP4. Bathing the liposomes in quenchers of the encapsulated fluorophore supports a reduction of fluorescence intensity if the fluorophore is released from the liposomes or the quencher is allowed to enter the liposome as a result of membrane disruption permitting contact between the quencher and the fluorophore (Figure 5A). In contrast, quenching is not observed if the membrane remains intact and the quencher and the fluorophore are unable to cross the membrane bilayer [24]. This experimental system provides a highly tractable approach to characterize the membrane disruption properties of VP4.
Large unilamellar vesicles (LUV) mimicking the lipid compositions of bacterial, and mammalian plasma and nuclear membranes were prepared to examine the membrane disruption activity of GST-VP4 (Table 1) [25], [26], [27]. GST-VP4 efficiently disrupted mammalian plasma (PM-like) and nuclear membrane-like (NuM-like) liposomes (Figure 5B). The activity against PM-like liposomes was approximately double that which was directed against NuM-like liposomes. The viral protein did not affect liposomes mimicking the lipid composition of the bacterial inner membrane (BcM-like). Furthermore, regardless of the liposomes tested, the hydrophobic domain was found to be essential for the membrane disruption activity of VP4, as VP4ΔHD and GST alone displayed no significant membrane disruption. These results demonstrated that VP4 possessed membrane-permeabilizing activity that required its hydrophobic domain and its activity was optimal against membranes that represented the lipid composition of the mammalian plasma membrane.
To explore if the differences observed for liposome disruption were caused by the ability of VP4 to bind to the various membranes, a membrane-binding assay was employed. A liposome flotation assay was used to isolate liposomes and monitor VP4 binding. Bound and unbound fractions of VP4 were quantified by immunoblotting. VP4 bound efficiently to both PM-like and NuM-like liposomes and did not display any association with BcM-like liposomes (Figure 5C, lanes 6, 9 and 12). These results were consistent with the fluorescence-based liposome disruption assay. In addition, VP4ΔHD mutant and GST did not bind any of the liposomal membranes (Figure 5C). Therefore, the liposome binding results indicate that the membrane disruption activity required stable membrane binding.
To investigate the origin for the difference in the activities of VP4 to the various membranes, liposomes comprised of different lipid compositions were tested to determine which lipids affected the membrane disruption properties of VP4. VP4 was most active against PM-like liposomes, followed by NuM-like liposomes, while it displayed marginal activity against BcM-like liposomes. Since BcM-like liposomes were rich in phosphatidylethanolamine (PE) and contained no cholesterol, the influence of PE and cholesterol on the activity of VP4 was tested. Given that VP4 supported 50% fluorescence quenching in NuM-like liposomes, the composition of these liposomes was modified in an attempt to optimize membrane disruption by VP4.
VP4-mediated membrane disruption increased when PE was excluded from the NuM-like liposomes (Figure 6A, NuM-like-noPE). This suggested that PE might inhibit membrane perturbation by VP4. The addition of cholesterol to the NuM-like liposomes (NuM-like+Chol) caused a significant decrease in membrane permeabilization. Cholesterol did not appear to produce a direct inhibitory effect as no difference was observed between the quenching of phosphatidylcholine (PC) liposomes and cholesterol+PC-containing liposomes (Figure 6B, Chol+PC). VP4 showed highest activity against PM-like membranes that contained 50% cholesterol, again suggesting that cholesterol did not have a direct inhibitory effect on VP4 activity (Figure 5B).
Increasing cholesterol levels in NuM-like liposomes was also associated with a decrease in the levels of PC and sphingomyelin (SM) (Table 1). Therefore, the level of SM and PE was varied while keeping the cholesterol content constant. First, the addition of PE to the cholesterol+PC liposomes produced a decrease in quenching (Figure 6B), consistent with the previously discussed inhibitory effect of PE. Interestingly, this inhibitory effect was reversed by the addition of SM (Figure 6B). As VP4 efficiently bound to all liposomes tested besides the bacterial-like liposomes that were rich in PE (Figure 5C, 6C and 6D), high concentrations of PE appeared to inhibit VP4 binding. Altogether these results showed that VP4 permeabilizing activity was dependent on the membrane lipid composition and VP4 was most active against PM-like liposomes due to what appeared to be a combined effect of lower PE and higher SM levels.
The observed release of liposomal content by VP4 led to the question of whether VP4 lysed the liposomes by creating discrete pores in the membrane or by a non-specific membrane solubilization or a detergent-like effect. In support of the pore formation hypothesis, LUVs average diameter or their total scatter intensity did not change significantly after incubation with VP4 (Figure 5D, 5E and data not shown). This indicated that the liposomes were not solubilized by the viral protein.
To further explore the characteristics of the pores formed by VP4, the pore forming activity of VP4 on cellular membranes was tested by analyzing the osmoprotection capabilities of different size polyethylene glycols (PEGs). The rationale for this approach is that pores at the plasma membrane produce cell lysis by uncontrolled water and ion influx. PEGs can serve as osmotic protectants and prevent water and ion influx, only if the PEG molecules are larger than the size of the VP4 formed pore. In contrast, if PEGs are small enough to pass through pores formed by VP4, their concentration equilibrates rapidly across the membrane and no osmotic protection is conferred, resulting in cell lysis. Therefore, the presence of discrete-sized pores in the membrane and their approximate diameter can be estimated by determining the minimum size PEG that confers osmoprotection [28].
The hemolytic activity of GST-VP4 was examined in the presence and absence of 30 mM PEGs. Smaller PEGs (1 and 4 kD) had negligible osmoprotection effect on hemolysis, whereas PEGs of 6 kD and larger reduced hemolysis (Figure 7A). These results supported the hypothesis that VP4 formed pores in RBC membranes. A sharp transition in osmoprotection was observed between PEGs of 4 and 6 kD with estimated hydrodynamic diameters of 3.8 nm and 6.4 nm, respectively. This suggested that VP4 formed ∼4–6 nm diameter pores in the plasma membrane of bovine RBCs.
The pore formation properties of VP4 were examined using Cos 7 cells, a SV40 permissive host cell line. Cell lysis can be followed by the release of cytosolic proteins. After VP4 treatment, the release of cytoplasmic lactate dehydrogenase (LDH) from Cos 7 cells was used to probe cell lysis. Incubation of Cos 7 cells at 37°C for 30 min led to the complete release of intracellular LDH (Figure 7B). Consistent with hemolysis and liposome disruption results, the deletion of the hydrophobic domain of VP4 abolished its cytolytic activity.
The cytolytic activity of VP4 on Cos 7 cells was analyzed using different size PEGs. PEGs larger than 3.35 kD efficiently inhibited cytolysis (Figure 7C). There was a gradual reduction in VP4-mediated cytolysis in the presence of PEG from 1 to 2 kD in size. These results suggested that VP4 formed smaller pores in the plasma membrane of Cos 7 cells. The transition in osmoprotection occurred with PEGs between 2 and 3.35 kD in size, which corresponded to hydrodynamic diameters of 2.8 and 3.5 nm, respectively. This indicated that VP4 formed pores of ∼3 nm in diameter in the cell membrane of the SV40 permissive host cells.
VP4 is a late expressed SV40 hydrophobic protein proposed to play a role in viral release [9]. Here, the membrane disruption properties of purified VP4 were thoroughly characterized. VP4 efficiently bound and permeabilized bovine RBCs. The hydrophobic domain of VP4 was required for these activities, which were optimal at 37°C and neutral pH in the presence of calcium. The disruption of LUV by VP4 demonstrated that VP4 had a preference for liposomes comprised of lipid compositions representing plasma and nuclear membranes. The membrane disruption activity of VP4 supported the formation of ∼3 nm pores in the plasma membranes of RBCs and the permissive SV40 host (Cos 7 cells). These results are consistent with VP4 causing the lysis of infected cells to support the poorly understood process of nonenveloped viral release.
VP4 disrupted membranes by directly interacting with the membrane lipid bilayer and its membrane disruption ability was dependent upon the composition of the bilayer. VP4 efficiently lysed bovine RBCs, which are rich in SM [29]. In liposome (or LUV) studies, SM rescued VP4 activity in the presence of the inhibitory lipid PE. Bacterial membranes contain high levels of PE. The inability of GST-VP4 to lyse PE-rich or BcM-like liposomes likely fortuitously contributed to the expression of large amounts of the protein in bacteria. PE is a conical lipid that induces negative membrane curvature, which might contribute to its inability to support VP4 pore formation [30], [31]. Pore formation by the antimicrobial peptide from the Xenopus skin, magainin 2, was induced by positive membrane curvature and inhibited by negative membrane curvature [30]. The membrane disruption activity of VP4 did not appear to be influenced by cholesterol, whereas high levels of PC favored efficient pore formation. Both nuclear and plasma membranes are rich in PC. While GST-VP4 disrupted liposomes that were representative of both plasma and nuclear membranes, it displayed a strong preference for PM-like liposomes. This was consistent with its ability to efficiently lyse RBCs and Cos 7 cells. The membrane disruption activity of VP4 was highly influenced by the lipid composition of the LUV.
Enveloped and nonenveloped viruses utilize proteins termed viroporins (viral-encoded membrane pores) to mediate membrane disruption events during the viral life cycle. Currently over a dozen viroporins are known and their functions appear to be similar to the well-studied toxins from bacterial pathogens such as Bacillus anthracis protective antigen (PA63) and E. coli (hemolysin) that form membrane pores [32]. Typically, viroporins are small (60–120 residues), hydrophobic proteins that form oligomeric structures in lipid bilayers of infected cells. These viral-encoded proteins form hydrophilic pores in host cell membranes to modify their permeability or stability. The membrane pores support the movement of ions or small molecules across membrane bilayers, to potentially aid in the viral entry and penetration steps, or promote the efficient release of virions by compromising the integrity of host cell membranes [33].
We demonstrated that VP4 acts as a viroporin. In vitro translated VP4 efficiently bound to GST-VP4 suggestive of the ability of VP4 to oligomerize (Figure S2B), a property shared by viroporins. Viroporins also frequently contain stretches of basic amino acids that are proposed to act as detergents by potentially binding to anionic lipid head groups [33]. VP4 possesses a basic pI of 10.2 and a large number of basic residues disproportionately clustered to the C-terminal side of the hydrophobic domain. These basic residues include a nuclear localization sequence, shared by both VP2 and VP3. While the lipid composition and the hemolytic activity of VP4 are consistent with its ability to associate with and disrupt plasma membranes, it will be of interest to determine what cellular membranes VP4 interacts with when it is expressed in the cytoplasm of host cells, as is the case during viral infection. Previous results found that VP4 appeared to accumulate at the nuclear periphery upon transfection of the SV40 viral genome lacking VP2 and VP3 [9]. The directionality of the VP4 membrane binding and disruption in the hemolysis reaction differs from that utilized during the viral life cycle. The plasma membrane is asymmetric, possessing higher concentrations of phosphatidylserine (PS) on the inner facing membrane leaflet. Since PS did not affect VP4 pore forming activity, this membrane asymmetry was not likely to affect the membrane permeabilization assay. These results indicated that VP4 effectively disrupted membranes that mimicked plasma and nuclear membrane compositions, with a preference for plasma-like membranes.
The protein-mediated mechanisms for membrane lysis have been studied extensively using antimicrobial peptides, which are lytic substances secreted by cells for a means of defense against microbial pathogens [34]. Antimicrobial peptides can contain short cationic and hydrophobic sequences that lack Cys residues, like VP4. Antimicrobial peptides have been proposed to disrupt bacterial membranes using three possible mechanisms. First, the barrel-stave model involves the formation of transmembrane pores created by alpha-helices integrated into the bilayer. This mechanism of membrane perturbation is unlikely for VP4 as VP4 was extracted from membranes after alkaline treatment, indicating that it was not fully integrated into the bilayer (Figure 3B). Secondly, in the carpet model, peptides accumulate on the membrane surface through electrostatic forces (cationic proteins binding anionic lipid head groups). At high concentrations, it is proposed that these peptides disrupt the membrane in a detergent-like manner resulting in the formation of micelles. The sharp size distribution in the pores formed by VP4 (Figure 7A and C), and the invariable diameter of the LUVs observed by dynamic light scattering (Figure 5D and E) after VP4 treatment are not consistent with VP4 acting through a carpet model since complete membrane disruption or lysis was not observed. Finally for the toroidal-pore model, proteins insert into the membrane and by interacting with the lipid head groups force curvatures in the interacting lipids, resulting in the fusion of the inner and outer leaflets at the lipid-protein interaction site. The toroidal-pore model differs from the barrel-stave model in that the protein interacts mostly with the lipid head groups and is not directly inserted through the hydrophobic core of the membrane [35]. The sharp pore size distribution and the ability to extract VP4 from membranes after alkaline treatment are consistent with VP4 forming a toroidal-pore within the membrane. However, further studies will be needed to fully delineate the membrane disruption mechanism of VP4.
Cytolytic viruses such as the nonenveloped polyomaviruses and picornaviruses release their viral progeny by initiating the timely lysis of host cells. The release of viral particles by cell lysis after adequate numbers of viral particles have been assembled ensures the efficient spread of the virus. Previously, we identified SV40 VP4 as a protein encoded within the viral genome that is expressed in the host cell at later times during infection that coincide with viral release [9]. The SV40 virus has a diameter of 50 nm therefore it is too large to be directly translocated through the ∼3 nm VP4 pores formed in Cos 7 cells. Whereas GST-VP4 efficiently forms pores in mammalian cells, at this time we cannot rule out the possibility that the size of the pore is influenced by the GST attached to its N-terminus. However, we favor the explanation that VP4 pores alter the cytoplasmic concentration of ions or other small molecules, which leads to cell lysis. This appears to be the mechanism for the 2B protein mediated release of picornavirus [36]. Alternatively, VP4 may form heterocomplexes involving other viral proteins that influence the size of the pores formed. Recently for JC virus (a human polyomavirus), the viral encoded agnoprotein was identified as a viroporin that aids in the release of JC virus [37]. As the SV40 genome also encodes for the hydrophobic agnoprotein, it is possible that these proteins work in concert to initiate viral release.
VP4 is an N-terminal truncation of the late structural viral proteins VP2 and VP3 [9]. Our previous studies found that VP3 and VP4 co-expression supported bacterial lysis suggesting that heterocomplexes between other late proteins and VP4 may influence lipid specificity or the size of the pores formed. While VP4 is solely found in infected cells [9], VP2 and VP3 are minor structural components of the viral particle [5], [6]. Upon internalization, SV40 traffics to the endoplasmic reticulum, the proposed site of uncoating and penetration [38]–[41]. An important issue for the penetration of nonenveloped viruses is how does a subviral particle or the viral genome cross endomembranes without disrupting cellular homeostasis so that the cell can be exploited for viral production for subsequent hours or days [42], [43]. VP2 and VP3 have been shown to insert into ER membranes [22]. This leads to the provocative possibility that since VP2 and VP3 both possess the VP4 sequence involved in membrane disruption, the exposure of VP2 and VP3 after viral uncoating supports membrane disruption to permit viral penetration. Interestingly, VP2 and VP3 from SV40 and polyomavirus have been shown to exhibit membrane disruption activities [44], [45]. Future studies will be required to address the properties of VP2/VP3 and VP4-heterocomplexes, and their roles in viral penetration and release.
The GST-Tag (12G8) mouse monoclonal antibody was purchased from Abmart (Arlington, MA). CytoTox 96 cytotoxicity assay kit for LDH release determination was purchased from Promega (Madison, WI). All phospholipids were obtained from Avanti Polar Lipids (Alabaster, AL), while cholesterol was obtained from Steraloids (Newport, RI). AcTEV protease was purchased from Invitrogen (Carlsbad, CA). All other reagents were purchased from Sigma (St. Louis, MO).
The pGEX-6P-1 plasmid (Amersham Bioscience; Piscataway, NJ) was modified to include a tobacco etch virus (TEV) protease site and a C-terminal 6xHis epitope upstream and downstream of the multiple cloning site, respectively, to create pGEX-6P-1-TEV-His. The GST-tagged version of full length VP4 was created by PCR cloning into the bacterial expression plasmid pGEX-6P-1-TEV-His using standard techniques. VP4 contained an N-terminal GST tag and a C-terminal His tag (GST-TEV-VP4-His). The QuikChange mutagenesis primer design program was used to delete the hydrophobic domain of VP4 (amino acids 65–83, PQWMLPLLLGLYGSVTSAL) to create VP4ΔHD. Mutagenesis was confirmed by sequencing.
The BL21 E. coli Rosetta strain (DE3: pLysS) (Novagen) was transformed with GST-Tev-VP4-His and grown at 37°C to an OD of 0.4 at 600 nm. NaCl (300 mM) was added to increase the osmolality of the nutrient medium. Simultaneous with the osmotic increase, the medium was supplied exogenously with 20 mM proline and the culture was induced with 1 mM IPTG for 4 hr at 30°C. Cells were centrifuged, resuspended in PBS (pH 7.4)/10 mM DTT with 200 µg/ml lysozyme and protease inhibitors, and rotated for 30 min at 37°C. Triton X-100 (1%) was added to the cells, which were then sonicated and the insoluble debris was sedimented by centrifugation for 20 min at 12,000× g. The clarified supernatant was then added to GST Sepharose 4B (Amersham Biosciences) matrix pre-equilibrated with PBS (pH 7.4)/10 mM DTT/1% Triton X-100. The matrix was washed three times with PBS (pH 7.4), and the protein was eluted with freshly prepared 10 mM reduced glutathione (in 50 mM Tris-HCl, pH 8.0). The eluate from GST Sepharose resin was further purified by adding the eluate to Ni-NTA His Bind resin (Novagen) pre-equilibrated with PBS (pH 7.4)/10 mM imidazole and additional 150 mM NaCl. The matrix was washed three times with PBS (pH 7.4)/50 mM imidazole, and the protein was eluted with 250 mM imidazole/PBS (pH 7.5). Protein purity was confirmed by SDS-PAGE and protein concentration was determined using a Bradford assay (Biorad). The expression and purification of GST-TEV-VP4ΔHD and GST-TEV-His were performed similarly.
Bovine RBCs were washed repeatedly in cold PBS immediately before use. Reactions (700 µl) were incubated at 37°C in hemolysis buffer (PBS, pH 7.4, 1 mM DTT and 200 ng BSA) with 0.5% RBCs, without or with 30 mM PEG (Fluka) and 10 µg/ml VP4, unless otherwise noted. Each time point for time course studies represented a separate reaction. End-point samples were removed after 30 min. Hemolysis was carried out in 10 mM Tris (pH 7.5)/150 mM NaCl containing 1 mM DTT and 200 ng BSA when the effect of ions was studied. Metal chlorides were added to the hemolysis reaction buffer at a final concentration of 40 mM. Reactions were centrifuged at 6,000× g for 5 min at 4°C to pellet unlysed cells. The absorbance of the supernatant was measured at 414 nm. Percentage of hemolysis was calculated as [(A414 (sample) − A414 (blank))/(A414 (water) − A414 (blank))] ×100. The blank reaction contained all components except protein and PEG. RBCs were hypotonically lysed by adding 50% water. Protease-treated RBCs were created by incubating cells at a concentration of 10% (v/v) in 10 mM Tris (pH 7.5)/150 mM NaCl, 20 mM MgCl2 containing 0.5 mg/ml proteinase K, 0.5 mg/ml trypsin, and 1.5 mM CaCl2 for 90 min at 37°C with gentle agitation. Proteases were inactivated with 2 mM PMSF and the cells were washed twice with cold 10 mM Tris (pH 7.5)/150 mM NaCl, 20 mM MgCl2 with 2 mM PMSF.
For determining RBC cell surface binding, hemolysis was carried out as described above at 37°C for 30 min. The cells were then centrifuged at 6,000× g for 5 min and the supernatant and pellet fractions were separated. Alkaline extractions were performed by resuspending the cell pellet in 700 µl of ice-cold 0.1 M Na2CO3 (pH 11.5), followed by a 30 min incubation on ice. The solution was layered on top of a 100 µl sucrose cushion, and the membrane-bound fraction was isolated by ultracentrifugation for 20 min at 65,000× g at 4°C. The membrane-bound pellet was resuspended in sample buffer, and the supernatant containing the peripherally associated proteins was TCA precipitated, washed with acetone, and resuspended in sample buffer for SDS-PAGE (13% acrylamide) and immunoblotting.
LUVs or liposomes were generated using a Liposofast extruder (Avestin Inc., Ottawa, Canada) [24]. Chloroform solutions of lipids [PC (POPC, 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine); PE (POPE, 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine); PS (POPS, 1-palmitoyl-2-oleoyl-sn-glycero-3-(phospho-L-serine); PG (POPG, 1-palmitoyl-2-oleoyl-sn-glycero-3-(Phospho-rac-1-glycerol); SM (sphingomyelin)] were mixed in the respective ratios and chloroform was evaporated at 37°C with a mild N2 flow. The lipid film was kept under vacuum for at least 3 h to eliminate traces of the organic solvent. To hydrate the lipid mixture, 0.25 ml of Hepes buffer saline (HBS, pH 7.5) was added to the dried phospholipid/sterol mixture (final total lipid concentration 10 mM), and the samples were incubated for 30 min at 37°C. The lipids were then resuspended by vortexing. The suspended lipid mixtures were frozen in liquid N2 and thawed at 37°C a total of five times to reduce the number of multilamellar liposomes and to enhance the trapped volumes of the vesicles [24]. Then the samples were passed at room temperature 21 times through the extruder equipped with a 100 nm pore size polycarbonate filter. The resulting liposomes were stored at 4°C and used within 2 weeks of production. Liposomes containing the fluorophore, Terbium-Dipicolinic acid [Tb(DPA)33-], were prepared as above, except that HBS buffer included 3 mM TbCl3 (Alfa Aesar, Ward Hill, MA), and 9 mM 2,6-pyridinedicarboxylic acid (DPA, neutralized to pH 7). The resulting liposomes were separated from non-encapsulated [Tb(DPA)33-] by gel filtration (Sepharose CL-6B-200, 0.7 cm inner diameter ×50 cm) in HBS buffer.
Liposomes (100 µM total lipids) were suspended in 300 µl of buffer A (PBS, pH 7.4) containing 5 mM EDTA. The net initial emission intensity (F0) was determined after equilibration of the sample at 25°C for 5 min. Aliquots of 3 µl containing the amount of protein that gives the final concentration of 5 µg/ml were added and the samples were incubated for 30 min at 37°C. After re-equilibration at 25°C, the final net emission intensity (Ff) of the sample was determined (i.e., after blank subtraction and dilution correction) and the fraction of marker quenched was estimated using (F0-Ff)/(F0-FT), where FT is the net emission intensity obtained when the same liposomes are treated with 3 mM Triton X-100 (i.e., under conditions of maximal release of the fluorophore).
Intensity measurements were performed using the Fluorolog 3–21 spectrofluorimeter equipped with a 450 W xenon arc lamp, a double excitation monochromator, a single emission monochromator, and a cooled PMT. The excitation wavelength/bandpass, and the emission wavelength/bandpass were respectively: 278/2 and 544/4 nm for [Tb(DPA)33-]. For [Tb(DPA)33-] measurements, a 385 nm longpass filter was placed in the emission light path to block any second-order scatter emission light. Measurements were done in 4×4 mm quartz microcells stirred with a 2×2 mm magnetic bar as described previously [24].
Binding reactions (75 µl) containing LUVs (400 µM) and protein (the amount of protein added was such that the ratio of LUVs to protein was same as that used in liposome permeabilization assay) were incubated at 37°C for 30 min. LUVs-bound and unbound proteins were separated by flotation through sucrose gradients, as liposomes float in the gradient when a g-force is applied, while free proteins sediment. Sucrose/HBS (225 µl of 67%) was added to the binding reactions and thoroughly mixed. The samples were overlaid with 360 µl of 40% sucrose, followed by 240 µl of 4% sucrose. Centrifugation was carried out for 50 min at 90,000× g at 4°C. Three 300 µl fractions (upper, middle, and bottom) were collected from the gradient. After trichloroacetic acid precipitation and resuspension in SDS sample buffer, samples were analyzed by SDS-PAGE and immunoblotting.
The average size of the liposomes, before and after incubation with VP4, was determined by dynamic light scattering. Measurements were made at room temperature using a PDDLS Coolbatch 90T/PD2000DLSPlus instrument (Precision Detectors, Inc., Franklin, MA) employing a 30-mW He-Ne laser source (658 nm) and a photodiode detector at an angle of 90°. VP4 and liposome concentrations were the same as that employed for the liposome permeabilization assays.
The assay was carried out by plating Cos 7 cells (10,000 cells/well) and incubating the cells with 10 µg/ml protein for 30 min at 37°C. Aliquots of media (50 µl) were removed for the determination of LDH release. The CytoTox 96 cytotoxicity assay kit was used to determine the level of LDH released from the cells according to the manufacturer's instructions. After allowing 30 min of incubation with substrate, the A490 was determined using a BioTek Synergy 2 multi-mode microplate reader. As a control for total cell-associated LDH, Cos 7 cells in selected wells were lysed with 0.9% Triton X-100. Percentage LDH release was calculated by dividing the A490 released from samples by total cell-associated LDH release and multiplying by 100.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) accession number for SV40 VP4 is DAA06058.1.
|
10.1371/journal.ppat.1004697 | Adhesive Fiber Stratification in Uropathogenic Escherichia coli Biofilms Unveils Oxygen-Mediated Control of Type 1 Pili | Bacterial biofilms account for a significant number of hospital-acquired infections and complicate treatment options, because bacteria within biofilms are generally more tolerant to antibiotic treatment. This resilience is attributed to transient bacterial subpopulations that arise in response to variations in the microenvironment surrounding the biofilm. Here, we probed the spatial proteome of surface-associated single-species biofilms formed by uropathogenic Escherichia coli (UPEC), the major causative agent of community-acquired and catheter-associated urinary tract infections. We used matrix-assisted laser desorption/ionization (MALDI) time-of-flight (TOF) imaging mass spectrometry (IMS) to analyze the spatial proteome of intact biofilms in situ. MALDI-TOF IMS revealed protein species exhibiting distinct localizations within surface-associated UPEC biofilms, including two adhesive fibers critical for UPEC biofilm formation and virulence: type 1 pili (Fim) localized exclusively to the air-exposed region, while curli amyloid fibers localized to the air-liquid interface. Comparison of cells grown aerobically, fermentatively, or utilizing an alternative terminal electron acceptor showed that the phase-variable fim promoter switched to the “OFF” orientation under oxygen-deplete conditions, leading to marked reduction of type 1 pili on the bacterial cell surface. Conversely, S pili whose expression is inversely related to fim expression were up-regulated under anoxic conditions. Tethering the fim promoter in the “ON” orientation in anaerobically grown cells only restored type 1 pili production in the presence of an alternative terminal electron acceptor beyond oxygen. Together these data support the presence of at least two regulatory mechanisms controlling fim expression in response to oxygen availability and may contribute to the stratification of extracellular matrix components within the biofilm. MALDI IMS facilitated the discovery of these mechanisms, and we have demonstrated that this technology can be used to interrogate subpopulations within bacterial biofilms.
| Bacteria are commonly found in multicellular communities known as biofilms. Biofilms can form on a variety of surfaces, both outside and within living things, and can have detrimental effects on human health. The characteristics of bacteria occupying different areas within biofilms are not well understood, and such knowledge is critical for understanding how biofilms form and for developing strategies to treat biofilm-related infections. Here, we adapted a technique to sample how proteins cluster within bacterial biofilms as a means to identify the location of bacteria with differential protein expression within the community. We observed that with uropathogenic E. coli, which is the major cause of urinary tract and catheter-associated urinary tract infections, bacteria close to the air-exposed region of the biofilm expressed different adhesive fibers compared to those at the liquid interface. We went on to show that lack of oxygen shuts down the production of fibers known to be critical for adherence to host bladder cells and to catheter material. This discovery was enabled by a new application of an existing technology that allowed us to gain insights into the spatial regulation of proteins within bacterial biofilms and to elucidate pathways that could be targeted to inhibit bacterial adherence.
| In nature, bacteria predominantly exist in a biofilm state [1] forming mutualistic or parasitic associations with other living organisms [2,3]. Within vertebrate hosts, the resident microbiota are essentially multi-species biofilms that play a key role in preventing colonization by pathogens [4]. Conversely, pathogenic bacteria exploit biofilm formation to colonize prostheses, catheters, as well as extracellular and intracellular host niches resulting in potentially life-threatening infections that are often difficult to treat [5]. Both single and multi-species biofilms are heterogeneous in nature, comprised of bacterial subpopulations with distinct tasks, such as expression of matrix components or a specific metabolic activity [6–9]. This “division of labor” within the community contributes to recalcitrance of the biofilm to antibiotic treatment. Biofilm subpopulations can be transient in nature, and arise in response to alterations in nutrient and oxygen availability of the surrounding microenvironment that in turn leads to local changes in bacterial gene expression [6–9]. However, little is known about the expression and distribution of individual protein species within a single multicellular community that results from this differential gene expression and how such differences may shape the characteristics and the fate of the biofilm.
Traditional techniques used to visualize protein distribution within intact biofilms rely on microscopy-based methods that require the use of either fluorescently labeled proteins or the application of antibodies specific to a protein of interest [10,11]. These techniques are limited to previously identified protein targets and can typically only accommodate one or two species in a single analysis. Conversely, more global genomic and proteomic-based analyses necessitate the destruction of biofilm architecture, leading to complete loss of spatial information.
Matrix-assisted laser desorption/ionization time-of-flight imaging mass spectrometry (MALDI-TOF IMS) is a surface-sampling technology that can determine spatial information and relative abundance of analytes directly from biological samples [12]. Samples are treated with a matrix that absorbs ultraviolet light from a laser source to ionize analytes of interest. The generated ions are accelerated along a time-of-flight (TOF) mass analyzer for separation and detection [13]. Using this technique, spectra are collected in a defined array across the sample, and each peak intensity in the spectra is then extrapolated to generate an ion intensity map, allowing for a two-dimensional representation of analyte distribution within the imaged array (Fig. 1 and [14]). This label-free technology does not require prior knowledge of sample composition or analyte distribution and provides an unbiased approach for the simultaneous localization analysis for multiple analytes within a single biological sample.
Here, we used MALDI-TOF IMS to examine the in situ distribution and localization of low molecular weight proteins within biofilms formed by uropathogenic Escherichia coli (UPEC). UPEC, one of the extra-intestinal E. coli pathotypes and the primary cause of urinary tract infections, can form extracellular biofilms on host cells and urinary catheters, as well as intracellular biofilm-like communities within host bladder epithelial cells [15–19]. These UPEC virulence mechanisms dictate multiple disease outcomes [20], including urosepsis that can have life-threatening complications [21]. MALDI IMS detected distinct protein localization patterns within the surface-associated UPEC biofilms imaged in these studies. Subsequent, conventional proteomic approaches led to the identification of several of the distinctly localized ion species. Among the proteins identified were CsgA and FimA, which comprise the primary structural subunits of curli and type 1 pili fibers respectively. Type 1 pili, encoded by the fim gene cluster, are chaperone-usher pathway (CUP) pili [22] that facilitate adherence to mannosylated moieties and are the primary determinant that enables a) UPEC attachment to the bladder urothelium, and b) inter-bacterial interactions in both extracellular and intracellular biofilms [15,23].
MALDI IMS revealed that, while curli subunit signatures are found at the air-liquid interface of the biofilm, which is consistent with their primary role in extracellular matrix infrastructure, type 1 pili subunit signatures predominantly localize to the air-exposed regions of the biofilm. Subsequent studies investigating the effects of anaerobiosis on expression of type 1 pili in UPEC led to the discovery of two regulatory mechanisms controlling expression of type 1 pili in response to the presence of oxygen. Together, these data demonstrate how MALDI IMS can be used to dissect the spatial proteome of an intact bacterial biofilm, and highlight how the information obtained can provide new insight into protein regulation relating to biofilm infrastructure.
In order to assess the utility of MALDI-TOF IMS for evaluating protein localization within bacterial biofilms, we adapted a simple surface-associated biofilm setup that enabled the sampling of single-species biofilms formed by uropathogenic Escherichia coli (UPEC) [24]. We optimized growth conditions to promote biofilm formation onto indium tin oxide (ITO) coated glass slides, given that our MALDI IMS must be performed directly from an electrically conductive surface for high voltage analyses [25]. Slides were placed vertically into culture media seeded with bacteria, such that only half of the slide was submerged within the media. This setup created an environmental gradient of oxygen and nutrients that induced biofilm formation at the air-liquid interface (Fig. 1A). We hypothesized that MALDI IMS would enable detection of distinct bacterial subpopulations resulting from the induced gradient (Fig. 1A).
MALDI-TOF IMS requires the application of a UV-absorbing matrix for analyte ionization [25] (Fig. 1B). Typical sample preparation methods begin with solvent washes to decrease ion suppression from lipids and salts within the sample in order to enhance protein ionization [25]. Here, we selected a sequential washing procedure of 70%, 90%, and 95% ethanol for 30 seconds each. Following washes, we evaluated biofilm integrity using three different techniques: crystal violet staining, scanning electron microscopy (SEM), and optical profilometry (S1 Fig). SEM analysis of the air-exposed, the air-liquid interface, and liquid-exposed regions of the biofilm indicated that the tertiary structure, along with cell shape and surface features, were preserved post-washing (S1 Fig). Crystal violet staining [26] and subsequent quantitation showed that the preparative ethanol washes did not significantly reduce biofilm levels (S1 Fig). Finally, optical profilometry [27] was used to assess the biofilm depth on the surfaces analyzed by MALDI IMS (S1 Fig). Combined, these approaches indicated that the sample preparation methods for MALDI IMS did not significantly perturb biofilm integrity.
A schematic for the MALDI-TOF IMS analysis of UPEC biofilms is shown in Fig. 1B. The MALDI methods and matrix selected for these studies were optimized for lower molecular weight protein species; therefore, all analyses were carried out over an ion range of mass-to-charge ratio (m/z) 2,000–25,000. Within this range, we observed 60 UPEC protein ion species that were detected reproducibly in at least 5 biological replicates (S1 Dataset). The relative abundance and localization patterns for representative ion species are shown in Fig. 2. Each panel depicts a heat-map intensity plot for a unique ion species within the biofilm, where red/white indicates the highest levels of relative abundance, and black/blue the lowest levels (Fig. 2). All observed ion species displayed one of the following localization/distribution patterns: diffuse distribution throughout the biofilm, localization specific to the air-exposed or liquid-exposed region, or localization to the air-liquid interface (Fig. 2). Overlay analysis of ion images demonstrated that we could differentiate localization patterns for different protein species within the same region of the biofilm (Fig. 2, ion overlay of m/z’s 5,596-red and 13,036-yellow).
Following MALDI IMS spatial analysis, enzymatic digestion of biofilm lysates and tandem mass spectrometry were used to identify select ion species observed (Table 1). These analyses identified the histone-like global transcriptional regulators HU-α (UniProt KB Q1R5W6, m/z 9,535) and HU-β (UniProt KB Q1RF95, m/z 9,226), which co-localized throughout the biofilm and were most abundant in the air-exposed region (Fig. 3A and S2 Fig). The acid stress-response chaperone protein, HdeB (UniProt KB Q1R595, m/z 9,064), and the uncharacterized protein YahO (UniProt KB Q1RFK1, m/z 7,718) were also identified (Table 1). HdeB localized to the air-liquid interface and was most abundant towards the liquid-exposed surface, while YahO localized throughout the biofilm (Fig. 3A and S2 Fig). Finally, two of the IMS signals identified by proteomics corresponded to major subunits of two UPEC adhesive organelles: The major curli subunit CsgA (UniProt KB Q1RDB7, m/z 13,036), an essential determinant for UPEC biofilm formation under the culture conditions used for these studies [7,28], and; the major subunit of type 1 pili, FimA (UniProt KB Q1R2K0, m/z 16,269).
Based on the MALDI IMS results, CsgA signatures were predominantly found at the air-liquid interface of the biofilm (Fig. 3A-B and S2 Fig), consistent with the role of curli as the primary extracellular matrix (ECM) component under the biofilm conditions tested. Conversely, FimA localized uniquely to the air-exposed region of the biofilm (Fig. 3A-B, and S2 Fig). Under the biofilm growth conditions used for these studies, type 1 pili have been shown to play an accessory role to biofilm infrastructure, and loss of type 1 pili impairs integrity but does not abolish biofilm formation [7]. Thus, we took advantage of a fim deletion mutant (UTI89ΔfimA-H) to validate the identification of the m/z 16,269 ion as FimA. MALDI IMS analysis of UTI89ΔfimA-H biofilms showed a loss of the ion at m/z 16,269 (S3 Fig), confirming the ion m/z 16,269 as FimA. Similarly, the ions m/z 9,535 and m/z 7,718 were validated as HupA and YahO respectively, through MALDI analysis of UTI89 mutants lacking the respective gene (UTI89ΔhupA and UTI89ΔyahO) (S3 Fig).
Given that curli are essential for UPEC biofilm formation under the conditions tested, we utilized a more traditional immuno-fluorescence approach with an antibody against CsgA to visualize curli-expressing bacteria within the biofilm and validate CsgA localization to the air-liquid interface. Combining immunohistochemistry with super-resolution structured illumination microscopy (SIM), we observed that the majority of curli-producing bacteria localized to the air-liquid interface of the biofilm, with only sparse populations found at the air- and liquid-exposed regions (Fig. 4, S1 Video). These data confirmed the IMS observations of CsgA localization to the air-liquid interface of the biofilm. As an orthologous approach, we took advantage of small peptidomimetic molecules that interfere with curli biogenesis in UPEC [29]. We hypothesized that treatment of pre-formed biofilm with one such compound, FN075 [29], should block curli fiber subunit incorporation leading to an abundance of CsgA monomers within the biofilm that could be detected by IMS. To test this hypothesis we cultured UPEC biofilms for 24 hours, at which time we added FN075 or DMSO (vehicle control) at previously reported concentrations [29]. Biofilms were allowed to grow in the presence of compound/vehicle for 24 hours prior to quantitation by crystal violet staining and imaging by MALDI IMS (S4 Fig). Consistent with previous observations [30], DMSO treatment increased biofilm levels and CsgA expression compared to untreated controls (S4 Fig). Though these experiments were carried out under atmospheric conditions, DMSO can serve as an alternative terminal electron acceptor for E. coli [31]. This ability of DMSO may be contributing to the observed increase in biomass, though additional studies are needed to dissect the basis of biofilm increase in response to DMSO treatment. Colorimetric quantitation of biofilm levels also revealed a significant reduction in biomass with FN075-treatment of biofilms (p = 0.0089), compared to the DMSO-treated controls (S4 Fig). Consistent with the difference in biofilm levels, average MALDI IMS spectra normalized to the total ion current (TIC) indicated a higher level of overall signal within the DMSO-treated samples (S4 Fig). To account for the differences in biofilm levels between non-treated 48 hour biofilms, DMSO-treated, and the FN075-treated samples, mMass [32] software was used to normalize the overall intensity of the average spectra of each sample to the most abundant ion in the analysis (m/z ~7,280). These normalization parameters revealed an apparent increase in detection of the ion species corresponding to CsgA (m/z 13,036) within the FN075-treated sample, despite the reduction in overall biofilm levels (S4 Fig). IMS ion images for CsgA also appeared to show an increase in detectable CsgA monomers within the liquid-exposed region of the biofilm (S4 Fig). This is consistent with our hypothesis that FN075 treatment of a pre-formed biofilm would lead to an increase in monomeric CsgA, which would be more readily ionized and thus detected. Having validated the identity and localization of CsgA and FimA, we next sought to understand the basis of the spatial segregation of type 1 pili within UPEC biofilms.
The observation that type 1 pili-producing bacteria make up the top-most layer of the biofilm led us to the hypothesis that oxygen tension, at least in part, regulates the expression of type 1 pili. The fim gene cluster is under the control of a phase-variable promoter region (fimS), the orientation of which in UTI89 is directed by the action of site-specific recombinases FimB, FimE, and FimX (Fig. 5A) belonging to the lambda integrase family [33]. At least two other global transcriptional regulators, Lrp and IHF, have been proposed to bend the fimS DNA in order to bring the invertible repeats in close proximity to each other and allow for recombination [33,34]. We used a previously developed PCR-based “phase assay” [35] that can distinguish between the transcription-competent ON (fimON) and transcription-incompetent OFF (fimOFF) orientations of the fim promoter (Fig. 5A), along with immunoblot analysis and transmission electron microscopy to evaluate whether oxygen is requisite for fim expression.
UTI89 was grown statically in either the presence or absence of oxygen in two different growth media (YESCA and Luria Bertani (LB)) and in two different temperature conditions (room temperature and 37°C) to evaluate the possibility that Fim localization to the air-exposed region was due to a nutritional or a temperature cue (S1 Table). Static growth at 37°C in LB media under atmospheric conditions enhances expression of UPEC type 1 pili [36–38]; these conditions were used as a positive control. UTI89ΔfimA-H was used as a negative control. Given the static nature of all culture methods, cultures grown in the presence of oxygen were termed “semi-aerobic”.
When starting these experiments from UPEC cultures that were primarily fimOFF, we observed that sub-culturing statically in the presence of oxygen induced expression of type 1 pili (Fig. 5B—“semi-aerobic”, S5 Fig, S6 Fig). However, regardless of growth medium or temperature, the fim promoter remained in the fimOFF orientation when bacteria were cultured in the absence of oxygen (fermentative conditions) (Fig. 5B and S5 Fig). When oxygen is not present, E. coli can utilize alternative terminal electron acceptors, such as nitrate, DMSO, TMAO, or fumarate [31]. Given that nitrate is the preferred alternative electron acceptor for E. coli, we assayed how anaerobic growth in the presence of nitrate (in the form of 40 mM sodium nitrate, NaNO3) would impact expression of type 1 pili. We observed that static cultures started fimOFF remained largely fimOFF during anaerobic growth in the presence of NaNO3 similar to what was observed with cultures grown fermentatively (Fig. 5B). When populations grown fermentatively or anaerobically with nitrate were sub-cultured into semi-aerobic conditions for 18 hours, the phase-variable promoter returned predominantly to the fimON orientation, leading to increased FimA protein levels (Fig. 5B). These results suggested that the phase-switch from fimOFF to fimON is affected by the bacterial respiration state, favoring aerobic respiration.
Previous studies indicated that multiple static sub-cultures under aerobic conditions enhance expression of type 1 pili by enriching for UPEC populations in which the fim promoter is fimON [36,37]. We thus repeated our experiments starting from cultures that were pre-enriched for fimON populations to test whether this would influence piliation in the absence of oxygen. Phase assays, FimA western blot analyses, and transmission electron microscopy (TEM) revealed that under fermentative conditions, the promoter actively inverted to the fimOFF orientation (Fig. 5C), leading to significantly fewer pili on the cell surface (Fig. 5C and S7 Fig). These data suggest that under fermentative conditions the phase-switch is preferably in the fimOFF orientation. Interestingly, growth of fimON cells in the presence of nitrate partially preserved the fimON state and production of type 1 pili on the surface (Fig. 5C and S7 Fig). The partial preservation observed under anaerobic growth in the presence of nitrate for populations starting fimON suggests that anaerobic respiration does not impact the fimON to fimOFF phase-switch. Together, these data suggest a regulatory mechanism that actively senses and responds to environmental oxygen levels, and/or bacterial respiration state, to control the expression of type 1 pili in UPEC by altering fimS promoter orientation.
In previous studies we created a UPEC strain (UTI89_LON) in which the fim promoter element is genetically locked into the transcription-competent fimON orientation [38]. We postulated that if oxygen/respiration state only impacts the phase-state of the fim promoter, then UTI89_LON would be piliated when cultured in the absence of oxygen. When cultured under fermentative conditions, UTI89_LON exhibited a marked reduction in type 1 pili production, similar to wild-type (WT) UTI89, despite the “locked on” position of the promoter (Fig. 5D and S7 Fig). The phase state of the fim promoter in UTI89_LON was verified by phase assays (Fig. 5D) to exclude the possibility of mutations affecting the phase state under the conditions tested. These data point towards an additional regulatory mechanism that influences production of type 1 pili in a manner that is independent of the fim promoter switch.
Interestingly, anaerobic growth in the presence of nitrate induced fim gene expression in UTI89_LON (Fig. 5D), similar to the fimON population shown in Fig. 5C (Fig. 5C-D and S7 Fig). Taken together, these observations suggest that the absence of oxygen impacts the phase state of the fim promoter element, and demonstrate that if the promoter is found in the fimON orientation, the presence of an alternative electron acceptor is sufficient to induce transcription.
Previous studies indicated that reduction in the expression of type 1 pili induces the expression of S pili under type 1 pili-inducing conditions [39–41]. We therefore evaluated the presence of S pili on the surface of the cell. Type 1 pili are characterized by their ability to bind mannosylated moieties [42]. An assay to evaluate the extent of type 1 pili in a UPEC population involves the agglutination of guinea pig red blood cells in the presence and absence of mannose. In bacteria that solely express type 1 pili, hemagglutination can be abolished by the addition of mannose to the agglutination reaction [42]. S pili bind sialic acid residues; therefore desialylation of red blood cells using neuraminidase prior to the agglutination assay abrogates S pili-dependent hemagglutination [43,44]. We combined these two approaches to establish the identity of the pili produced by UTI89 under anaerobic growth with cultures started from populations primarily fimON. As expected, when WT UTI89 was grown statically in the presence of oxygen, hemagglutination (HA) was abolished in the presence of mannose and was unaffected by neuraminidase treatment (Fig. 6A), suggesting high numbers of type 1 pili. However, WT UTI89 grown under fermentative conditions exhibited lower HA titers that were inhibited by both mannose and by neuraminidase treatment (Fig. 6B), indicating that the observed agglutination was mediated by both type 1 and S pili. Given the inverse relationship between these two chaperone usher pathway (CUP) pili systems, the observable increase in S pili-mediated agglutination under fermentative growth conditions is an orthologous approach to demonstrate the down-regulation of type 1 pili in response to the lack of oxygen.
WT UTI89 grown anaerobically in the presence of nitrate exhibited overall lower HA titers compared to semi-aerobic and fermentative conditions (Fig. 6C). However, this agglutination was inhibited by mannose and was not significantly impaired by neuraminidase treatment, confirming the de-repression of type 1 pili expression by addition of nitrate and the subsequent down-regulation of S pili. UTI89_LON exhibited an HA profile that was similar to WT UTI89, suggesting that when the fim promoter is genetically locked in the fimON orientation, it exerts a negative effect thereby repressing S pili expression (Fig. 6). UTI89ΔfimA-H yielded low HA titers under the three growth conditions tested and agglutination was not inhibited by mannose but was abolished when treated with neuraminidase, verifying that pili observed by TEM with the UTI89ΔfimA-H mutant are S pili (Fig. 6 and S7 Fig). These data demonstrate that the inverse relationship previously reported for type 1 and S pili [41] is maintained during growth in the absence of oxygen and that depletion of oxygen does not repress expression of all CUP pili systems.
This work shows MALDI IMS to be a strong analytical technology to study the spatial proteome of intact bacterial biofilms. Using a surface-associated biofilm setup that allowed for the formation of a biomass spanning two environmental niches (liquid versus air), we show that this imaging technology can be applied towards the interrogation of biofilm heterogeneity without a priori knowledge of protein targets of interest. Various mass spectrometric techniques have previously been applied for the study of microbial systems [45]. Laser desorption post-ionization mass spectrometry has been applied to analyze peptides involved in sporulation and bacterial competence [46], and secondary ion mass spectrometry (SIMS) was successfully used to analyze peptides involved in bacterial swarming [47]. MALDI IMS has been used successfully for the analysis of small molecules and metabolites within bacterial communities [48–51]. To date, only one other study has utilized MALDI IMS for the direct analysis of protein species within a bacterial community [52]. M.T. et al. used MALDI IMS for the analysis of peptides and proteins found at the site of interaction between E. coli and Enterococcus faecalis biofilms co-cultured on an agar surface, as well as within each individual biofilm [52]. Other than this initial study, little has been done to define the stratification of proteins within intact biofilms by IMS. Therefore, the application of MALDI IMS for the analysis of the intact spatial proteome of a single-species bacterial community represents an emerging approach that has the potential to offer new insights into the role and regulation of protein stratification within biofilms.
One caveat to MALDI-TOF IMS analyses of intact protein localization is that the species observed are typically limited to those most abundant within the sample or those that crystallize and ionize best with the MALDI matrix selected [25,53,54]. This limitation can restrict the sensitivity and dynamic range of the analytes observed by IMS. In turn, large molecular weight proteins or large polymeric protein complexes vital to biofilm formation, which are harder to ionize by MALDI and detect by time-of-flight mass analysis could be intrinsically excluded from the data. This caveat is exemplified by our curli fiber studies, where FN075 treatment increased the amount of detectable CsgA. Thus, orthologous approaches are still critical for validating MALDI IMS findings.
The profile of protein species observed can be expanded by varying the UV-absorbing matrix used for the analysis and by extending the overall m/z ion range analyzed (i.e. from 2,000–25,000 m/z to 2,000–40,000 m/z, and so on) [55]. The sensitivity of MALDI IMS can be refined further by increasing the spatial resolution at which the biofilm is imaged from the current resolution of 150 μm to as low as 20μm in order to better define stratification of subpopulations. We are currently developing both methods to enhance the number and type of protein species that can be localized within a single biofilm. While our approach clearly did not capture the global biofilm proteome, it simultaneously detected the spatial localization of up to 60 protein species within a single analysis; this represents a significant advancement compared to more traditional antibody- or fluorescent tag-based approaches that have been largely limited in the number of protein species visualized per analysis. In addition, the localization of proteins such as FimA and CsgA, which have been shown to play a crucial role in UPEC biofilm formation and pathogenesis but cannot be epitope-tagged due to their incorporation in macromolecular structures, also highlights the strength of this application.
MALDI IMS analyses revealed that type 1 pili-producing bacteria stratify above curli fiber-producing bacteria within the UPEC glass slide surface-associated biofilms interrogated in our studies (Fig. 3B). Similar UPEC biofilms have been previously shown to consist of an extracellular matrix comprised of curli and cellulose [7,28], with type 1 pili playing an accessory role in biofilm tensile strength [7]. The study by Hung et al., revealed that the bacteria on the air-exposed layer of a floating pellicle biofilm (formed during growth in the same media used in our studies), are morphologically distinct from those at the liquid interface [7]. In the same study, they also reported that disruption of fim-mediated adhesion did not ablate biofilm formation, but rather impaired biofilm integrity through the formation of large holes on the air-exposed side of the biomass [7]. Here, MALDI IMS demonstrated that type 1 pili are produced by the bacteria forming the topmost, air-exposed layer of the biofilm. In our studies, we observed that a pellicle biofilm typically surrounded the UPEC slides cultured for IMS analysis within 72–96 hours of starting the culture. If the slide-associated biofilm analyzed by MALDI IMS, is representative of a cross-section of the growing pellicle biomass biofilm, stratification of type 1 pili observed in surface-associated biofilms by IMS could help to explain the loss in tensile strength upon disruption of fim-mediated adhesion observed by Hung et al. However, it is important to note that the type of surface to which the bacteria adhere and the nutrient or surrounding environmental conditions can alter the genetic expression profiles within the biofilm community. Therefore, we recognize that the conclusions drawn here are representative of biofilms formed on a glass surface in a laboratory setting and may bear differences from cross-sections obtained from floating pellicles.
Bacterial biofilms constitute a serious problem in the healthcare setting. The unique heterogeneous architecture of the biofilm, combined with the composition of a self-secreted extracellular matrix, greatly hampers the penetrance and efficacy of bactericidal drugs and limits treatment options in the case of biofilm-related infection [21]. It is thus imperative to identify new strategies to combat or re-program how bacteria form these multicellular structures. Numerous studies identified the presence of bacterial subpopulations within bacterial biofilms and identified that these subpopulations execute unique “tasks” [56,57]. For example, in the benign B. subtilis biofilms, specific subpopulations produce extracellular matrix while others undergo sporulation [57,58]. Further studies indicated that B. subtilis biofilms are coated with a hydrophobin that renders the biofilm colony impervious to penetration [58]. In E. coli and other pathogens, metabolically inactive “persister” cells within the biofilm re-seed the infection upon cessation of antibiotic treatment [8,56,59]. Identifying the spatial proteome of biofilms may uncover markers for distinct subpopulations, thereby aiding in the development of new strategies for thwarting biofilm formation.
Our analyses so far revealed that induction of type 1 pili expression likely occurs on the topmost layer of the imaged biofilm due to the increased oxygen levels in this region. Previous studies reported that UPEC strains rely on the TCA cycle during infection [39,60] and that TCA cycle perturbations lead to a repression of fim gene expression and abrogation of intracellular bacterial community formation [39]. The studies described here show that there are at least two regulatory mechanisms that control expression of type 1 pili in the absence of oxygen; one that exerts its regulatory effect by influencing the fim promoter switch and another that acts independently of the fim promoter switch. Both of these mechanisms are engaged under fermentative growth, strongly suggesting that loss of the ability to use the electron transport processes imposes an energetic cost to the bacteria and necessitates the down-regulation of energetically expensive structures.
In probing the basis of these mechanisms, we have found that under fermentative conditions, there is no significant change in steady-state mRNA transcripts of the two main fim recombinases FimB and FimE (S8 Fig). We have also ruled out the involvement of the Anaerobic Respiration Control (Arc) two-component system (S8 Fig). It is likely that the effects on the phase-state of the fim promoter result from effects on the function of FimB and/or FimE as previously described [61]. Muller et al. elegantly demonstrated that CRP impacts fim gene expression by interfering with FimB function and repressing the expression of Lrp [61]. Other studies indicated that mutants deleted for the global regulator FNR had increased levels of Lrp under anaerobic growth conditions, suggestive of FNR down-regulating lrp expression in the absence of oxygen [62,63]. In the UPEC strain CFT073, Barbieri et al. have demonstrated that deletion of FNR suppresses expression of the FimB recombinase under atmospheric conditions [63]. We are currently investigating the involvement of FNR on modulating fim promoter switching in UPEC strain UTI89.
Use of alternative electron acceptors affords E. coli the ability to continue the electron transport processes under a variety of growth conditions, extending the range of environmental conditions they can withstand. Here we show that while incorporation of an alternative terminal electron acceptor (nitrate) partially preserved piliation in cells that had the promoter fimON, it was unable to restore production of type 1 pili in cells with the promoter in the fimOFF orientation. We have attributed this effect to the ability of nitrate to serve as an alternative terminal electron acceptor. However, it is important to note that nitrate itself, as well as byproducts of nitrate respiration, specifically nitric oxide (NO), can also serve as a signaling molecule within the biofilm community [64–66]. NO has also been shown to have anti-biofilm abilities, suggesting possible role within biofilm signaling and maintenance [67]. We are currently in the process of confirming our results and examining the impact of the other preferred alternative terminal electron acceptors of E. coli (DMSO, TMAO, and fumarate), on type 1 pili expression under oxygen-deplete conditions.
Overall, the results of our nitrate studies are in agreement with our previous studies, in which a non-functional TCA cycle threw the fim switch in the fimOFF orientation [39]. Pathogenic extra-intestinal E. coli strains, such as UPEC, typically thrive in the gastrointestinal tract of humans and other warm-blooded animals where oxygen is limited. As UPEC exit the gut and ascend the urethra to eventually colonize the urinary tract, they undergo multiple metabolic transitions between aerobic and anaerobic growth states. Each of these transitions is accompanied by fluctuations in oxygen tension from strictly anaerobic to highly oxygenated, to semi-aerobic. The bacterial cells respond to these fluctuations by modulating central metabolic pathways for carbon and energy flow, which in turn impact expression of a battery of targets including virulence factors. Together with previous reports [39,60], the studies described here corroborate a direct link between respiration state and the expression of adhesive fibers that has multiple regulatory checkpoints, possibly to account for the diverse fluctuations in oxygen tension encountered by UPEC. Our study also suggests that oxygen gradients determine fiber stratification within the biofilm, which may contribute to overall integrity.
Collectively, our studies used MALDI IMS to begin to define the spatial stratification of distinct bacterial subpopulations within UPEC biofilms based on differential protein expression profiles. Extrapolating from observations made by MALDI IMS, we discovered that type 1 pili-producing bacteria constitute the uppermost layer of UPEC biofilms under the conditions tested, and we identified two new UPEC regulatory mechanisms that control the expression of type 1 pili in response to oxygen and/or bacterial respiration state. These findings highlight how MALDI IMS can drive the identification and characterization of biofilm subpopulations, leading to a greater understanding of their role and regulation within the biofilm.
For these studies we used the UPEC cystitis isolate UTI89 [24]. Previously constructed UTI89 mutants used in this study are UTI89ΔfimA-H (gift from Dr. Scott Hultgren); UTI89_LON [38]; and UTI89ΔarcA (gift from Dr. Matthew Chapman). UTI89ΔhupA and UTI89ΔyahO were created using the previously established λ Red recombinase methods [68] and the following primers (Integrated DNA Technologies): hupA_Fwd (5’–TTACTTAACTGCGTCTTTCAGTGCCTTGCCAGAAACAAATGCCGGTACGTGTGTAGGCTGGAGCTGCTT–3’) / hupA_Rev (5’-ATGAACAAGACTCAACTGATTGATGTAATTGCAGAGAAAGCAGAACTGTCCATATGAATATCCTCCTTAG-3’); yahO_Fwd (5’-ATGAAAATAATCTCTAAAATGTTAGTCGGTGCGTTAGCGTTTGCCGTTACGTGTAGGCTGGAGCTGCTTC-3’) / yahO_Rev (5’-TTACTTCTTCTTATAAATATTTGCCGTGCCGTGAATCTTATTGTCAGTTTCATATGAATATCCTCCTTAG-3’).
All strains were grown overnight in Lysogeny broth (LB) (Fisher), pH 7.4, at 37°C with shaking, unless otherwise specified. Overnight cultures were then sub-cultured in 1.2x Yeast-Extract/Casamino Acids (YESCA) broth [43]. Bacterial suspensions were then dispensed in 50 mL conical tubes containing ITO-coated glass slides (Delta Technologies) and cultured for 48 hours at room temperature. After culture, slides were removed, rinsed with water to remove non-adherent bacteria and stored at -80°C until analysis.
Biofilms were quantified as previously described [43]. Crystal violet stained biofilms were removed from ITO slides using 35% acetic acid and transferred to 96-well plates for absorbance readings. Absorbance at 570 nm was determined using a BioRad Model 680 microplate reader (BioRad). Data are presented as the average absorbance from at least three independent experiments. Statistical analysis was performed using a two-tailed unpaired Student’s t-test (GraphPad Prism 6).
Scanning electron microscopy (SEM). Bacterial biofilms grown as described for MALDI IMS were treated for SEM as previously described [69]. Samples were dried at the critical point, mounted onto aluminum sample stubs and sputter coated with gold-palladium. A small strip of silver paint was applied to the sample edge, and biofilms were imaged with an FEI Quanta 250 Field-emission gun scanning electron microscope (FEI). At least two biological replicates were imaged for each sample preparation and representative images were collected.
Transmission electron microscopy (TEM). TEM analyses were performed as outlined previously [40]. Briefly, 100 μL of normalized bacterial cultures (OD600 = 1.0) from each condition were centrifuged at 4,000 rpm for 10 minutes and resuspended in 50 μL of TEM fixative (2.5% glutaraldehyde in 100mM sodium cacodylate (Electron Microscopy Sciences)) for 1 hour at room temperature. Samples were then deposited onto glow-discharged formvar-/carbon-coated copper grids (Electron Microscopy Sciences) for 60 seconds and stained with 1% uranyl acetate for 90 seconds. Samples were then analyzed on a Phillips/FEI T-12 Transmission Electron Microscope (FEI).
Immuno-fluorescence by Super-resolution Structured Illumination Microscopy (SIM). The α-CsgA antibody was provided by Dr. Matthew Chapman at the University of Michigan. UPEC biofilms were grown for 48 hours as previously described. Biofilms were fixed in 4% paraformaldehyde in phosphate-buffered saline (PBS) for 30 minutes at room temperature and blocked in 5% BSA overnight at 4°C. Biofilms were immuno-stained with α-CsgA (1:1000) for 1 hour at room temperature, followed by 3 washes in PBS and secondary detection with Alexa Fluor-555 goat anti-rabbit (1:1000) (Life Technologies) for 1 hour at room temperature. Samples were washed 3 times in PBS and mounted under a 1.5 size coverslip (Fisher Scientific) using ProLong Gold antifade reagent containing DAPI for DNA counterstain (Life Technologies). Cells were imaged using a GE/Applied Precision DeltaVision OMX in SIM mode with 1.516 immersion oil at 63X magnification. Post-data acquisition processing was performed using SoftWorx for OMX. Images were processed for contrast enhancement and cropping in Photoshop. With the exception of x-y sections (z stacks), images are shown as maximum intensity projections through the entire imaged area (ranging from 3–6 μm in z, 40 μm in x-y). Videos depicting three-dimensional reconstruction of biofilms were generated using the Volume Viewer in Progressive mode in SoftWorx for OMX.
Surface analysis was performed on crystal violet stained biofilms using a Zeta-20 True Color 3D Optical Profilometer (Zeta Instruments) at 20x magnification. Fifty microns were z-stacked to create the profiles at 0.2 microns/step. Images were reconstructed using a 10% optical overlap in stitching. Optical images of crystal violet stained biofilms were obtained using a Leica SCN400 Digital Slide Scanner (Leica Microsystems) at 20x magnification in manual bright field mode.
Biofilms grown on ITO-coated glass slides were washed to remove interfering salts and lipids in sequential 30-second washes of 70, 90, and 95% HPLC-grade ethanol (Fisher Scientific). Matrix comprising 15 mg/mL 2,5-dihydroxybenzoic acid (DHB) (Fisher Scientific) and 5 mg/mL α-Cyano-4-hydroxycinnamic acid (CHCA) (Sigma-Aldrich) was applied using a TM-Sprayer (HTX Imaging), and samples were vapor rehydrated with 10% acetic acid. Samples were analyzed using a Bruker Autoflex Speed mass spectrometer (Bruker Daltonics) in linear positive ion mode. Each pixel contains an average of 200 spectra. Images were collected at 150 micron (μm) lateral resolution. Data were analyzed using FlexImaging 3.0 Build 42 (Bruker Daltonics). Datasets were normalized to total ion current unless otherwise indicated. Ion intensity maps were extracted for each range of interest and were plotted using the maximum intensity within the range. (Detailed MALDI-TOF IMS methods are found in S1 Methods).
To identify 48-hour UPEC biofilm m/z ion species observed by IMS, multiple slide-associated biofilms were lysed and pooled together. Lysates were sonicated, centrifuged, and supernatants dried by vacuum centrifugation (Thermo Scientific). Samples were resuspended and fractionated using C8 (Grace Vydac) or C18 (Phenomenex) reversed-phase high performance liquid chromatography (HPLC) (Waters). Fractions were analyzed for m/z ions corresponding to those observed in the IMS analyses, subjected to in-solution tryptic digestion, and submitted to the Vanderbilt University Mass Spectrometry Research Center Proteomics Core for LC-MS/MS identification (Detailed methods in S1 Methods). For validation of FimA protein identification, 48-hour biofilms of the UTI89ΔfimA-H were cultured as described above and analyzed by MALDI IMS. For validation of HupA and YahO protein identifications, 48-hour static liquid cultures (in 1.2x YESCA) of UTI89ΔhupA and UTI89ΔyahO were grown. After 48 hours, an aliquot of liquid culture was removed and pelleted. Pellets were then lysed with a volume of 35% acetic acid, and centrifuged to pellet debris. Lysates were then analyzed by MALDI-TOF MS (Bruker Daltonics) using the same matrix and parameters for IMS analyses.
FN075 was prepared and characterized as described previously [29,70]. UPEC biofilms were cultured as described above for 24 hours. After 24 hours the preformed biofilm was treated with either 125 μM FN075 dissolved in 100% dimethyl sulfoxide (DMSO), an equivalent volume of 100% DMSO (vehicle control), or an equivalent volume of fresh YESCA media (negative control) and allowed to develop for another 24 hours. Slides were then removed and processed as described above. Biofilms were quantified and analyzed by MALDI IMS as described above.
WT UTI89 and mutant strains were cultured under media and growth conditions listed in Supplemental Table 1 (S1 Table). Cultures starting fimOFF were begun from overnight shaking cultures, and cultures starting fimON were begun from overnight statically grown cultures, both in LB media at 37°C. Oxygen-deplete cultures were grown in an anaerobic chamber maintained at 0% oxygen with between 2–3% hydrogen. Alternative terminal electron acceptor samples were treated with 40mM sodium nitrate (NaNO3) (Sigma-Aldrich). All cultures were grown for 48 hours to mimic biofilm growth conditions used in IMS analyses. After 48 hours, cultures were normalized to an OD600 of 1.0 with sterile PBS for phase assay and immunoblot analysis.
Phase assays were performed as previously described [35] using 100 ng of genomic DNA, or an aliquot of normalized cells (OD600 1.0) and with the following modifications: Primers Phase_L (5’-GAGAAGAAGCTTGATTTAACTAATTG-3’), and Phase_R (5’-AGAGCCGCTGTAGAACTCAGG-3’) were used and the PCR was performed using the following parameters: 95°C—5min, 30 cycles (95°C—45sec, 50°C—20sec, 72°C—45sec), 72°C—5min. To determine the proportion of the population fimON vs. fimOFF, mean pixel intensity of the bands at 489 bp (fimON) and 359 bp (fimOFF) was determined within each sample using Adobe Photoshop CS6 (Adobe Systems). Background taken from a blank area of the gel at a position equivalent to each band, was subtracted. The mean intensity of the fimON and fimOFF band for each sample was then summed, and the percentage ON vs. OFF was then determined for each sample. The percentage of each sample fimOFF was then plotted with GraphPad Prism 6 (GraphPad Software Inc.), and statistical analysis was performed using a one-way ANOVA with Bonferroni’s multiple comparisons test.
Immunoblots probing for FimA were performed as previously described [43]. Briefly, cultures were normalized to an OD600 = 1.0 and 1 ml of normalized cultures was pelleted by centrifugation. Normalized cell pellets were suspended 1x Laemmli sample buffer (BioRad) containing 5% 2-mercaptoetahnol (Sigma-Aldrich). Samples were acidified with 1M hydrochloric acid (HCl), heated at 100°C for 10 minutes, and then neutralized with 1N sodium hydroxide (NaOH). Samples were then resolved on a 16% SDS-PAGE gel. Following SDS-PAGE, proteins were transferred to nitrocellulose using the Trans-Blot Turbo Transfer System (BioRad), (7 minute transfer at 1.3A and 25V). Transfer efficiency was verified with Ponceau S (Sigma-Aldrich). Stains corresponding to blots shown in Fig. 5 are included in S6 Fig. Following transfer, membranes were blocked with 5% non-fat milk in 1x TBST overnight at 4°C. After blocking, membranes were washed 2x with 1x TBST and incubated with primary anti-FimA antibody [1:5,000] [43] for 1 hour at room temperature, washed 2x with 1x TBST, and incubated with HRP-conjugated goat—anti-rabbit secondary antibody (Promega) for 30 minutes at room temperature. Following secondary antibody application membranes were washed 3x with 1x TBST, treated with SuperSignal West Pico Chemiluminescent Substrate (Thermo Scientific), and bands visualized on x-ray film (MidSci). Immunoblots probing for CsgA were performed in a similar fashion with the exception that cell pellets were first solubilized in 100% formic acid, which was then evaporated prior to re-constitution in 1x SDS sample buffer, as previously described [29]. The anti-CsgA antibody was used at a 1:10,000 dilution.
Hemagglutination assays were performed as described previously [43]. Guinea pig erythrocytes were obtained from the Colorado Serum Company. Erythrocyte de-sialylation was performed using Clostridium perfringens neuraminidase (New England BioLabs) for 2 hours at 37°C with gentle agitation.
RNA extraction, reverse transcription, and real-time quantitative PCR were performed as previously described [71]. qPCR analysis was performed with three concentrations of cDNA (50 ng, 25 ng, 12.5 ng) each in triplicate for each sample, and internal DNA gyrase (gyrB) levels were used for normalization. The following primers (Integrated DNA Technologies) were used for amplification; fimB_Fwd (5’—GCATGCTGAGAGCGAGTCGGTA—3’), fimB_Rev (5’—GGCGGTATACCAGACAGTATGACG—3’), fimE_Fwd (5’—ATGAGCGTGAAGCCGTGGAACG—3’), fimE_Rev (5’—TATCTGCACCACGCTCAGCCAG—3’), gyrB_L (5’—GATGCGCGTGAAGGCCTGAATG—3’), gyrB_R (5’—CACGGGCACGGGCAGCATC—3’). The following probes (Applied Biosystems) were used for quantitation; fimB (5’– 6FAM-TCATCCGCACATGTTAC-MGBNFQ—3’); fimE (5’—NED-CGGACCGACGCTATAT-MGBNFQ—3’); gyrB (5’—VIC-ACGAACTGCTGGCGGA-MGBNFQ—3’).
|
10.1371/journal.ppat.1000567 | O-Antigen Delays Lipopolysaccharide Recognition and Impairs Antibacterial Host Defense in Murine Intestinal Epithelial Cells | Although Toll-like receptor (TLR) 4 signals from the cell surface of myeloid cells, it is restricted to an intracellular compartment and requires ligand internalization in intestinal epithelial cells (IECs). Yet, the functional consequence of cell-type specific receptor localization and uptake-dependent lipopolysaccharide (LPS) recognition is unknown. Here, we demonstrate a strikingly delayed activation of IECs but not macrophages by wildtype Salmonella enterica subsp. enterica sv. (S.) Typhimurium as compared to isogenic O-antigen deficient mutants. Delayed epithelial activation is associated with impaired LPS internalization and retarded TLR4-mediated immune recognition. The O-antigen-mediated evasion from early epithelial innate immune activation significantly enhances intraepithelial bacterial survival in vitro and in vivo following oral challenge. These data identify O-antigen expression as an innate immune evasion mechanism during apical intestinal epithelial invasion and illustrate the importance of early innate immune recognition for efficient host defense against invading Salmonella.
| The mammalian host recognizes infection by the detection of particular microbial structures. Recognition of these structures leads to activation of host defense effector mechanisms that in turn combat infection. A very potent activating microbial structure is lipopolysaccharide, a cell wall component released by many bacteria such as Salmonella, one of the most frequent causative agents of foodborne infection of the gut. We previously showed that cells lining the gut surface require uptake of bacterial lipopolysaccharide for its detection. The functional consequence of lipopolysaccharide uptake, however, was unknown. Here, we demonstrate that the uptake of lipopolysaccharide released by Salmonella is impaired by its extensive sugar modification. Impaired lipopolysaccharide uptake prevents early activation of host defense mechanisms and thereby allows Salmonella to better survive and proliferate within the host's intestinal cells. Thus, this lipopolysaccharide modification represents a mechanism by which Salmonella impairs recognition by the mammalian host to more efficiently cause infection of the intestinal mucosa.
| Lipopolysaccharide (LPS) is an obligate constituent of the outer membrane of gram-negative bacteria. It is composed of three parts - a conserved lipid A, a short core carbohydrate, and the O-antigen assembled by a variable number of highly polymorphic carbohydrate subunits [1]. The lipid A consists of a hexa-acylated disaccharide. It is the ligand of the innate immune receptor Toll-like receptor (TLR) 4 and represents one of the most potent immunostimulatory molecules. TLR4-mediated LPS recognition provides an important signal for activation of the antimicrobial host defense during bacterial infection [2],[3]. The O-antigen confers resistance to serum complement activation during systemic infection and represents the chemical basis of bacterial serotyping [4].
Due to its amphiphilic character, LPS forms aggregates in watery solution. The serum protein LPS-binding protein (LBP) retrieves LPS from aggregates or intact bacteria and transfers it to the GPI-anchored surface-bound or soluble form of CD14. CD14 in turn presents the LPS molecule to the MD-2/TLR4 receptor complex. Alternatively, LPS bound to soluble MD-2 can bind to the TLR4 receptor facilitating efficient recognition of even minute amounts of LPS [5],[6]. The structural basis of this intense interaction has recently been resolved [7]. Ligand binding induces a conformational change of the TLR4 dimer and leads to signal transduction, transcriptional activation, and the production and secretion of proinflammatory mediators. Beside professional immune cells, also other cell types such as epithelial cells express functionally active innate immune receptors [8],[9]. Lack of TLR4 signaling has been associated with enhanced susceptibility to microbial challenge, increased tissue destruction during mucosal injury and cancerogenesis within the intestinal tract [10],[11],[12].
Strikingly, the subcellular localization of TLR4 has been demonstrated to differ between macrophages and intestinal epithelial cells (IEC) [13]. Myeloid cells harbor TLR4 on the cell surface and ligand recognition and signling occur from the plasma membrane [14]. In contrast TLR4 is restricted to an intracellular compartment in IECs [13],[15]. LPS is rapidly internalized, reaches the TLR4-positive compartment and initiates signal transduction [16]. Although LPS internalization has been noted since many years [13],[14],[17],[18] and the intracellular TLR4 localization has been confirmed in pulmonary, renal and corneal epithelial cells as well as endothelial cells [15], [17], [19]–[22], the functional consequence of the different cellular localization of the TLR4 molecule and the functional role of ligand internalization is unknown.
Here we report a strikingly delayed recognition of wildtype Salmonella as compared to O-antigen deficient Salmonella by IECs but not macrophages. Delayed recognition of wildtype Salmonella is caused by lack of early TLR4-mediated cell activation associated with impaired LPS internalization. Importantly, lack of early epithelial activation significantly promotes intraepithelial bacterial survival and O-antigen expression is linked to enhanced numbers of intraepithelial Salmonella after oral infection in vivo. The data show that O-antigen expression contributes to bacterial virulence during apical epithelial invasion prior to contact with serum complement and illustrate the susceptibility of Salmonella to antibacterial defense activation before it reaches and establishes its protected intracellular niche.
In order to evaluate a possible biological effect of LPS glycosylation on epithelial cell stimulation, differentiated and polarized intestinal epithelial m-ICcl2 cells were coincubated with wildtype Salmonella, isogenic O-antigen deficient mutants, or their respective complemented strains. The waaG (rfaG) gene encodes a UDP-glucose:(heptosyl)LPS α1,3-glucosyltransferase and mutants exhibit a rough Rd1 LPS phenotype with only the inner core sugars attached to the lipid A molecule [23],[24]. waaL (rfaL) encodes the O-antigen ligase, the last step in the LPS biosynthesis. WaaL functions within the periplasmic space at the cytoplasmic membrane to ligate the presynthesized O-antigen chain onto the lipid A core molecule [1],[24]. waaL mutants therefore express the complete core sugars but completely lack the O-antigen (Ra LPS). O-antigen expression was confirmed using silver staining of LPS extracts (Fig. S1A).
Cellular activation was evaluated using (i) visualization of nuclear translocation of the NF-κB subunit p65/RelA, (ii) a stably transfected transcriptional NF-κB luciferase reporter construct, and (iii) quantification of the secreted proinflammatory chemokine MIP-2. Strikingly, a significant difference in the kinetics of cellular activation was recognized after challenge with wildtype and LPS mutant strains. Whereas no difference in the overall magnitude of epithelial cell activation was noted, waaL mutants induced a significantly earlier p65/RelA translocation (Fig. 1A, the earliest detectable p65/RelA translocation is marked with arrows) and an accelerated course of chemokine secretion and NF-κB reporter gene transcription in epithelial cells as compared to wildtype Salmonella (Fig. 1B and C). This difference in p65/RelA translocation (Fig. 1D), chemokine secretion (Fig. 1E), and NF-κB reporter gene activation (Fig. 1F) was similarly observed using waaG- and waaL-deficient mutants and reversed by the complemented strains carrying an expression plasmid encoding the waaG and waaL gene, respectively. Thus lack of O-antigen expression leads to a significantly accelerated recognition of Salmonella by IECs.
A similar delay in epithelial activation by wildtype Salmonella was also noted using heat-killed or UV-treated Salmonella suggesting structural impairment of LPS recognition by the O-antigen rather than O-antigen-mediated active inhibition of epithelial cell activation (Fig. S1B and data not shown). To examine the contribution of TLR4-mediated epithelial cell activation and exclude indirect effects of the O-antigen on cellular activation, the role of TLR4 in Salmonella recognition during coculture with IECs over 6 hours was examined. First the stimulatory activity released in the cell culture medium at bacterial numbers corresponding to a multiplicity of infection (MOI) of 10∶1 (106 CFU/mL) and 1∶1 (105 CFU/mL) during one hour was completely inhibited by addition of the LPS-inhibiting agent polymyxin B (Fig. 2A). Also, inhibition of Tlr4 (Fig. 2B) or Myd88 (Fig. 2C) expression by small interfering (si) RNA technique inhibited epithelial activation by Salmonella to a similar degree as epithelial activation by LPS. In fact, early recognition of the O-antigen deficient waaL mutant Salmonella was almost abolished in epithelial cells treated with Tlr4 siRNA (Fig. 2D). In contrast, inhibition of TLR2, TLR5, or TLR9 expression did not reduce the epithelial response to bacterial exposure (Fig. 2E). Consistently, no early epithelial stimulation was observed after apical exposure to other innate immune receptor ligands released by gram-negative bacteria such as flagellin, di- or tri-acylated lipopeptides, or CpG oligonucleotides (data not shown). The important role of LPS for the observed effect of delayed recognition of wildtype Salmonella was finally confirmed using LPS purified from O-antigen positive (smooth-type LPS, sLPS) as well as O-antigen negative (rough-type LPS, rLPS) Salmonella. Indeed, a similar pattern as compared to exposure to whole wildtype and mutant Salmonella with delayed epithelial activation in response to smooth LPS at early time points (Fig. 2F and G) but similar levels of epithelial activation at later time points (Fig. 2H) was observed. Thus epithelial activation early during the time course of coculture is predominantly caused by TLR4-mediated cell stimulation. The observed delay in the recognition of smooth Salmonella is not related to an O-antigen-mediated suppressive effect on early epithelial activation but rather caused by an inhibitory effect of the O-antigen on LPS recognition by epithelial TLR4.
Myeloid cells like macrophages carry the TLR4 receptor complex on the cell surface and signaling is initiated at the plasma membrane [14]. This is in contrast to IEC lines and isolated primary IECs that exhibit restriction of the TLR4 molecule to an intracellular compartment [13],[15]. In these cells, receptor activation requires ligand internalization and signaling is initiated at the intracellular TLR4-positive compartment [17]. Using the protein delivery reagent PULSin in combination with TLR4/MD2 blocking antibodies, the different receptor localization could be functionally demonstrated. Whereas activation of myeloid cells was readily blocked by addition of the blocking anti-TLR4 antibody MTS510 to the cell culture medium, antibody-mediated inhibition of epithelial activation was only observed in the presence of the protein delivery reagent PULSin (Fig. 3A and B). Interestingly, both, sLPS - as well as rLPS - stimulated RAW 264.7 cells at early time points to a similar degree and with very similar kinetics (Fig. 3C and D). Also, early p65/RelA nuclear translocation was similarly induced in macrophages by all strains, wildtype as well as ΔwaaL and ΔwaaG mutant Salmonella as well as the respective complemented strains (Fig. 3E and F). Thus the delayed recognition of wildtype sLPS as compared to rLPS is restricted to IECs that are devoid of plasma membrane expression of TLR4 and rely on ligand internalization. Yet we cannot exclude that macrophage activation additionally occurs by LPS release during phagocytosis.
Genes encoded by the so called Salmonella pathogenicity island 1 (SPI-1) confer the ability to invade epithelial cells. Bacterial invasion is induced by direct translocation of effector proteins into the host cell cytoplasm which causes actin polymerization and membrane protrusions. Within one hour, this mechanism leads to bacterial internalization and localization within an endosomal compartment named Salmonella containing vacuole (SCV). Initially, the kinetics of Salmonella invasion was examined using constitutive GFP-positive wildtype Salmonella followed by immunostaining with anti-O-antigen Salmonella O4/O5 antibodies without prior cell membrane permeabilization. This technique allows the differentiation of extracellular (simultaneously green and red = orange) and intracellular (green) Salmonella. Exposure of confluent polarized m-ICcl2 cells revealed bacterial invasion starting approximately 20 minutes after challenge with significant numbers of intracellular bacteria at 2 hours after infection (Fig. 4A). Fig. 4B provides a more detailed illustration of the actin-dependent mode of Salmonella invasion at 30 minutes after infection (left panel) and the intracellular localization after 2 hours (right panel). Importantly, the O-antigen-mediated delay in LPS recognition was also observed in invasion-mutants: An isogenic pair of smooth and rough invC-mutants exhibited a similar difference in the kinetics of epithelial stimulation as compared to invasion-competent Salmonella (Fig. 4C and D). Also, hilA- and pho-24 (PhoPc) deficient smooth Salmonella, both significantly impaired in epithelial invasion, exhibited a similar pattern of reduced activation at early time points but cellular stimulation at later time points after infection (Fig. S2A and B). Thus the observed delay in epithelial activation by wildtype bacteria is not dependent on their ability to exhibit an epithelial cell-invasive phenotype but rather result from extracellular ligand exposure.
Viable bacteria continuously release LPS in the surrounding medium. In accordance, significant amounts of 102 EU/mL (approximately 10 ng/mL) LPS were found to be released from viable wildtype Salmonella into the cell culture medium within 30 minutes. The concentration increased up to 103 EU/mL (approximately 100 ng/mL) during the observed time period of 2 hours. No significant difference in the degree of endotoxin release between wildtype and waaL- and waaG-mutants or their respective complemented Salmonella strains was noted (Fig. 5A). As expected, inhibition of CD14 and LBP expression by siRNA significantly impaired LPS and Salmonella-mediated activation of m-ICcl2 cells in accordance with the literature (Fig 5B) [25]. Strikingly, LPS internalization studies using biotinylated rLPS or sLPS preparations revealed a marked difference in the kinetics of ligand uptake. Whereas detectable amounts of rLPS were observed after 30 minutes, wildtype LPS remained undetectable until many hours after exposure (Fig 5C). Previous characterization of intestinal epithelial stimulation with rLPS identified a clathrin- and lipid raft-dependent pathway of LPS internalization and receptor activation [15]. Inhibition of lipid raft formation with filipin previously linked to recognition of rLPS abolished early recognition of rLPS but left the more delayed cellular activation induced by wildtype sLPS unaffected (Fig. 5D). Also, a significant inhibitory effect of clathrin siRNA on early recognition of rLPS was noted (data not shown) and dynamin inhibition by dynasore significantly reduced activation by rough, ΔwaaL Salmonella consistent with this rapid internalization pathway for rLPS uptake (Fig. 5E). These data suggest that qualitative differences in the uptake and intracellular transport mechanism between sLPS and rLPS exist and account for the observed delay in the epithelial recognition of wildtype, O-antigen-positive Salmonella. Similar results obtained using viable invasive and non-invasive Salmonella, heat-killed Salmonella, or purified LPS suggest involvement of plasma membrane-to-Golgi traffic. Yet we cannot exclude that transport pathways from the SCV to the Golgi apparatus are also affected.
To examine the functional consequences of early epithelial activation, a standard Gentamicin protection-assay was performed. Strikingly, both O-antigen negative mutants but not their respective complemented strains exhibited a significantly reduced number of viably intracellular bacteria two hours after infection (Fig. 6A). This was confirmed by immunofluorescence (Fig. 6B) as well as by flow cytometry (Fig. 6C) using GFP expressing wildtype and waaL-deficient Salmonella. Both, the relative number of Salmonella-positive epithelial cells (6.3±0.5% versus 1.7±0.1%, p<0.01) as well as the mean fluorescence intensity (MFI) of positive cells indicating the number of bacteria per cell (MFI 487.8±14.1 versus MFI 343.9±7.4, p<0.01) were significantly enhanced 2 hours after infection with wildtype as compared to waaL-deficient Salmonella (Fig. 6C). Notably, this difference in the number of viable Salmonella was not due to impaired invasion of waaL-deficient Salmonella since similar numbers of intracellular bacteria were obtained 30 minutes after infection (1.4±0.1% versus 1.9±0.1%) (Fig. 6D). In fact flow cytometric quantification of intracellular bacteria after epithelial cell lysis revealed an approximately 2-fold enhanced invasion rate of the O-antigen deficient waaL mutant Salmonella as compared to wildtype bacteria (Fig. S1C and D). The increase of the number and fluorescence intensity of wildtype Salmonella-infected epithelial cells together with the marked clusters of intracellular wildtype Salmonella 2 hours after infection suggest significant intraepithelial proliferation early after invasion. In contrast, no signs of bacterial growth were noted for the waaL-deficient Salmonella strain. In addition, reduced bacterial numbers in epithelial cells did not appear to result from general growth or viability defects of O-antigen deficient Salmonella. Wildtype, ΔwaaL, and ΔwaaG Salmonella as well as the complemented mutants exhibited comparable growth rates in LB medium, or m-ICcl2 cell lysate (Fig. S1E). Also, both wildtype and ΔwaaL or ΔwaaG Salmonella were able to induce persistent intracellular infection in m-ICcl2 cells (Fig. S1F). In accordance with the different phenotype observed in epithelial cells and macrophages (Fig. 1 versus Fig. 3), the intracellular survival illustrated by enhanced fluorescence of infected cells with wildtype Salmonella was only found in epithelial cells. In contrast, the fluorescence of Salmonella-infected RAW 264.7 macrophages was not significantly altered during the first 2 hours of infection, irrespective of the Salmonella strain used (Fig. S2C). Of note, an enhanced internalization of the waaL-deficient Salmonella mutant was observed after macrophage infection in accordance with Ilg et al. [26] (Fig. S2C).
To confirm that the observed antibacterial effect was directly linked to early innate immune recognition and cell activation, epithelial cells were infected with wildtype bacteria in the presence or absence of rLPS. Indeed, the number of intracellular bacteria as measured by invasion assay (Fig. 6E), or immunofluorescence (Fig. 6F) was significantly reduced in rLPS-stimulated epithelial cells illustrating the critical importance of early cell activation to restrict intracellular bacterial growth. Wildtype Salmonella in sLPS-stimulated epithelial cells were significantly less affected (Fig. S2D). The dramatic nature of this antibacterial effect was illustrated by flow cytometric quantification of intracellular bacteria in cell lysate between 30 and 60 minutes after infection. Whereas invasion of naïve epithelial cells allowed immediate intracellular bacterial growth, rLPS-stimulated epithelial cells were able to restrict the number of Salmonella (Fig. 6G and H). Thus early activation of intestinal epithelial cells by O-antigen-deficient Salmonella is associated with significantly reduced intraepithelial survival.
Salmonella has been shown to invade IECs in vivo after oral challenge [27]. Intestinal epithelial invasion from the luminal side occurs without prior contact with tissue macrophages or complement. To examine a possible effect of O-antigen expression on intraepithelial survival in vivo, mice were orally challenged and highly pure IECs were isolated and examined for the presence of viable Salmonella. Similar numbers of intracellular wildtype and waaL-deficient Salmonella were noted at early time points following infection (Fig. 7A). Interestingly, a significant reduction of O-antigen-deficient (waaL) Salmonella as compared to wildtype as well as the respective complemented Salmonella was detected in highly pure IECs later during the course of infection (Fig. 7B). The presence of intraepithelial wildtype Salmonella after oral challenge was also confirmed by immunohistology (Fig. 7C). Thus, lack of O-antigen expression does not influence intestinal epithelial invasion but intraepithelial survival of Salmonella in vitro and in vivo. These results identify O-antigen expression as innate immune evasion strategy to enhance intraepithelial survival. O-antigen expression might thereby promote intraepithelial proliferation and mucosal spread.
S. Typhimurium is one of the leading causative agents of enteritis in humans. Infection is acquired by oral ingestion of contaminated food. In the intestine, Salmonella firmly attaches to the epithelial surface and induces membrane protrusions that surround the bacterium and form an endosomal vesicle called Salmonella-containing vacuole (SCV). This process has been extensively studied in vitro but also confirmed in vivo [27],[28]. Intestinal epithelial invasion from the enteric lumen occurs prior to contact with serum complement or professional phagocytes such as macrophages. It plays an important role in the induction of enteritis and mucosal damage in vivo and thus represents an essential step in Salmonella pathogenesis [29],[30].
Similar to professional immune cells, also intestinal epithelial cells express receptors of the innate immune system, and thus might contribute to recognition of microbial infection and antibacterial host defense during the initial phase of infection. Indeed, the LPS structure was shown to significantly influence epithelial invasion [26],[31]. Also, innate immune recognition via TLR4 was reported to play a significant role in the host defense against Salmonella infection in vivo [10], [32]–[35]. Strikingly, the subcellular localization of TLR4 in myeloid versus epithelial cells is markedly different. Whereas the receptor molecule is situated on the cell surface of macrophages and ligand recognition and cell signaling occurs at the cell membrane, TLR4 in IECs is restricted to the intracellular compartment and ligand recognition requires uptake and intact cell traffic [13],[15],[16]. We could previously show that internalization of rLPS results in significant intracellular accumulation within minutes after exposure [16]. In the present study we show that qualitative differences in the uptake and intracellular transport mechanism between rLPS and sLPS might significantly contribute to immune evasion of wildtype Salmonella during the early phase of mucosal infection. Thus our results for the first time report on a biological consequence of intracellular TLR4 localization in IECs. Since apical invasion of enterocytes by Salmonella occurs prior to contact with serum complement or professional phagocytes such as tissue macrophages, the epithelial specific delay in wildtype Salmonella LPS recognition significantly contributes to bacterial virulence at the mucosal surface in addition to what has been described as serum resistance during systemic spread of the bacteria (Fig. S2E).
LPS is composed of the hydrophobic lipid A, the core polysaccharides, and the highly polymorphic and hydrophilic immunodominant O-antigen [1]. The LPS receptor TLR4 specifically interacts with and recognizes the lipid A part of the LPS molecule. Therefore, even small variations observed in the lipid A structure and their influence on TLR4-mediated recognition have extensively been studied [36]. The O-antigen is not required for the immunostimulatory activity of LPS and variations of the O-antigen and their impact on TLR4-mediated recognition have only recently attracted attention [37],[38]. The O-antigen is composed of up to 100 repetitive structurally variable carbohydrate subunits and the distinction of different O-antigen subunits has been used in the serotyping of various gram-negative bacteria. It is synthesized separately from the rest of the LPS molecule on a lipid carrier by enzymes encoded by the rfb/waa locus. The O-antigen chain is subsequently transferred to the periplasmic space where ligation to the lipid A-core polysaccharide precursor takes place. Only then, the completed LPS molecule is transferred to the bacterial cell surface [39]. The O-antigen is the major determinant of complement resistance and thus represents an important virulence factor [40]. Indeed, gram-negative enteropathogenic bacteria isolated from fecal samples of diseased patients such as Yersinia enterocolitica, Salmonella enterica, Shigella dysenteriae, as well as enterohemorrhagic (EHEC) or enteropathogenic (EPEC) Escherichia coli exhibit long O-antigen chains on their respective LPS molecule. Modifications within the lipid A portion of the molecule have been described to alter the stimulatory potential of LPS [41]. The presence or absence of the O-antigen, however, has not been linked to alterations in the TLR4-mediated signaling cascade leading to MAP kinase and NF-κB activation [38].
Using X-ray diffraction of dried LPS, Kastowsky and collegues estimated the size of the lipid A molecule to measure approximately 2.4 nm in length [42]. Addition of the inner core carbohydrates (corresponding to the LPS produced by the waaG mutant Salmonella) would result in a length of approximately 3.5 nm, addition of the outer core carbohydrates (corresponding to the LPS produced by the waaL mutant Salmonella) in a length of approximately 4.4 nm. Their analysis further suggested an additional length of 1.1 to 1.6 nm per repeating carbohydrate unit of the O-antigen. A full length O-antigen with up to 100 repeating units may therefore extend the molecular length to more than 100 nm. The tertiary structure and orientation of long chain O-antigen in respect to the outer cell membrane is not fully understood [42]. Although the O-antigen might be heavily coiled and allow (and actually favor) some degree of lateral bending [42],[43], addition of this long chain hydrophilic residue might dramatically enhance the spacial extension of the LPS molecule [42]–[44]. In accordance, electron microscopic images from the membrane of gram-negative bacteria suggest that the O-antigen extends from the outer cell membrane for 40–100 nm [43],[44]. The length of the extending O-antigen structures are also illustrated by reports on O-antigen mediated impairment of efficient type III secretion in enteropathogenic Shigella [45]. Taking into account that the inner diameter of clathrin coated vesicles is strictly defined; one explanation for the delayed internalization of smooth LPS by epithelial cells might therefore be its physical size.
Both in vitro as well as in vivo experiments revealed comparable intestinal epithelial invasion by wildtype and O-antigen-deficient bacteria at early time points after challenge. Once inside the epithelial cell, Salmonella is able to interfere with cellular processes of endosomal maturation altering the molecular composition of its surrounding membranous compartment for its own benefit [46]. Well-established virulence determinants such as the PhoPQ regulon and the SPI-2 type III secretion system contribute to this immune evasive behavior [47]. Avoidance of epithelial activation might significantly contribute to bacterial survival since innate immune signaling has been suggested to promote maturation of endosomal compartments and to influence intracellular bacterial proliferation [48],[49]. Indeed the intracellular viability of O-antigen-deficient Salmonella in intestinal epithelial cells in vitro and in vivo was significantly reduced. Previous animal studies have indicated a significant effect of Salmonella O-antigen expression after oral but not intraperitoneal or intravenous infection [50],[51]. Our results provide an explanation for these findings and demonstrate that the O-antigen-modification of LPS significantly contributes to mucosal immune evasion and thus bacterial virulence in the intestine. Our data further point towards a role of intestinal epithelial infection for enteric bacterial multiplication and fecal excretion and thereby transmission of enteropathogenic bacteria like Salmonella.
In conclusion, we for the first time provide a functional consequence of internalization-dependent ligand recognition by TLR4 as compared to surface recognition in myeloid cells. We demonstrate that O-antigen modification of Salmonella LPS hinders rapid epithelial internalization and delays TLR4-mediated recognition. Evasion from early innate immune activation of IECs markedly enhances intracellular proliferation of wildtype Salmonella. This novel immune evasion mechanism might thus significantly contribute to mucosal virulence of enteropathogenic bacteria.
Animals were handled in strict accordance with good animal practice as defined by the relevant local animal welfare bodies, and all animal work was approved by the appropriate committee (Landesamt für Lebensmittelsicherheit und Verbraucherschutz, Oldenburg, 07/1334).
Antibodies against p65/RelA were purchased from Santa Cruz Biotechnology, Inc. (Santa Cruz, CA). The rat monoclonal anti-TLR4/MD-2 antibody (MTS510) and the mouse monoclonal anti-O4 and anti-O5 antigen antibody were kindly provided by K. Miyaki (Saga Medical School, Nabeshima, Saga, Japan) and M. Kim (Kim Laboratories Inc., Champaign, IL), respectively. All fluorophore-conjugated secondary antibodies and Cy5-conjugated streptavidin were from Jackson ImmunoResearch (West Grove, PA). Filipin, dynasore, and polymyxin B were purchased from Sigma (Taufkirchen, Germany). Plasmid DNA was prepared using the EndoFree Plasmid kit from Qiagen (Hilden, Germany). High purity grade smooth and rough form LPS were purchased from List Biochemicals (Campbell, CA) and Alexis Biochemicals (Lausen, Switzerland) and tested for its TLR4-specific activity using Tlr4-deficient C57BL/10ScN-Tlr4lps-del/JthJ mice (The Jackson Laboratory, Bar Harbor, USA) (Fig. S1G–I). LPS was biotinylated using EZ link biotinylation kit from Thermo Scientific (Rockford, IL). Endotoxin was quantified using the chromogenic QCL-1000 Limulus amebocyte lysate system from Lonza (Basel, Switzerland). All siRNA probes used (Tlr2, Tlr4, Tlr5, Tlr9, Myd88, Cd14, Lbp, Clathrin and control siRNA) were from Qiagen (Hilden, Germany). For plasmid transfection, siRNA transfection, and intracellular antibody delivery Lipofectamin 2000 (Invitrogen, Carlsbad, CA), INTERFERin (Polyplus Transfection, New York, NY) and PULSin (Polyplus Transfection), respectively, were used according to the manufacturer's instructions. Cell culture reagents were purchased from Invitrogen. All other reagents were obtained from Sigma (Taufkirchen, Germany) if not stated otherwise.
Salmonella enterica subsp. enterica sv. Typhimurium (S. Typhimurium) ATCC 14028 was used as wildtype strain. Isogenic mutant strains (ΔwaaL and ΔwaaG) were generated by Red recombinase mediated deletion and chromosomal insertion of a Kanamicin antibiotic resistance cassette as described elsewhere (Zenk et al., submitted). The construction of plasmids for the complementation of mutant strains is described in (Zenk et al. submitted). The LPS profiles of the various strains were analyzed using SDS-PAGE and silver staining (Fig. S1A). Bacteria were incubated at 70°C for 10 min to produce heat-killed Salmonella. The non-invasive isogenic pho-24 (PhoP constitutive) and ΔhilA mutants were a generous gift from Mikael Rhen (Karolinska Institute, Stockholm) and the isogenic ΔinvC and ΔinvC ΔwaaL double mutants were generated as described above. HilA is a central regulator of SPI-I mediated epithelial cell invasion, the PhoP/PhoQ two component system is a central regulator in Salmonella virulence, and InvC is required for type III secretion of SPI-1-encoded virulence determinants. The phenotype of the non-invasive pho-24 mutant is designated PhoPconstitutive (PhoPc). All three mutants exhibit strongly impaired epithelial invasion. Fluorescent bacteria were generated by transformation with a constitutively GFP expressing plasmid. For all experiments, bacteria were routinely grown in Luria-Bertani (LB) broth, supplemented with antibiotics if required. Murine small intestinal epithelial m-ICcl2 cells and m-ICcl2 cells stably expressing a NF-κB luciferase reporter construct were cultured as described previously [52]. RAW 264.7 macrophages were purchased from ATCC and cultured in RPMI 1640 medium (Invitrogen) supplemented with 20 mM Hepes, 2 mM L-glutamine, and 10% FCS.
For all coculture experiments, wildtype or mutant Salmonella were grown overnight at 37°C, diluted 1∶10 and subcultivated with mild agitation at 37°C, until mid-logarithmic growth was reached (OD600: 0.5). Bacteria were adjusted by dilution, added to polarized and differentiated intestinal epithelial m-ICcl2 cells at a multiplicity of infection (MOI) of 10∶1 and centrifuged at 300×g for 5 min. Following incubation for one hour, the medium was replaced with fresh medium supplemented with 50 µg/mL Gentamicin. Cell culture supernatants, as well as cell lysates, were collected after the indicated periods of time and stored at −20°C. The chemokine MIP-2 was analyzed using a commercial ELISA from Nordic Biosite (Täby, Sweden). Luciferase in cell lysates was quantified using a luciferase reporter kit (Promega, Madison, WI). Pharmacological inhibitors were added to the cell medium 30 min prior to stimulation. NO production was determined by measurement of nitrite in cell culture supernatant using Griess reagent [53]. Stimulation with mouse recombinant TNF (R&D Systems GmbH, Wiesbaden, Germany) was performed at 100 ng/mL. To quantify bacterial invasion coculture for one hour was followed by one hour incubation in fresh cell culture medium, supplemented with 50 µg/mL Gentamicin. After washing, cells were lysed in H2O/Tween 0.1% and the number of intracellular bacteria was determined by serial dilution and plating. Specific or control siRNA was transfected with INTERFERin at a final concentration of 1 or 10 nM 48 hours prior to functional analysis. TLR4 blocking experiments using the rat monoclonal anti-TLR4/MD2 antibody MTS510 were conducted in the presence or absence of the protein delivery reagent PULSin according to the manufacturer's recommendations. R-Phycoerythrin (R-PE) is a fluorescent protein to visualize efficient protein transfection by PULSin (data not shown).
m-ICcl2 or RAW 264.7 cells were grown on 8-well chamber slides (Nunc, Rochester, NY), and infected with constitutively GFP-expressing wildtype S. Typhimurium or isogenic mutants as indicated at a MOI of 10∶1, or exposed to biotinylated LPS (100 ng/mL), and incubated for the specifically mentioned period of time. Visualization of the cellular distribution of p65/RelA was performed as previously described [15]. Discrimination of extra-and intracellular bacteria was achieved using Salmonella strains carrying a GFP expression construct (green) in combination with a mixture of two mouse monoclonal anti-Salmonella O-antigen (anti-O4 and anti-O5) antibodies visualized with a Texas-Red-conjugated anti-mouse secondary antibody (red) in the absence of cell permeabilization. Due to impaired penetration of the anti-Salmonella antibodies into the cells, intracellular bacteria appear green, whereas extracellular bacteria exhibit an orange (green plus red) color. Biotinylated LPS was detected using Texas-Red conjugated Streptavidin (Jackson ImmunoResearch). For permeabilization of eukaryotic cell membrane, saponin was adjusted to a final concentration of 0.5%. After the indicated time periods, cells were fixed in 5% PFA and counterstained with MFP488- or MFP647-phalloidin (MoBiTec, Göttingen, Germany). Cells were subsequently mounted in DAPI containing Vectashield (Vector Laboratories) and visualized using an ApoTome-equipped Axioplan 2 microscope connected to an AxioCam Mr digital Camera (Carl Zeiss MicroImaging, Inc. Jena, Germany). Flow cytometric detection of intracellular GFP-expressing bacteria in intact epithelial cells was carried out in Trypsin-EDTA 0.05% treated, fixed m-ICcl2 cells using a FACS Calibur® apparatus (BD Pharmingen). In addition, flow cytometry was used to quantify the number of GFP-expressing bacteria in cell lysates. To standardize the volume examined, a defined quantity of Cy5-labelled particles was added to all samples and the data acquisition on GFP-positive bacteria (recorded in channel Fl-1) was limited until a simultaneously recorded number of 10.000 events in the far red channel (Cy5, Fl-4) was reached.
Mice were purchased from Charles River Breeding Laboratories (Sulzfeld, Germany) and housed under specific pathogen-free conditions. 6–8-week-old female Balb/c mice were orally infected with 1×108 CFU S. Typhimurium in 20 µl phosphate-buffered saline (PBS). 4 and 24 h post infection, mice were sacrificed and the small intestine was removed. Highly pure IECs (>98% E-cadherin+/CD45−) were isolated using a recently described protocol [15], incubated in the presence of Gentamicin (50 µg/mL) and washed. The number of viable intraepithelial bacteria was determined by serial plating.
All experiments were performed at least three times and results are given as the mean±SD of one representative experiment. Statistical analyses were performed using the Student's t test. A p value<0.05 was considered significant.
|
10.1371/journal.pcbi.1006586 | Moth olfactory receptor neurons adjust their encoding efficiency to temporal statistics of pheromone fluctuations | The efficient coding hypothesis predicts that sensory neurons adjust their coding resources to optimally represent the stimulus statistics of their environment. To test this prediction in the moth olfactory system, we have developed a stimulation protocol that mimics the natural temporal structure within a turbulent pheromone plume. We report that responses of antennal olfactory receptor neurons to pheromone encounters follow the temporal fluctuations in such a way that the most frequent stimulus timescales are encoded with maximum accuracy. We also observe that the average coding precision of the neurons adjusted to the stimulus-timescale statistics at a given distance from the pheromone source is higher than if the same encoding model is applied at a shorter, non-matching, distance. Finally, the coding accuracy profile and the stimulus-timescale distribution are related in the manner predicted by the information theory for the many-to-one convergence scenario of the moth peripheral sensory system.
| Sensory neural systems of living organisms encode the representation of their environment with remarkable efficiency. We study the dynamic coding of naturalistic olfactory stimulation by pheromone-specific antennal neurons. The analysis reveals that the representation is optimal from several complementary information-theoretic perspectives. (1) Pheromone encounters are best detected if the concentration follows the naturally intermittent time course. (2) Antennal neurons dynamically adjust to the local stimulus statistics. (3) The coding accuracy profile and the stimulus-timescale distribution are in the relationship predicted by both information theory and statistical estimation theory.
| Orienting towards food and mates in insects is an olfactory-controlled behavior that relies on detecting odorant molecules delivered from the source. Atmospheric turbulence causes strong mixing of air and creates a wide spectrum of spatio-temporal variations in the signal (Fig 1). The largest eddies may be hundreds of meters in extent and take minutes to pass a fixed point, while the smallest spatial variations could have a size of less than a millimeter and last for milliseconds [1–3]. The mean concentration of the odorant decreases with distance from the source, however, a signal with a large instantaneous magnitude can be found in a wide range of distances from the source, though their frequency decreases with distance [1]. Hence, an important characteristic of the detected signal is its intermittency, i.e., the fraction of time during which the odorant can be detected [1, 4, 5]. Rapid behavioral responses of male moths tracking plumes in turbulent flows [6] and the ability of neurons from the first two layers of the olfactory system to encode the temporal dynamics of pheromone plumes at any distance from the source [7] suggest efficient coding of olfactory plume dynamics.
Recently, the statistical distributions of odorant fluctuations was described [3], namely the statistics of time intervals with the presence of an odorant at a given point in space, denoted as whiffs, and intervals when the odorant concentration is zero, blanks. The distributions of whiff and blank durations change with the distance of a detector from a source (Fig 2) and provide together an important statistical description of the local spatio-temporal properties of the pheromone plume.
The local statistics of many natural stimuli differs from the average global distribution, and the limited coding range of neurons does not cover the wide range of all possible stimulus values [8–10]. The efficient coding hypothesis [11] states that neuronal responses are adjusted, through evolutionary and adaptive processes, to optimally encode such stimulus statistics that match the local sensory environment [12–15]. The hypothesis thus predicts that coding accuracy is highest for the most commonly occurring events to minimize overall decoding error. Such situations have been reported in auditory coding of sound intensity [8, 9, 16, 17], of interaural level differences [18] and time differences [19], and also for primary visual cortex [10] and primary somatosensory cortex [20]. To the best of our knowledge, an analogous study has not been done yet in odor detection, partially due to the difficulties associated with the description of the natural stimulus statistics and its changes [2–4, 21].
In this work, we study how pheromone-sensitive olfactory receptor neurons (ORNs) adjust their responses to the local stimulus statistics (Fig 1). Our results show that ORN responses are adjusted in such a way that pheromone encounters are encoded best after blanks that have the most common duration. We also found that the average accuracy of pheromone detection is better if an encoding scheme is adapted to the stimulus statistics of a particular distance from the source than if the same scheme is applied closer to the source. In addition, ORNs’ coding properties support an idea of efficient population information transmission from the ORNs to the antennal lobe neurons.
Experiments were performed with laboratory-reared adult males of Agrotis ipsilon fed an artificial diet [22]. Pupae were sexed, and males and females were kept separately at 22 °C in an inversed light-dark cycle (16 h–8 h light-dark photoperiod). Adults were given access to 20% sucrose solution ad libitum. Experiments were performed on virgin 4- or 5-day-old (sexually mature) males.
Insects were restrained in a Styrofoam block with the head protruding. One antenna was fixed with adhesive tape on a small support. Electrodes were made from electrolytically sharpened tungsten wires (TW5-6, Science Products, Hofheim, Germany). The recording electrode was inserted at the base of a long pheromone-responding sensillum trichodeum located on an antennal branch. The reference electrode was inserted in the antennal stem. The electrical signal was amplified (×1000) and band-pass filtered (10 Hz–5 kHz) with an ELC-03X (npi electronic, Tamm, Germany), and sampled at 10 kHz via a 16-bit acquisition board (NI-9215, National Inst., Nanterre, France) under Labview (National Inst.). One sensillum was recorded per insect.
ORNs were stimulated with the major sex pheromone component of A. ipsilon, (Z)-7-dodecenyl acetate (Z7-12:Ac). Pheromone stimuli were diluted in decadic steps in hexane and applied on a filter paper introduced in a Pasteur pipette at doses ranging from 10−6 to 100 ng. The antenna was constantly superfused by a humidified and charcoal-filtered air stream (70 L ⋅ h−1). Air puffs (10 L ⋅ h−1) were delivered through a calibrated capillary (Ref. 11762313, Fisher Scientific, France) positioned at 1 mm from the antenna and containing the odor-loaded filter paper (10 × 2 mm). An electrovalve (LHDA-1233215-H, Lee Company, France) was controlled by custom Labview programs reading sequences generated using Matlab scripts. The time resolution of the sequences was 1 ms. The characteristic response time of the valves, i.e. the time to go from open to close (close to open) is <5 ms.
Durations of whiffs (puffs) and blanks were set to mimic the turbulent dynamics of the odorant plume in a real environment according to the model by Celani et al. [3] at 5 virtual downwind distances from the pheromone source (d = 8, 16, 32, 64 and 128 m). The virtual crosswind distance was always 0, hence the positions were virtually in the centre of the pheromone plume. The geometric progression of distances was chosen to emphasize the effect of turbulence on puff/non-puff statistics. Other parameters of the model were U = 1 m ⋅ s−1 (mean wind velocity), δU = 0.1 m ⋅ s−1 (wind fluctuations), a = 0.1 m (size of the pheromone source), χ = 0.4 (intermittency factor), yielding the probability density function of blank (B) and whiff (W) durations
f B ( x )= x - 3 / 2 2 ( 1 / τ - 1 / T B ) , x ∈ [ τ , T B ] , (1)
f W ( x )= x - 3 / 2 2 ( 1 / τ - 1 / T W ) , x ∈ [ τ , T W ] , (2)
where τ = a2d/[d(δU)2] is the shortest possible blank (whiff), TW = d/U is the longest possible whiff and TB = TW(1/χ − 1) is the longest possible blank. Throughout the paper, we report the results with respect to decadic logarithms of the durations of blanks. The logarithm of a blank represents a transformed random variable Y = g(B) = log10 B and hence the corresponding probability density function is derived using the formula
f Y ( y ) = f B ( g - 1 ( y ) ) | d d y g - 1 ( y ) | . (3)
Plugging g−1(y) = 10y and dg−1/dy = 10y ln 10 yields
f log 10 B ( x )= 10 - x 2 ln 10 2 ( 1 / τ - 1 / T B ) , x ∈ [ log 10 τ , log 10 T B ] . (4)
Sequences of whiffs and blanks were tested only once on a single recorded ORN. The dose of pheromone was constant throughout one recording session.
For the two largest virtual downwind distances from the source, 64 and 128 m, we selected the generated sequences, excluding those exhibiting extremely long stimuli, which led to the complete shutdown of ORN spiking activity. Thus, the statistics for 64 and 128 m were biased from the pure turbulence by removal of extremely rare events (puffs >30 s).
The data were analyzed using the R programming environment [23]. In total, we analyzed recordings of 217 moth ORNs obtained at 5 virtual distances and for 7 levels of pheromone dose. For each combination of virtual distance and pheromone dose we had 3-11 recordings of distinct ORNs, with the exception of 128 m and 1 ng dose, a category that was not studied because the occurrence of extremely long whiffs induced a complete interruption of the spiking activity at this high pheromone dose. Because the activity of ORNs is independent of other neurons [24], all the recordings obtained with a particular dose of pheromone and at a particular virtual distance were pooled and analyzed together.
The experimental setup emulated the fluctuating delivery of pheromone at 5 virtual distances (8, 16, 32, 64, 128 m) from the pheromone source. Pheromone dose was set to one of 7 levels, (10−6 to 100 ng) and a pheromone of constant concentration was released in puffs (whiffs), separated by blanks (see Materials and methods). The lengths of whiffs and blanks were generated randomly from the distributions of blanks and whiffs in real plumes [3] to mimic the natural pheromone fluctuations at a given downwind distance from the pheromone source. Henceforward, the distribution of blank durations is also referred to as the stimulus-timescale distribution. The durations of blanks and whiffs are restricted to the intervals [τ, TB] and [τ, TW], respectively (Eq 1). As the distance from the source increases, the range of possible blank (whiff) durations becomes wider (Fig 2), but the shortest blanks and whiffs always appear with the highest frequency.
ORN firing rate in response to a plume encounter was determined from the number of action potentials fired within the first 150 ms after each whiff arrival. The whiff onset is marked by a higher firing rate, which increases with the length of the immediately preceding blank (duration-rate relationship). The duration-rate relationship captures the sensitivity of the response with respect to the blank preceding the whiff onset and it is used as the encoding model for the whiff detection (Fig 3).
The duration-rate relationship was not stable throughout the whole recording. At the beginning, before the neurons became adjusted to the stimulation protocol, the responses were higher and became stabilized approximately after 100 s. Throughout the paper, we analyze only the behavior of ORNs in the adjusted state based on the recordings done between 100 s and 500 s.
The duration-rate relationship also changes with concentration of the odorant and the virtual distance. A higher pheromone dose leads to a higher maximum firing rate and higher slope of the duration-rate curve (Fig 3A–3D). The dependency on the virtual distance is less straightforward, nevertheless, we observe a systematic change of the slope of the curve, the maximum firing rate changes too, but the variance does not seem to be substantially affected (Fig 3E).
We investigated what the ORN duration-rate relationship reveals about the coding accuracy of pheromone encounters. Decoding accuracy is commonly evaluated by means of the stimulus-reconstruction paradigm, that is, by answering how well an ideal observer may determine the stimulus value from a noisy neuronal response [32]. Coding accuracy is quantified and interpreted by employing Fisher information (see Materials and methods, Eq 8) in a standard way, i.e, we use the fact that the inverse of the Fisher information is the mean square error of decoding by an ideal observer [16, 28, 29, 33–38]. Hence, the value of the Fisher information reflects the ultimate decoding accuracy and the maximum of the Fisher information corresponds to the optimum conditions for decoding. The approximation of the Fisher information is the square of the slope of the mean response divided by the variance of responses at each point. Thus, Fisher information is high when the firing rate has a low variability and changes rapidly with respect to the blank duration.
We observe that the profiles of the coding accuracy (Fisher information) and of the stimulus-timescale distributions are matched (Fig 4) in the sense that the Fisher information reaches high values for events of stimulation with high probability of occurrence and has low values for rare stimulations. Most importantly, the modes of the corresponding Fisher informations and blank distributions coincide in most cases (Fig 5). The correlation between the mode of the Fisher information and the mode of the corresponding stimulus probability density function is R = 0.6. If the results obtained with the smallest dose of 10−6 ng are excluded, the correlation coefficient increases to R = 0.8. This implies that the sensitivity of neuronal responses is adjusted to the most frequent temporal patterns of stimulation.
To assess the match of the complete Fisher information profiles to the stimulus-timescale statistics, we introduce the notion of average decoding accuracy. Each duration-rate relationship defines a specific encoding model for pheromone detection. We calculate the average decoding accuracy of an encoding model with respect to a given timescale distribution by integrating the whole profile of the Fisher information, where each value of the Fisher information is weighted proportionally to the frequency of the corresponding blank in the given timescale distribution (Eq 10 in Materials and methods).
For each tested dose and virtual distance, we calculated the average decoding accuracy assuming a) the stimulus-timescale statistics to which the encoding model is adjusted and b) other stimulus-timescale statistics corresponding to nonmatching virtual distances to which the encoding model was not adjusted (Fig 6). Among all possible nonmatching distances, we considered only the distances shorter than the matching one, since the ranges of possible blanks at longer distances than the matching one are wider than the range of blanks for which the Fisher information was calculated, and therefore the average Fisher information cannot be determined.
We observe that the average decoding performance is highest for the stimulus-timescale statistics of the matching distance. And conversely, the same encoding model applied to the statistics of a non-matching distance always resulted in a lower overall decoding accuracy. The only exception was the encoding model obtained for 16 m with the largest dose of 1 ng.
The first two layers of the moth olfactory system are organized so that the first-layer neurons (ORNs) converge onto a much smaller number of second layer neurons [24]. The signal-to-noise ratio (S/N) of the pooled signal increases with the square root of the number of pooled ORNs [39, 40]. Typically, hundreds of ORNs converge onto a single second-order neuron, resulting in a high S/N information transmission scheme. Assuming a homogeneous population of ORNs, information theory predicts that the optimal encoding scheme, i.e. a scheme that maximizes the mutual information between stimuli and responses [41–45], is such that the stimulus becomes a Jeffreys prior [46–51]. A Jeffreys prior is defined as a distribution that is proportional to the square root of the Fisher information. Vice versa, the Fisher information is then proportional to the second power of the stimulus distribution. Although the definition of the Jeffreys prior might evoke the idea that the stimulus distribution is to be adjusted in order to correctly correspond to the Fisher information, it is not the stimulus distribution, but the encoding model that must be tuned in order to establish this relation.
We constructed stimulus-timescale distributions that would satisfy the definition of the Jeffreys prior, based on the empirical Fisher informations, and compared them with the real stimulus-timescale distributions. In most cases these two appear to be in a close agreement (Figs 7 and 8), which is more evident for short blanks. As the blanks get longer, the predicted distributions decrease more slowly than the real distributions of blanks, however, we should bear in mind that Fisher information is most reliably calculated for short blanks, for which we had most of the data, whereas it may be inaccurate for long blanks due to influential outliers. The real and the predicted stimulus-timescale distribution differ also for observations made at 128 m virtual distance, reasons for which are given in the Discussion.
We demonstrated that responses of ORNs are adjusted to the spatio-temporal statistics of pheromone plumes at variable distances from the source as predicted by the efficient coding hypothesis. This is manifested mainly by the fact that the peak decoding accuracy, quantified by Fisher information, aligns with the most frequent timescale of blanks in the plume. The match of the maximum Fisher information and the mode of the distribution of blanks is less convincing only for the distance of 128 meters, possibly due to two reasons. First, whiffs at 128 m can be relatively very long and ORNs can become temporarily insensitive to the pheromone delivery. Second, neuronal recordings obtained for 128 meters typically contain a smaller number of blanks and whiffs, which can last longer, and therefore fewer responses are available, yielding possibly erroneous estimates of the Fisher information.
We report that not only the peak but also the overall decoding accuracy is adjusted to the local stimulus-timescale statistics. That is, the average decoding accuracy of pheromone encounters with the matching stimulus-timescale statistics of the particular distance is higher than if the same encoding model is applied for the non-matching statistics at a shorter distance. This suggests that there might exist processes, e.g. adaptation, driving ORNs to a response behavior optimal for the local stimulus distribution. Unfortunately, we cannot evaluate the coding accuracy in the non-adjusted state to assess if it improves in time, because the construction of Fisher information requires much more data than can be extracted from the beginnings of ORN recordings. Besides, the dynamic change of neuronal responses at the very beginning of the recordings might also be eventually influenced by the initial dynamics of the pheromone concentration, which can be neither traced nor controlled. Hence, we purposely do not infer the dynamical changes of coding properties, but only the adjusted state.
Another important finding is that the distribution of the stimulus timescale is close to the one that would be a Jeffreys prior with respect to the Fisher information. Such a relationship has important implications from a perspective of information theory [52, 53]. Under the assumption of vanishing response variability, which is essentially the case when many independent noisy “sensors” provide the signal for the decoder, the Jeffreys prior is the optimal stimulus distribution in terms of maximizing the mutual information between stimuli and responses [41–45]. We speculate that such situation in fact corresponds to the anatomy of the moth peripheral olfactory system, where the output of hundreds of ORNs converges onto a single antennal lobe neuron [24, 39]. The optimality of the Jeffreys prior has been theoretically predicted but never actually experimentally observed, to the best of our knowledge.
The fact that the stimulus-timescale distribution is close to the Jeffreys prior might have also some “technical” implications supporting the robustness of the reported results with respect to the chosen unit system of the duration of blanks. It is known that the Fisher information is not invariant with respect to the physical scale on which the stimulus is quantified. It has been demonstrated [54] that the change of the scale may shift the location of the maximum Fisher information, which could disrupt the match with the mode of the stimulus distribution. However, if the stimulus is distributed according to the Jeffreys prior, the match of the two modes is preserved after any arbitrary rescaling, i.e. for any choice of stimulus measurement units. Therefore our observation of the matching peaks of the timescale distributions and the Fisher informations does not depend on the chosen unit system.
Turbulence erases global gradients pointing towards the source, whereas local gradients point in random directions, so that the temporal structure of the sensory input is the unique information about the location of the source. The temporal pattern of odor encounters by a male moth is constantly changing as it flies to the pheromone source. ORNs may constantly adapt their coding to the temporal statistics of the odor signal, a process that would contribute to the efficient tracking by flying moths of pheromone plumes from large distances.
How ORNs dynamically adapt to a particular temporal pattern of odor encounters is elusive as a comprehensive picture of the insect olfactory transduction does not emerge yet. In particular, whether moth pheromone-responding receptors, which belong to the so-called OR family of insect odorant receptors, are ionotropic and/or metabotropic remains a matter of controversial discussion [55–59]. OR-expressing ORNs adapt to strong and/or prolonged stimuli [60]. Adaptation in insect ORNs covers a broad range of timescales, allowing a dynamic adjustment of their responsiveness: Drosophila ORNs can adapt to odorant pulses as brief as 35 ms on timescales as fast as 500 ms [61]. In A. ipsilon, ORNs exhibit short-term (timescale lower than a second) and long-term adaptation (timescale of minutes) in response to dynamical stimuli [7]. Adaptation occurs both at the level of receptor potential and action potential generators [62, 63]. Sliding adjustment of odor response threshold and kinetics has several molecular actors including ion channels, second messengers and ORs. ORs form non-selective cation channels which are also permeable for Ca2+. OR activation leads to Ca2+ influx into ORNs. Adaptation in Drosophila OR-expressing ORNs is mediated by the Ca2+ influx during odor responses [61]. First, Ca2+-dependent channels, such as BK channels which underlie the largest current density in moth ORNs [64], may serve for odor adaptation as in vertebrate ORNs [65]. Second, G protein signaling cascades can increase (adenylyl cyclase-dependent signaling [55], phospholipase C-dependent signaling [57]) or decrease (guanylate cyclase-dependent signaling [57]) the ORN sensitivity. Finally, ORs also adjust their sensitivity according to previous odor detections. Insect ORs are heteromers formed by an odor-specific OrX protein and an ubiquitous odorant co-receptor, Orco. Orco plays a central role both in down- and up-regulating the ORN sensitivity. Orco dephosphorylation upon prolonged odor exposure reduces the OR sensitivity [66]. On the other hand, Orco activation that depends on Ca2+, Ca2+-dependent proteins (protein kinase C and calmodulin) and cAMP production contribute to OR sensitization after moderate odor stimulation [55]. In moth pheromone-sensitive ORNs, Orco was proposed to function as a pacemaker channel, controlling the kinetics of the pheromone responses [67]. In addition, to expand the dynamic range of olfactory detection and thus allow to encode the temporal structure of odor plumes independent of their concentration [68], one or a combination of mechanisms of modulation of ORN sensitivity may contribute to adjust their coding efficiency to temporal statistics of pheromone fluctuations. Ca2+ plays a central role in tuning ORN sensitivity and fine adjustments of the Ca2+ concentration at the receptor potential and/or spike initiation generator site may be the principal mechanism of this adjustment of coding efficiency.
|
10.1371/journal.pntd.0005278 | A Novel Human T-lymphotropic Virus Type 1c Molecular Variant in an Indigenous Individual from New Caledonia, Melanesia | Human T-Lymphotropic Virus type 1 (HTLV-1) is endemic among people of Melanesian descent in Papua New Guinea, Solomon Islands and Vanuatu, and in Indigenous populations from Central Australia. Molecular studies revealed that these Australo-Melanesian strains constitute the highly divergent HTLV-1c subtype. New Caledonia is a French overseas territory located in the Southwest Pacific Ocean. HTLV-1 situation is poorly documented in New Caledonia and the molecular epidemiology of HTLV-1 infection remains unknown.
Studying 500 older adults Melanesian natives from New Caledonia, we aim to evaluate the HTLV-1 seroprevalence and to molecularly characterize HTLV-1 proviral strains.
Plasma from 262 men and 238 females (age range: 60–96 years old, mean age: 70.5) were screened for anti-HTLV-1 antibodies by particle agglutination (PA) and indirect immunofluorescence assay (IFA). Serological confirmation was obtained using Western blot assay. DNAs were extracted from peripheral blood buffy coat of HTLV-1 seropositive individuals, and subjected to four series of PCR (LTR-gag; pro-pol; pol-env and tax-LTR). Primers were designed from highly common conserved regions of the major HTLV-1 subtypes to characterize the entire HTLV-1 proviral genome.
Among 500 samples, 3 were PA and IFA positive. The overall seroprevalence was 0.6%. The DNA sample from 1 New Caledonian woman (NCP201) was found positive by PCR and the complete HTLV-1 proviral genome (9,033-bp) was obtained. The full-length HTLV-1 genomic sequence from a native woman from Vanuatu (EM5), obtained in the frame of our previous studies, was also characterized. Both sequences belonged to the HTLV-1c Australo-Melanesian subtype. The NCP201 strain exhibited 0.3% nucleotide divergence with the EM5 strain from Vanuatu. Furthermore, divergence reached 1.1% to 2.9% with the Solomon and Australian sequences respectively. Phylogenetic analyses on a 522-bp-long fragment of the gp21-env gene showed the existence of two major clades. The first is composed of strains from Papua New Guinea; the second includes strains from all neighboring archipelagos (Solomon, Vanuatu, New Caledonia), and Australia. Interestingly, this second clade itself is divided into two sub-clades: strains from Australia on one hand, and strains from Solomon Islands, Vanuatu and New Caledonia on the other hand.
The HTLV-1 seroprevalence (0.6%) in the studied adult population from New Caledonia appears to be low. This seroprevalence is quite similar to the situation observed in Vanuatu and Solomon Islands. However it is very different to the one encountered in Central Australia. Taken together, these results demonstrated that Australo-Melanesia is endemic for HTLV-1 infection with a high diversity of HTLV-1c strains and a clear geographic clustering according to the island of origin of HTLV-1 infected persons.
| The human T-lymphotropic virus type 1 (HTLV-1) infects at least 5 to 10 million individuals worldwide. In Australo-Melanesia, a south Pacific region including Papua New Guinea, Solomon Islands, Vanuatu archipelago and Australia, previous studies have shown that HTLV-1 is present in limited remote areas among few ancient Aboriginal populations. The molecular characterization of the HTLV-1 viruses present in such Indigenous individuals indicates that they belong to a specific HTLV-1 genotype called the Australo-Melanesian subtype c. In the present study, we provide evidence that the HTLV-1 endemicity among elderly individuals from New Caledonia is low and quite similar to Vanuatu and Solomon Islands, yet very different to the situation encountered in Central Australia. Furthermore, the newly described full-length HTLV-1 genomic sequences, from two Melanesian natives from New Caledonia and Vanuatu, both belong to the HTLV-1c genotype but are distinct from those of Aboriginal individuals living in neighboring countries. These results suggest that HTLV-1 viral strains were probably introduced among Melanesian native populations during multiple ancient human migration events to these archipelagos.
| The Human T-lymphotropic virus type 1 (HTLV-1) is a human oncoretrovirus, which causes two major diseases: adult T-cell leukemia/lymphoma and tropical spastic paraparesis/HTLV-1-associated myelopathy. This virus infects at least 5 to 10 million people worldwide [1]. Clusters of high endemicity have been described in certain geographic areas and ethnic groups, in particular in southwestern Japan, sub-Saharan Africa, South America, the Caribbean basin and localized areas in Iran and Australo-Melanesia [1]. Seven main HTLV-1 molecular subtypes are currently reported: the Cosmopolitan a-subtype, five African subtypes (b, d-g) and a Melanesian/Australian c-subtype found in Australia and neighboring Melanesian islands such as Papua New Guinea (PNG) and the Solomon and Vanuatu archipelagos. HTLV-1 exhibits a high genetic stability. The polymorphism observed among the viral strains is linked to the geographic origin of the infected individuals. This low genetic drift can be used as a molecular tool to monitor viral transmission and the movements of ancient infected populations [2–4].
In the context of ongoing studies among Aboriginal individuals in Australo-Melanesia [4–6], we extended our researches on HTLV-1 to native individuals from New Caledonia archipelago.
Therefore, the detection of HTLV-1 molecular variants from southwestern Pacific territories may constitute a powerful epidemiological and phylogenetic tool to reveal the ancient spread of this oncogenic virus, whose presence in this region has not yet been fully elucidated.
The aim of the present study was hence to characterize the HTLV-1 strains that infect Melanesian natives from New Caledonia, and to compare them with those previously characterized from Melanesian natives originating from neighboring territories.
New Caledonia (NC), a French overseas territory, is a group of islands located in the Southwest Pacific, east of Australia. (Fig 1). Our work was performed on plasma samples obtained from 500 elderly Melanesian adults, presenting to the blood-taking center at the Institut Pasteur de Nouvelle-Calédonie between July and October 2007. People attending this center usually come from the surroundings of Noumea, the capital city, located in the south of New Caledonia’s main island in the southern province. Briefly, plasma samples were taken from 238 women and 262 men (mean age 70.5 years, range 60–96 years) with the following stratification by age: 48.8% from 60–69 years of age, 39.4% from 70–79 years of age and 11.8% over 80 years of age.
This survey received administrative and ethical clearance in New Caledonia from the “Direction des Affaires Sanitaires et Sociales de Nouvelle-Calédonie” and in metropolitan France from the “Comité consultatif sur le traitement de l’information en matière de recherche dans le domaine de la santé” (N° 08.511) and the “Commission nationale de l’informatique et des libertés “(N° 908425). Information regarding the blood collection and the molecular characterization of HTLV-1 proviral strains was provided to the participants presenting to the blood-taking center.
HTLV-1 antibodies in plasma were first detected by a particle agglutination (PA) technique (Serodia HTLV-1, Fujirebio, Tokyo, Japan) at the Institut Pasteur (IP) in New Caledonia and then transferred to the IP in Paris for serological confirmatory assays. An indirect immunofluorescence assay (IFA) using the HTLV-1 and HTLV-2 transformed cell lines MT2 and C19 respectively, was performed. Anti-HTLV-1 antibody titers were determined by successive 2-fold dilutions. All positive samples were further tested by Western blot (WB) assay (HTLV-I/II Blot 2.4, MP Diagnostic, Illkirch, France). A sample with reactivity to the two gag proteins (p19 and p24) and both env-coded glycoproteins (the HTLV-1 recombinant gp46-I peptide [MTA-1] and the HTLV-1/HTLV-2 recombinant [rGD21] protein) was considered to be positive for HTLV-1 antibodies.
High-molecular weight DNA was extracted from peripheral blood buffy coat using the QIAamp DNA Blood Mini Kit (Qiagen Gmbh, Hilden, Germany). DNA samples were subjected to a first polymerase chain reaction (PCR) using human beta-globin specific primers, to ensure that DNA was amplifiable [7]. DNA samples were then subjected to four series of PCR to characterize the entire HTLV-1 proviral genome using primers designed from highly common conserved regions to the major HTLV-1 subtypes. These amplifications have been extensively described [5]. Briefly, we obtained four different HTLV-1 proviral genomic regions: F1, LTR-gag; F2, pro-pol; F3, pol-env and F4, tax-LTR. We also added a proviral DNA obtained from a native woman from Vanuatu (EM5) in the frame of our previous studies, and only partially molecularly characterized [4, 5]. Following electrophoresis, PCR products were purified and sequenced on both strands using a series of specific primers [5]. The Clustal W algorithm (MacVector 6.5 software, Oxford Molecular) was implemented to align forward and reverse sequences of each segment to obtain a consensus sequence of the full HTLV-1 proviral genome. Phylogenetic trees were generated from multiple alignments using HTLV-1 prototypic sequences available in Genbank, and the new sequences generated in this work.
Seven of the 500 plasma samples studied were tested positive by PA. Among them three were IFA positive with higher immunofluorescence assay titers on MT-2 (HTLV-1) than on C-19 (HTLV-2) cells (Table 1).
These three plasma samples exhibited high PA assay titers (≥ 1:2,048) and displayed full reactivity on WB (Fig 2). Taken together, these results revealed a low HTLV-1 prevalence (0.6%, 3/500) in indigenous Melanesian adults. Furthermore, the three positive samples were drawn from women.
Amplification of HTLV-1 proviral genome was tested for the 3 HTLV-1-seropositive women (NCP91, NCP173 and NCP201) and a sample from Vanuatu (EM5). Complete HTLV-1 genomic sequences (9,033-bp), derived from 4 PCR fragments of the appropriate size, were obtained for a unique sample from New Caledonia (NCP201; age = 73 years) and the sample from Vanuatu (EM5; age = 76 years). The New Caledonian NCP201 strain exhibited a 0.3% nucleotide divergence with the EM5 strain from Vanuatu. The pairwise comparison of these two new strains with other available complete sequences from Solomon Islands (Mel5) and Australia (Aus-CS, Aus-DF, Aus-NR and Aus-GM) revealed an overall nucleotide polymorphism of up to 2.9% (1.1% to 2.9% respectively) [8].
To include in our analyses most of the Central Australian sequences characterized to date [5], we performed additional phylogenetic analyses using both neighbour-joining (NJ) and maximum likelihood (ML) methods on the concatenated gag-tax genomic fragment (2,346-bp). First, the results obtained from NJ and ML (data not shown) methods confirmed that the NCP201 strain belongs to the HTLV-1 Australo-Melanesian c-subtype. Second, two clades supported by high bootstrap values (100%) exist within the HTLV-1c subtype. The first clade comprises strains from New Caledonia (NCP201), Vanuatu (EM5, PE376 and ESW44) and Solomon Islands (Mel5), and the second clade includes the sequences from Australia (Fig 3).
As most of available HTLV-1 sequences from PNG were generated on a 522-bp fragment of the gp21-env gene, we further analyzed this region to better understand the relationships within the HTLV-1c subtype strains. Thus, the nucleotide homology of the NCP201 strain reached [97.1%-97.3%] with the PNG strains (Mel1, Mel2 and Mel7). Interestingly, the New Caledonian NCP201 strain exhibited a nucleotide similarity range of [99.6%-100%], [99%-99.4%] and [98.1%-98.5%] with the Vanuatu, Solomon Islands and Australian strains respectively (Fig 4). Phylogenetic analyses using the 522-bp fragment revealed the existence of two major clades. The first comprised strains from Papua New Guinea, while the second included strains from all neighboring archipelagos; Solomon Islands, Vanuatu, New Caledonia, and Australia. Interestingly, this latter clade itself is divided into two sub-clades, with strains from Australia on one side, and strains from Solomon Islands, Vanuatu and New Caledonia on the other side (Fig 4). Both clades and sub-clades are supported by high bootstrap values (≥87%).
This study extends to New Caledonia the HTLV-1 endemicity described in native populations from Australo-Melanesia. HTLV-1 seroprevalence (0.6%) in the studied adult population appears to be low but is quite similar in Vanuatu and Solomon Islands. Indeed, previous studies showed that HTLV-1 seroprevalence in Vanuatu archipelago reached 0.74% among individuals aged over 60 years old, while among native Solomon islanders over 50 years old, the reported seroprevalence was 2.3% [4, 9]. This result contrasts with the situation in Australia where HTLV-1 prevalence among Indigenous Australians reached 40% in adults over 50 [10]. Of note, we investigated New Caledonian natives mostly living in the southern part of the main island, and clusters of higher HTLV-1 endemicity may exist elsewhere, as reported in other HTLV-1 endemic areas [11, 12]. Furthermore, some specific risk factors like cultural practices performed in the context of initiation rites, may only be present in populations from Central Australia [10].
From a molecular point of view, the NCP201 sequence belongs to the HTLV-1c subtype, together with the Melanesian strains from Papua New Guinea, Solomon Islands and Vanuatu [4, 6, 13, 14], as well as strains infecting Indigenous individuals from Central Australia [5, 15–17]. Phylogenetic analyses on the gag-tax fragment revealed the existence of two major clades within the HTLV-1c subtype (Fig 3). The first one includes all strains from Australia, and the second one comprises the strains from Vanuatu, New Caledonia and Solomon Islands. Subsequent analyses with the gp21 env gene fragment revealed the existence of a PNG clade. Thus, even among HTLV-1c subtype, a clustering according to the geographic origin of the HTLV-1 infected individuals exists.
Though this study is the first demonstration that New Caledonia is endemic for HTLV-1 infection, it has some limitations. The small sample size precluded a detailed epidemiological analysis that could illustrate the geographical dispersal of HTLV-1 infection in the whole New Caledonian archipelago. A further limitation was our inability to molecularly characterize various HTLV-1 strains. This would certainly require HTLV-1-infected individuals exhibiting a higher proviral load. Nevertheless, our findings are consistent with previous studies that have found that HTLV-1 seroprevalence was low among Indigenous Melanesian natives from neighboring archipelagos (e.g. Solomon Islands, Vanuatu).
Current studies combining both host genetics markers and viral analyses on HTLV-1, but also on Kaposi's sarcoma-associated herpesvirus (KSHV)/human herpesvirus 8 (HHV-8), which infects the same aboriginal populations [18], are ongoing to get better insights into the peopling of such Melanesian territories [18–20].
|
10.1371/journal.pbio.0060011 | Distinct Cerebral Pathways for Object Identity and Number in Human Infants | All humans, regardless of their culture and education, possess an intuitive understanding of number. Behavioural evidence suggests that numerical competence may be present early on in infancy. Here, we present brain-imaging evidence for distinct cerebral coding of number and object identity in 3-mo-old infants. We compared the visual event-related potentials evoked by unforeseen changes either in the identity of objects forming a set, or in the cardinal of this set. In adults and 4-y-old children, number sense relies on a dorsal system of bilateral intraparietal areas, different from the ventral occipitotemporal system sensitive to object identity. Scalp voltage topographies and cortical source modelling revealed a similar distinction in 3-mo-olds, with changes in object identity activating ventral temporal areas, whereas changes in number involved an additional right parietoprefrontal network. These results underscore the developmental continuity of number sense by pointing to early functional biases in brain organization that may channel subsequent learning to restricted brain areas.
| Behavioural experiments indicate that infants aged 4½ months or older possess an early “number sense” that, for instance, enables them to detect changes in the approximate number of objects in a set. However, the neural bases of this competence are unknown. We recorded the electrical activity evoked by the brain on the surface of the scalp as 3-mo-old infants were watching images of sets of objects. Most images depicted the same objects and contained the same number of objects, but occasionally the number or the identity of the objects changed. As indicated by the voltage potential at the surface of the scalp, the infants' brains reacted when either object identity or number changes were introduced. Using a 3-D model of the infant head, we reconstructed the cortical sources of these responses. Brain areas responding to object or number changes are distinct, and reveal a basic ventral/dorsal organization already in place in the infant brain. As in adults and children, object identity in infants is encoded along a ventral pathway in the temporal lobes, although number activates an additional right parietoprefrontral network. These results underscore the developmental continuity of number sense by pointing to early functional biases in brain organization.
| Converging behavioural, brain-imaging, and neurophysiological results suggest that knowledge of number is an evolved competence of the animal and human brain, with a cortical basis in bilateral intraparietal cortex [1–8]. The number sense hypothesis [1] postulates that this cerebral system is available early on during development, possibly during infancy, and guides the learning of numerals and arithmetic in childhood. Indeed, an association of number processing tasks with intraparietal areas has been demonstrated in 4- and 5-y-old children [6,9]. At this age, a change in the cardinality of a set of objects (also called “numerosity”) leads to a response in the right intraparietal cortex, at a location comparable to adults [4,5]. This response cannot be ascribed to a domain-general attentional reaction to novelty inasmuch as it is not seen if the number stays constant while the identity of the objects changes [6].
It is tempting to speculate that this parietal sensitivity arises from a predisposition of parietal cortex for spatial and numerical transformations, possibly present since birth. In subjects below 4 y of age, however, neuroimaging explorations are limited, and the existence of early numerical abilities is mostly based on behavioural results. Habituation and violation-of-expectancy paradigms have revealed a clear sensitivity to large numbers in 4–6-mo-old infants. For instance, infants discriminate sets of eight versus 16 dots, even when nonnumerical parameters such as density and total surface are tightly controlled [10]. In the range of small numbers one, two, and three, however, behavioural evidence remains debated. Initial observations suggested discrimination in infants and even in newborns [11,12], but more recent studies demonstrated that this competence was mostly driven by confounded low-level perceptual dimensions such as continuous amount of stuff [10,13–15]. Nevertheless, Feigenson [16] demonstrated that infants can be driven to attend to numerosity, even in the range from one to three objects, when the sets comprise highly distinctive objects rather than replications of the same object. Moreover, successes at numerical tasks have been observed when the stimuli to be compared are presented in different modalities [17–19], therefore eliminating the possibility of relying on nonnumerical aspects of the stimuli.
Here, we aimed at bringing brain-imaging evidence to bear on the existence of a dedicated number sense system in early infancy. Our experiments probed the infant brain for a specific response to changes in the cardinal of sets of objects, distinct for the response to changes in object identity, and possibly common to small and large numbers. Identifying a brain response to numerosity in infancy would support the hypothesis of a developmental continuity in number sense and significantly extend the current evidence for a cerebral specialization for number, which is currently based almost exclusively on adult and children data. Only a single brain-imaging experiment to date has been conducted on infants' processing of number. Berger et al. [20] used event-related potentials to investigate the cerebral process underlying adults' and 7-mo-olds' reactions to correct and incorrect arithmetic operations (1 + 1 = 2 vs. 1 + 1 = 1). Compared to plausible outcomes, arithmetic violations elicited a cerebral reaction over anterior electrodes both in infants and adults. Because it emerged from the comparison of two arithmetical situations, which differed by their level of plausibility, this reaction probably indexes a general process of violation detection, underlying infants' behavioural response in terms of looking time to such implausible events.
Our experimental design relied on the habituation paradigm, which capitalizes on the phenomenon of neural adaptation: when a stimulus is repeatedly presented, the response of neural populations encoding this stimulus decays progressively over successive trials, but it recovers when a novel item is introduced. In particular, brain areas sensitive to number increase their activation in response to a change in number within a block where most stimuli have the same number [4–6]. Although predominantly used with functional magnetic resonance imaging (fMRI), habituation paradigms have also been successfully applied to young infants using event-related potentials (ERPs) [21,22]. We used ERP adaptation to investigate numerical abilities in 3-mo-old infants. This age is significantly younger than what is seen in most behavioural studies of infant numerical competence, which typically study 4½- to 6-mo-olds [10,13,18,23–26]. Although a few studies have reported number discrimination in newborns [11,12], these studies did not typically incorporate sophisticated controls over all nonnumerical parameters such as size, density, and occupied area.
In addition to probing numerical representation at this early age, we aimed to study whether the infant brain is already anatomically and functionally structured. In adults and 4-y-olds, fMRI adaptation has revealed a double dissociation between ventral regions sensitive to changes in object identity, but not in number, and dorsal regions sensitive to changes in number, but not in object identity [4,6]. Within the visual system, the ventral pathway is thought to be primarily concerned with object identity (“what”), and the dorsal pathway with object location, size, and motor affordance (“where” and “how”) [27]. At 4 mo of age, behavioural evidence suggests that this basic dorsal/ventral organization may be already present, since infants are able to selectively attend either to the identity of objects (e.g., faces) or to their location, yet fail to link these two types of information [28]. In line with these results, a brain-imaging experiment with 6-mo-olds revealed that holding a stimulus in memory relies on different brain mechanisms, depending on the nature of the stimuli, and therefore, presumably, the type of information retained in memory (identity vs. location) [29]. However, the presence of a ventral/dorsal organization in infancy has not been tested directly by looking at the anatomical localization of brain activation.
To address these questions, we recorded event-related potentials from 3-mo-old infants while they were presented with a continuous stream of images, each showing a set of objects. Within a given run, most sets had the same “standard” cardinality and object identity. However, we occasionally inserted test images that could differ from the habituation images in number and/or object identity, thus defining four types of test stimuli (Figure 1).We compared the visual event-related potentials evoked by unforeseen changes either in the identity of objects forming a set, or in the cardinal of this set. Because of the debated possibility of a discontinuity between the small and large number ranges, our tests involved both large and small numbers. Three groups of infants were tested respectively with a pair of small numbers (two versus three), distant large numbers (four versus eight), or very distant large numbers (four versus 12).
We examined the event-related potentials evoked by the same critical test images, which were defined as deviant or not, on both the numerical and object-identity dimensions, as a function of their relation to the context provided by the preceding images. ERPs showed a series of waveforms classically observed in infant studies with visual stimuli [21,30]. We observed two successive occipital negativities peaking at 88 and 192 ms after image onset, followed by an ample and long-lasting waveform peaking at 484 ms (P400), which was positive on posterior electrodes and negative on anterior electrodes. In order to evaluate the cerebral responses to changes in number and object identity, we computed the difference between the deviant number and standard number test images, and also between the deviant object and standard object test images. These differences were tested using a cluster-detection algorithm that detects spatiotemporal clusters with a consistent statistical difference over time and space, and assesses their significance against re-randomized data, with a correction for the large search space (electrodes and time points; see Methods). This cluster analysis indicated that the introduction of a change in number or in object identity generated discrimination effects that modulated the duration and amplitude of the P400.
For number change, a significant effect (corrected p = 0.044; uncorrected p = 0.0073) was due to more-negative voltages on number change compared to same-number trials over a cluster of 13 bilateral parietocentral electrodes (712–1,196 ms after stimulus onset), and more-positive voltages over two left frontal electrodes (992–1,196 ms). To further explore the number-change effect, voltages were averaged on a 100-ms time window (750–850 ms) and four groups of electrodes corresponding to the negative and positive maxima of this effect over parietal and prefrontal areas, and the symmetrical groups on the other hemisphere (see Methods and Figure 2A for the localisation of the electrode groups). These values were entered in two four-factors ANOVAs (number size, left vs. right hemisphere, parietal vs. prefrontal electrode group, and number change or object change as the last factor). On these electrodes, number change interacted significantly with the electrode group (F(1,33) = 18.7, p = 0.0001), but object change did not (F < 1). The effect did not differ significantly across hemispheres (interaction: F < 1) and was significant over both the left and the right electrode groups (left hemisphere: F(1,33) = 21.1, p < 0.0001; right hemisphere: F(1,33) = 8.3, p = 0.0069). The number-change effect was also consistent across groups as it did not interact with number size (F(2,33) < 1) and was significant within each experimental group, including the small-number group (four vs. 12: F(1,11) = 6.6, p = 0.026; four vs. eight: F(1,11) = 5.7, p = 0.036; and two vs. three: F(1,11) = 7.9, p = 0.017).
Object change also yielded a significant effect (corrected p = 0.035; uncorrected p = 0.0054), but slightly faster and with a distinct topography including a right anterior negative cluster (13 electrodes, 672–972 ms) and a bilateral temporooccipital positive cluster (ten electrodes, 636–868 ms). Similar to the analysis of the number effect, an ANOVA was performed on the voltage averaged during the same 100-ms time window and on four groups of electrodes (two groups located at the maxima of the dipole of the object-change response on occipital and central areas, and their symmetrical counterparts; see Figure 2B for the localisation of the electrode groups). On these electrodes, object change interacted with electrode localization (F(1,33) = 13.1, p = 0.001), whereas number change did not (F(1,33) = 1.2, p = 0.28). Voltages were more negative for deviant objects on central electrodes, and more positive on occipitotemporal electrodes. This effect did not interact with hemisphere (F(1,33) = 1.4 p = 0.24), and it reached significance over both the left and the right electrode groups (left hemisphere: F(1,33) = 7.0, p = 0.012; right hemisphere: F(1,33) = 16.9, p = 0.00024). The object change effect also did not interact with number size (F(2,33) = 1.4, p = 0.26). However, further analyses revealed a significant response to object change only in the two groups that saw the smallest numbers (two versus three group: F(1,11) = 6.7, p = 0.025; four versus eight group: F(1,11) = 13.4, p = 0.004); but not in the four versus 12 group (F < 1). A likely interpretation is that objects were presented at a smaller physical size in the latter condition and thus were harder to discriminate (see [4] for a similar observation).
Together, these results establish a double dissociation between the processing of number and of object identity at the scalp level: some electrode groups show an effect of number change, but not of object change, whereas others show the converse response. The dissociation indicates that neither of these responses can be ascribed solely to domain-general mechanisms such as attention to novel events.
Although event-related potentials have low spatial resolution, they can provide coarse information about brain localization, particularly given the dense sampling of the infant head provided by our 65-electrode net. We took advantage of the anatomical images obtained from our previous magnetic resonance experiments with infants [31,32] to compute a detailed model of the infant head and cortical folds (see Figure 3). We then used this forward model to reconstruct a plausible distribution of the cortical origins of our scalp recordings, using distributed cortical source modelling and a minimum norm constraint (see Methods). The reconstructed activations, while probably accurate only to within 1 or 2 cm, indicated a double dissociation of number and object identity at the cortical level (see Figure 3). Object change activated a stream of left ventral temporal regions, starting in posterior occipitotemporal regions around 300–400 ms and with a durable activation up to 800–1,000 ms in anterior temporal as well as posterior occipitotemporal regions. Conversely, number change led to the activation of a right network pertaining to the dorsal pathway, including the right inferior parietal and right inferior frontal region. We also observed effects of number change in temporal regions. An antagonistic relation was observed whereby number change led to a decrease in the left anterior temporal regions previously associated with object change, and a concomitant increase in right anterior temporal activation (see Figure 3).
We found a response common to small and large numbers on the scalp; we also searched the data for possible differences between them. We first examined whether the timing of the number-change effect varied with the range of numbers tested. An ANOVA was run on a larger time window (650–1,250 ms) spanning the whole duration of the effect as identified by the cluster analysis, on the groups of electrodes defined by the number-related analysis, including the same factors as previously analyzed (number size, hemisphere, electrode group, and number change) and a supplementary factor of time (six successive 100-ms–long time windows). The effect (interaction between electrode group and number change) did not interact with time or with number size (all Fs < 1), even when the two groups of infants that were presented with large numerosities were grouped together (Fs < 1). A cluster analysis directly comparing the responses to number change in the small-number range (two versus three group) versus the large-number range (four versus eight and four versus 12 groups) also did not identify any effect (corrected p > 0.42, uncorrected p > 0.11). Thus, our data do not show any discontinuity between small and large numbers.
By looking at brain electrical activity in 3-mo-old infants, our results establish a double dissociation between number and object identity processing. Visual evoked potentials were recorded as infants were presented with images of sets of objects, most of them depicting the same number of the same objects, but with occasional deviants in object identity and/or number. At the scalp level, some electrode groups showed an effect of number change, but not of object change, whereas others showed the converse response. Those brain responses reveal that infants are sensitive to both object identity and number by age 3 mo. For object identity, our results merely provide additional support to a considerable amount of previous behavioural evidence for visual object processing in infancy, including in newborns [33–35]. For number, however, our experiment goes beyond previous behavioural research by demonstrating numerical discrimination at an age younger than most previous reports [10,13,18,25,26], and by including strict controls over nonnumerical parameters that were typically lacking in previous experiments at this age [12,24].
The observed dissociation between object identity and number is particularly crucial for the interpretation of the functional role of the observed brain responses. It indicates that neither of them can be ascribed solely to domain-general mechanisms such as attention to novel events. Detection of rare events has been associated with a series of late waveforms in infants, and particularly a large negative deflection culminating 400–800 ms after stimulus onset on the anterior electrodes (Nc) [36–38]. Similarly, Berger et al. [20] observed a negative anterior waveform when 7-mo-old infants watched a movie that included an arithmetic violation (1 + 1 = 1 or 2 − 1 = 2). Although direct comparison of these results with ours is made difficult by the occasional use of different references for voltages, the domain-general Nc response bears some resemblance with the response we observed after changes of objects, which includes a negative component on frontal electrodes. The response to changes in number, however, is clearly distinct from this domain-general Nc. Our source analysis further suggests that neither of these components originates solely from an anterior attention network, which is considered to be the source of the Nc component [37], but that both our responses involve distributed and distinctly located sources.
The observed dissociation between object- and number-based responses also allows us to exclude simple interpretations of our results as artefacts of the experimental procedure. In particular, although parents were allowed to see the stimuli, it is very unlikely that the reaction we observed was a reaction to a parental signal rather than a reaction to our stimuli. First, the onset of the brain reaction occurred before 500 ms, which leaves little time for a whole chain of parental reaction, parent-to-infant signalling, and infant brain activation to occur. Second, most crucially, given the dissociation in ERPs associated with number and object change, it seems very improbable that all parents would have differentiated the two types of changes in a consistent way so as to elicit, through an unknown, yet differentiated feedback, a consistently distinct brain response in all infants.
Although they should be interpreted cautiously, as they represent only a tentative model with coarse spatial accuracy, our source reconstructions suggest, at a minimum, that distinct cortical pathways already exist in 3-mo-old infants for processing number and object identity. The sources of the object-change effect were located along the left temporal cortex, with an antagonistic response in the right temporal cortex. These results mesh well with fMRI observations from children and adults, where object-change responses have been recorded in the inferotemporal cortex, particularly in the left hemisphere [4,6]. In the present experiment, objects were defined both by their shape and by their colour, and therefore the observed response to object change could involve lower-level colour- and shape-sensitive areas as well as object-sensitive areas. Indeed, examination of the source reconstruction results brings support to this interpretation, since an entire stream of ventral occipitotemporal areas was activated in cascade in response to the presentation of deviant objects. The posterior areas, which responded first, might have encoded low-level features of the stimuli such as shape or colour, whereas the later activation of more anterior areas might relate to an encoding of object identity. Note, however, that the surface ERP component responding to object change was probably not sensitive solely to changes in colour, since it did not appear in the large-number condition in which the objects were smaller and therefore hardly discriminable in shape, but still markedly different in colour.
In contrast to changes in object identity, changes in number activated a parietoprefrontal network in the right hemisphere. Although bilateral areas have been associated with number processing in adults, the present lateralization to the right hemisphere is consistent with previous studies that suggest a greater right lateralization of intraparietal responses to number in 4-y-olds than in adults [6]. The activation of the left inferior parietal region increases with age in older children [39,40], and our results suggest that it might not be dominant in the first year of life.
In addition to activating a right parietoprefrontal network, changes in number elicited a decreased response within the left temporal region that was responsive to object change (as well an increase in the opposite right temporal region). This aspect of our results suggests that there may be an antagonistic relation between the dorsal network for number and space and the ventral network for object identity. Although unexpected, this tentative conclusion meshes well with several previous behavioural studies that observed a drop in infants' performance when object identity and either numerical or spatial information have to be jointly processed and integrated. For instance, Xu [41] and Xu and Carey [42] observed that infants were unable to use object-identity information to infer numerosity: when two distinct objects emerged successively from behind an occluder, infants were clearly able to discriminate these two objects, but they were not surprised if the occluder later dropped to reveal just one object. Furthermore, Mareschal and Johnson [28] explicitly demonstrated a competition between memory for object features (“‘what”‘) versus spatiotemporal trajectory (“‘where”‘) depending on the category of visual stimuli used. When the stimuli were faces, infants detected changes in their identity, but not their position, whereas when they were manipulable toys, infants detected changes in location, but not in identity [28]. Furthermore, Southgate et al. [29] observed a dissociation in the brain responses of 6-mo-old infants as they were holding either a toy or a face in memory. These results, together with our source localizations, support the hypothesis that during infancy, information pertaining to the “‘what”‘ and “‘where”‘ pathways is already processed by distinct networks that are initially not fully integrated in a coherent behaviour, but rather may interact according to an antagonistic mode. Further research will be needed to confirm this antagonistic effect and probe the roles of attention, language, and prefrontal cortex development in overcoming it. Some researchers have suggested that a full integration may not occur until much later during childhood [43].
Our results, which are based on a measure of infant cerebral activity, differ in part from previous conclusions based on behavioural results. We observed a shared brain response to numerical changes in both small (two vs. three) and large (four vs. eight and four vs. 12) number ranges. This finding is in apparent conflict with previous behavioural results suggesting that even 6-mo-old infants are unable to discriminate small numbers two and three when nonnumerical parameters are appropriately controlled [13,44,45], and are unable to discriminate ratios of 2:3, even in the large-number range, until 9 mo of age [46].
At this age, infants are thought to possess two separate systems conveying numerical information in the small- (<4) and large-number range, respectively. In the large-number range, numbers are represented by analogical internal magnitudes. In the small-number range, infants track sets of one to three items by means of attentional indexes attached to each object, and they can use one-to-one correspondence on these indexes to detect the absence or sudden apparition of objects [13,47]. With respect to this theoretical background, our results raise three questions: (1) why did we observe a capacity of the dorsal system to respond to both small and large numbers? (2) why did we not observe a distinct brain response to small numbers? and (3) why do brain measures seem to be more sensitive than behavioural measures?
Although behavioural results demonstrate the existence of a specific system for small numbers, no published study contradicts our present findings that the system of analog magnitude representation, which underlies infants' numerical competence for large numbers, can also respond to small numbers. In fact, many adult and animal studies reveal a continuity in behaviour between small and large numbers, suggesting that the analog system extends to small numbers. For instance, Cordes et al [48] found the same Weber signature across small and large numbers in a task adapted from the animal literature, in which human adults were required to tap a given number of times without counting. Contrary to the prediction of the small number system, responses showed some variability in the small-number range, in continuity with the large-number range. In tapping tasks, animals show a similar behaviour, with no discontinuity between small and large numbers [49]. In infants, a set-size signature characteristic of the object tracking system is obtained mostly in one type of task in which items are shown successively or retrieved successively from a box [50,51]. Hauser et al. [52] observed a similar set-size limit in untrained rhesus monkeys tested with a single trial in semiwilderness, but Beran [53] found no such discontinuity when using a similar task with multiple trials in trained laboratory animals: monkeys could select the larger of two sets based on their number, with a ratio-dependent performance and no discontinuity between small and large numbers. Under conditions of simultaneous presentation of a set of objects, as tested in the present work, monkeys' analog representation always appear to extend to small numbers. For instance, Brannon and Terrace [54,55] observed ratio-based generalization from small to large numbers: after having been trained on the comparison of small quantities (one to four), rhesus monkeys were able to generalize this training to larger numbers (up to nine).
At the brain level, neurons sensitive to numerosity have been found in the intraparietal sulcus as well as in the prefrontal lobe of monkeys. These neurons encode quantities from one to 30 objects and show a seamless increase in their tuning width with numerosity, corresponding to Weber's law, without any sign of a discontinuity at the boundary between numbers smaller or larger than three [7,56]. Similar observations have recently been made in another population of number-sensitive neurons located in lateral intraparietal area [57]. Our results accord with these observations, and suggest that human infants share with nonhuman primates an analog representation of numerosities that extends seamlessly across small and large numbers alike.
Given the behavioural evidence for a distinct small-number system in infancy, one might wonder why we did not observe an additional cerebral response in the small-number range. Several aspects of our design might have precluded this possibility, however. First, behavioural results show that infants can discriminate small numbers via an object tracking system, but that this competence disappears when the objects are identical and nonnumerical aspects of the sets, such as object size and density, are controlled for [58]. These factors were tightly controlled for in the present experiment, and thus our design may have effectively cancelled any response of the small-number system. Second, even if this system had been reactive, the neural habituation method that we used may not be appropriate to detect it. Repetition suppression occurs in neural populations tuned to the stimulus value presented repeatedly during the habituation phase. Thus, the habituation method is most appropriate to detect neural populations that encode an explicit representation of the dimension being tested. The object tracking system, however, is thought to represent numerical information only in an implicit way: in this system, there is no summary representation of “two”; instead, infants form a mental model of two objects by recruiting two attentional indexes or “object files” [47,59,60]. This system, therefore, may not show a repetition suppression effect and may remain invisible to the neural habituation method. In brief, our experiment was not targeted at detecting the cerebral bases of an object tracking mechanism. Consequently, the absence of a specific effect for small numbers in our results does not exclude the existence of such system, which may perhaps be detected by other means.
Finally, our event-related potentials revealed a capacity for discriminating numerosity, whereas behavioural studies typically show infants failing in similar conditions. Hence, although we observed a positive reaction to changes in number in brain activity, 5-mo-old infants are not able to discriminate numerosities when images were presented at the rate of one every 1,500 ms as here, but needed a longer presentation of 2,000 ms to succeed in a behavioural experiment [26]. Furthermore, we observed a reaction for numbers separated by a ratio of 2:1 (four vs. eight) and even 3:2 (three vs. two), although positive discrimination of 3:2 ratios is not achieved until 9 mo of age in behavioural measures [10,46]. Whereas behaviour is often a composite measure reflecting the combined effects of several processing stages, brain measures can provide a purer index of a given level of representation [61] and can track the response of a given system even when it does not lead to an overt behavioural response. Many examples of ERP-behaviour dissociations exist in the adult literature. For instance, high-density ERPs demonstrate a series of processing stages of subliminal visual stimuli that subjects deny seeing [62]. In a training task where subjects learn to attend to subtle phonetic differences, ERP evidence for stimulus discrimination (mismatch negativity) may appear as much as 24 h before a change in overt behaviour occurs [63]. These experiments exemplify the fact that cerebral measures can be more sensitive than behavioural measures.
A second factor, more specific perhaps to the developmental context, is that when a conflict exists between several levels of representations, infants and young children may lack the ability to resolve the conflict and integrate all of their sources of knowledge into a coherent behaviour. As a result, behavioural measures may not fully reflect the infant's competence. For instance, in the classical object permanence task, infants appear to lack knowledge of hidden objects when tested with reaching measures, but not when tested with looking time or eye-orienting methods [34,35,64,65]. In the case of small numbers, object-tracking representations may be so salient that they determine behaviour even when the analog quantity system is in possession of more-advanced information concerning numerical quantity. As discussed above, object identity and set numerosity pertaining to the ventral and the dorsal pathway may be not fully integrated during infancy. Furthermore, ERP responses to object and number change occur relatively late, with a peak around 800 ms, much later than the discrimination responses observed at the same age with auditory phonetic stimuli (around 200–400 ms for phonetic mismatch [22]) or visual faces (around 300 to 700 ms [38,66]). This slowness, perhaps related to the small size of the objects presented, may explain that numerical competence is not always expressed in behavioural measures such as looking time, which may be driven by faster computations of more salient perceptual dimensions. In particular, stimuli need to remain present for a longer time in order for infants to react to number [26]. Whereas ERP measures detect the on-line brain response to numerosity, more time may be needed for this response to guide overt behaviour.
We should also underscore the fact that although the stimuli used in behavioural and ERPs experiments can be similar, the constraints related to each type of procedure lead to important differences in experimental settings. Whereas behavioural results are based on a few measures obtained after several minutes of habituation to one type of stimulus, we recorded ERP to tens of object and number changes embedded in a continuous flow of standard stimuli. Such fast recurrent presentation provides more evidence to the infant about the value of the standard numerosity, and repeated measurements may also confer a greater sensitivity to ERP experiments. Both factors may ultimately explain why we observed a reaction to numbers separated by a smaller ratio (2:3), and presented at a faster pace.
Altogether, our results reveal a system representing both small and large numbers in infancy. They suggest developmental continuity, with a right parietal involvement for number and a basic ventral/dorsal organization already present at 3 mo of age. The parietal number system may constitute the neural substrate of infants' initial numerical competence, which increases in precision in the course of development and possibly guides the acquisition of arithmetic and more elaborate mathematical concepts.
High-density ERPs were recorded while infants were presented with a continuous stream of images, each depicting a set of colourful animal-like objects on a black background. Stimuli were projected onto a screen measuring 100 × 80 cm, and the infant was seated on a caretaker's lap at an approximate distance of 80 cm from the screen. Two categories of stimuli were used, respectively labelled “habituation” and “test.” Within a block, all habituation stimuli had the same number and depicted the same object. Test stimuli occasionally differed from the habituation stimuli in their number and/or in the identity of the object used, thus defining four types of test stimuli (see Figure 1).
In habituation experiments, it is important to control for nonnumerical parameters, such as the position of the objects and the physical parameters of the image, in order to ensure that a positive result can only be attributed to the discrimination of numbers rather than to the variation of any nonnumerical parameter. We applied a strategy similar to previous publications [4–6,10,67].
The stimuli were automatically generated using a variant of our laboratory's numerosity stimulus generation programs, which are described in depth at http://rd.plos.org/pbio.0060027 (78 KB DOC). The position of the objects and the physical parameters of the image (intensive parameters: object surface size, average area devoted to each object; extensive parameters: total luminance, total occupied area) varied across stimuli, following different rules for the habituation and test stimuli. For the test stimuli, on the one hand, the extensive parameters of the display (total luminance and total occupied area) were kept constant on average. Therefore, if infants encoded some extensive parameter of the stimuli rather than the cardinality of the sets, they would have the same reaction to all test stimuli. For the habituation stimuli, on the other hand, intensive parameters were kept constant on average. They were completely uncorrelated to number, as they were randomly drawn from a uniform distribution, ideally ranging between the minimal and maximal value these parameters take for the test stimuli (see Figure S1). If infants only encoded some intensive parameter of the stimuli, their reaction to the test stimuli should be independent of the numerical habituation context.
In the analysis of the results, we considered the difference between deviant and standard test stimuli in various numerical contexts, where two numerosities took alternatively the roles of deviant and standard numerosity. Reactions to intensive parameters would generate effects depending only on the numerosity of the test stimulus presented, not on the numerical habituation context (no effect of the deviant/standard factor). On the other hand, reactions to extensive parameters would be the same over all test stimuli, independent of their numerosity. Therefore, reactions to the number-change factor can only be imputed to numerosity, and not to the low-level attributes of the stimuli.
The total experiment consisted of eight blocks of 16 trials each, in which each trial consisted in the initial presentation of a variable number of habituation stimuli (two to five images) followed by a single test stimulus. Within each block, images were presented continuously, at the rate of one image every 1,500 ms, and no cue indicated the presence of a test stimulus or the beginning of a new trial. When the infant looked away from the screen, the experiment was stopped, the infant's gaze was attracted to the screen, and then additional habituation stimuli were presented before a test stimulus appeared.
The experiment ended after eight blocks, or when the infant showed signs of fussiness.
Throughout the whole experiment, each infant was presented with two different numbers whose respective roles (either deviant or standard) switched each time a new block started. Three different pairs of numbers were used in three different groups of infants: very distant large numbers (four versus 12), distant large numbers (four versus eight), and small numbers (two versus three). In order to maximize the infants' attention, in each of the eight blocks a new standard object was selected. The deviant object was chosen so that its shape and colour would be maximally different from the standard object.
Thirty-six healthy infants were included (14 females; mean age 103 d, range 92 to 124 d). Additional infants were rejected because of fussiness (77), excessive movements (31), excessive sweating artefacts (6), electrode net degradation (16), or other technical failure (3). The study was approved by the regional ethical committee for biomedical research, and parents gave their written informed consent.
Electroencephalography (EEG) was digitized continuously at 250 Hz using a 65-electrode geodesic electrode net (EGI) referred to the vertex. The recording was first digitally filtered between 0.5 and 20 Hz. For each trial, we then extracted an epoch starting 400 ms before the presentation of the test stimulus and lasting 2,000 ms after the onset of the test stimulus. Electrodes contaminated by eye or motion artefacts were automatically rejected, and trials with more than 25 contaminated electrodes were rejected. The remaining trials (on average 61.6 per participant; range 30 to 148) were averaged to obtain ERPs in each of four trial types (DN: deviant number; SN: standard number; DO: deviant object; SO: standard object). Note that experimental design crossed the two variables of number change and object change, so that in principle, we could have computed ERPs within smaller subcategories of trials (e.g., DN with or without a concomitant change in object identity). In practice, however, too few trials were available, so that only the main effects of object change and number change could be studied. ERP averages were digitally transformed to an average reference, corrected for eventual slow artefacts by removing a linear trend on the whole segment, baseline corrected on the 200 ms preceding the onset of the test stimulus, and finally spatially smoothed by convolution with a Gaussian, with a standard deviation corresponding to the distance between electrodes on an infant's head, e.g., about 1.5 cm.
In order to evaluate the cerebral responses to changes in number and object identity, we computed the difference between the DN and SN test images, and also between the DO and SO test images. Computing these differences cancelled out all the brain activity resulting from processes that are common between the two conditions, such as low-level perceptual processes, or processes resulting from the presentation of the previous habituation stimuli.
A customized cluster analysis coupled with a randomization procedure was used to identify effects and assess their level of significance after correction for the large number of electrodes and time points tested. The two conditions of interests were first compared separately, using Wilcoxon signed-rank test, for each electrode and each time sample. Levels of probability obtained were normalized into Z-scores, and then thresholded at ±1.96 (p = 0.05 two-tailed). These thresholded values were used to define spatiotemporal clusters of activation present for an extended time period and on contiguous electrodes. The entire time period extending between the onset of the test stimulus and the onset of the next habituation stimulus (1,500 ms) was used to detect the clusters. First, our procedure pooled above-threshold samples corresponding to spatially contiguous electrodes or adjacent time samples, separately for positive and negative above-threshold Z-scores, thus defining several positive and negative clusters. Each cluster was then attributed a weight equal to the sum of the absolute value of the Z-scores of all of its constitutive samples. Because our data were transformed to an average reference, significant effects were expected to present two near-simultaneous ERP components of opposite signs over two distinct electrode sets. Therefore, for each time sample, the total effect strength was measured as the sum of the weights of the largest positive and negative clusters that were found at this time.
To evaluate the significance of the effects obtained, we then recomputed the same analysis on 5,000 sets of randomly permuted data, for which no significant effect is expected. A permutation was defined by randomly attributing the label “deviant” or “standard” to the two conditions of interest for each subject. For each permutation, we extracted the distribution of effect strength (number of pairs of clusters of each strength), as well as the value of the maximum effect strength for the total time interval. Uncorrected p-values were obtained by averaging the distribution of effect strength over all the permutations, and then taking the rank of the real experimental data within this average distribution. Similarly, corrected p-values were given by the rank of the experimental data within the distribution of the maximal effect strength, divided by the total number of permutations [68].
The previous method allowed selecting the time windows and clusters of electrodes that were significantly affected by our experimental conditions. Possible effects of group and hemisphere on these effects were further explored with analyses of variance (ANOVAs) in order to assess (1) the consistency of the effect over different groups, (2) the independence of responses to number changes and object changes, and (3) possible asymmetries in the cerebral response. Voltages were extracted for each of the four experimental conditions (DN, SN, DO, and SO) within symmetrical groups of electrodes and a fixed time window (750–850 ms), identified by the above cluster analysis as a common window for number and object change effects. The electrode groups chosen corresponded to the six electrodes located at the negative and positive maxima of the effect, as well as their symmetrical electrodes (identical to themselves for medial electrodes). This way, two symmetrical groups of ten parietal electrodes and 12 prefrontal electrodes were defined for the number-change effects; and two groups of ten occipital electrodes and 12 central electrodes were defined for the object-change effect. Average voltages were then entered into two distinct ANOVAs testing respectively for numerical change and object change, with one between-subject factor (number size: 2/3, 4/8, and 4/12) and three within-subject factors of hemisphere, electrode group localization, and deviance (change or no change). Finally, in order to evaluate the possibility that these number pairs caused deviancy effects at different moments in time, another ANOVA, with a supplementary factor of time, tested six successive 100-ms–long time windows (from 650 to 1,250 ms).
Cortical current density mapping was obtained using a distributed model consisting of 10,000 current dipoles. Dipole locations and orientations were constrained to the cortical mantle of a generic head and brain model built from a 3-mo-old infant MRI anatomy using the BrainVisa software package (http://brainvisa.info/). The geometry of the EEG sensor net was then warped to the head mesh. EEG forward modelling was computed using an overlapping-sphere analytical model in which all sphere and conductivity parameters were adjusted to the typical tissue properties of the infant head [69]. Cortical current maps were computed from the EEG time series using a linear inverse estimator (weighted minimum-norm current estimate). All these procedures were conducted with the BrainStorm Matlab toolkit (http://neuroimage.usc.edu/brainstorm).
|
10.1371/journal.pntd.0005718 | Multiple introductions of the dengue vector, Aedes aegypti, into California | The yellow fever mosquito Aedes aegypti inhabits much of the tropical and subtropical world and is a primary vector of dengue, Zika, and chikungunya viruses. Breeding populations of A. aegypti were first reported in California (CA) in 2013. Initial genetic analyses using 12 microsatellites on collections from Northern CA in 2013 indicated the South Central US region as the likely source of the introduction. We expanded genetic analyses of CA A. aegypti by: (a) examining additional Northern CA samples and including samples from Southern CA, (b) including more southern US populations for comparison, and (c) genotyping a subset of samples at 15,698 SNPs. Major results are: (1) Northern and Southern CA populations are distinct. (2) Northern populations are more genetically diverse than Southern CA populations. (3) Northern and Southern CA groups were likely founded by two independent introductions which came from the South Central US and Southwest US/northern Mexico regions respectively. (4) Our genetic data suggest that the founding events giving rise to the Northern CA and Southern CA populations likely occurred before the populations were first recognized in 2013 and 2014, respectively. (5) A Northern CA population analyzed at multiple time-points (two years apart) is genetically stable, consistent with permanent in situ breeding. These results expand previous work on the origin of California A. aegypti with the novel finding that this species entered California on multiple occasions, likely some years before its initial detection. This work has implications for mosquito surveillance and vector control activities not only in California but also in other regions where the distribution of this invasive mosquito is expanding.
| Infectious diseases transmitted by Aedes aegypti, also known as the yellow fever mosquito, are of growing concern in tropical and subtropical regions. Dengue and Zika incidences are increasing, and no vaccines are currently available. Here we investigate the origin of California A. aegypti and find that this mosquito likely entered California on multiple occasions, at least once from the South Central US region and once from the Southwest US/northern MX region. The evidence suggests that the first invasion event likely occurred some years before its initial detection in 2013, despite California’s extensive and active surveillance program, implying that this invasive mosquito can go undetected. Understanding the invasion dynamics, gene flow, and population structure of A. aegypti can improve the monitoring of mosquitoes and prevent outbreaks of vector-borne disease.
| Dengue, Zika, and chikungunya are severe mosquito-borne infectious diseases that are of growing concern in tropical and sub-tropical regions, and yellow fever is re-emerging in many regions [1]. The viruses causing these diseases are primarily transmitted by the mosquito vector Aedes aegypti. The incidence of dengue has increased 30-fold in the past 50 years [2] with an estimated 96 million new cases annually [3]. More than 2 billion people are at risk of infection by one of the four dengue serotypes [4]. Following in the footsteps of a widespread chikungunya epidemic in Asia and the New World, Zika virus has rapidly emerged around the globe. It has spread to more than 40 countries, in some cases causing serious birth defects including microcephaly [5, 6]. With no effective vaccines and limited antiviral therapeutics available for these diseases, vector control remains critically important.
The ancestor of the domestic form of A. aegypti is a zoophilic sub-species called formosus [7]. It is likely from sub-Saharan Africa, where it can still be found [7]. Outside of Africa, A. aegypti is a domestic mosquito that primarily bites humans, lays eggs in manmade water-containers, and can disperse over long distances using human transportation systems [8]. As such, it is a highly successful invasive species that has colonized most tropical and subtropical regions [7, 8]. A. aegypti likely migrated to the Americas in European slave ships in the fifteenth through seventeenth centuries [7, 9], and these ships probably provided an environment that helped select for traits that increased the success of the domestic form of A. aegypti [7]. The first documented epidemic of yellow fever in the New World was in the Yucatan in 1648, and yellow fever was common in Atlantic seaports from the seventeenth through nineteenth centuries, presumably fueled by the arrival of African slaves and infected sailors [8]. Today, populations of A. aegypti are distributed throughout most of the southern United States, especially below the 33-degree north latitude line [10]. Populations are sporadically found in the Mid-Atlantic States and New England, including an overwintering population in Washington D.C. [10, 11]. Locally transmitted disease from A. aegypti is not common in the United States, but there were locally transmitted cases of Zika in Miami-Dade County, Florida and Cameron County, Texas in 2016 [12]. There has been local transmission of dengue and chikungunya in Florida as recently as 2013 and 2014, respectively [13, 14].
Population genetics plays important roles in both understanding the natural history of A. aegypti and in implementing vector control. Validated genetic markers can be used to determine the source of new invasions [15, 16], detect bottlenecks potentially caused by vector control or founder effect, determine a population’s susceptibility to different classes of insecticide, infer connectivity or isolation between populations, and determine if seasonal appearance of A. aegypti is due to new introductions each year or overwintering. Mosquitoes are more likely to be present and abundant in an area where they can survive the winter, and overwintering is especially concerning in cases where the vectored viruses can also persist through the winter, either in inseminated females or through vertical transmission from mother to eggs. (There is evidence for transovarial transmission of dengue and vertical transmission of Zika in A. aegypti [17, 18].) A stable, overwintering population would be expected to be genetically stable from year to year, although this can be difficult to distinguish from a population that is re-founded each year from the same source population.
California (CA) has an extensive mosquito-monitoring program, and historically A. aegypti were only occasionally detected near airports and other ports of entry [19]. Breeding populations were first reported in 2013 from Fresno, Madera, and San Mateo. Gloria-Soria et al. concluded that the likely origin of these populations was the South Central US, particularly Houston or New Orleans [20]. Through the end of 2016, A. aegypti had been found in 96 cities and census designated places from 12 different counties in CA [21, 22].
In this analysis, we have built upon our previous work by adding 13 new samples from CA, including 8 from southern CA (Fig 1A). The primary goals of the study are 1) to determine whether A. aegypti populations in CA originated from a single or from multiple introductions, 2) to characterize the genetic structure of CA A. aegypti populations, and 3) to determine if the genetic data are consistent with overwintering by A. aegypti, especially in the northern parts of CA. We found clear genetic differentiation between the Northern and Southern CA populations and found support for the hypothesis that at least two introductions of A. aegypti into CA are responsible for the current populations within the state.
A total of 34 samples of A. aegypti mosquitoes from 12 sites in CA and 16 sites from across the southern United States and northern Mexico were considered in analyses (Table 1). The mean sample size per collection was 39 individuals (range: 6–150). Ten of the California samples were collected between May and September of 2015, and an additional eight were collected in 2013–2014. In several cases, multiple collections were made from the same site in different years, or in different areas of the same site in a single year (as noted in Table 1). All mosquitoes from 2015 were collected as adults or eggs from traps and were shipped as adults to our laboratory for analysis. The collections made prior to 2015 are described elsewhere [15, 16, 20]. To avoid biased sampling of siblings, when ovitraps were the source of our sampling we used eggs from four or more traps from any locality with no more than six genotyped individuals per trap. Given that Aedes aegypti are “skip ovipositors” (normally laying one or a few eggs in multiple containers) [23], the use of multiple traps should be sufficient to minimize the sampling of siblings.
For convenience, we have grouped the samples into five broad geographic regions referred to throughout this paper as Southern California, Northern California, Southwest US, South Central US, and Southeast US. The regions are described in Table 1 and shown in Fig 1B.
Whole genomic DNA was extracted from 286 whole adult mosquitoes from ten CA sites collected in 2015 using the Qiagen DNeasy Blood and Tissue kit according to manufacturer instructions, including the optional RNAse A step. All individuals were genotyped at 12 highly variable microsatellites, as in Brown et al. [15]. The microsatellite loci are A1, B2, B3, A9 (tri-nucleotide repeats), and AC2, CT2, AG2, AC4, AC1, AC5, AG1, and AG4 (di-nucleotide repeats) [15, 24]. These loci have been validated previously for their ability to distinguish A. aegypti populations around the world [15].
A total of 107 individuals from ten CA samples (as noted in Table 1) were genotyped at 50,000 single-nucleotide polymorphisms using the high-throughput genotyping chip, Axiom_aegypti1 [25]. Cost prohibited the genotyping of all individuals, so we chose 5 Northern and 5 Southern CA populations instead. After excluding individuals that did not genotype at all microsatellites (likely due to poor DNA quality), we chose 6–12 arbitrary individuals from each population. These data were pruned as described below. Genotyping was subsequently conducted by the Functional Genomics Core at University of North Carolina, Chapel Hill.
All SNP data is available in S1 File as a VCF file, and all microsatellite data is available in S2 File. Additionally these data will be publicly available at Vectorbase.org, Population Biology Project ID: VBP0000177.
All microsatellite loci were tested for within-population deviations from Hardy-Weinberg equilibrium and for linkage disequilibrium among loci pairs using the online version of GENEPOP [26, 27] with 10,000 dememorizations, 1,000 batches, and 10,000 iterations per batch for both tests. To correct for multiple testing, a Bonferroni correction was applied at the 0.05 α level of significance.
Observed heterozygosity (HO) and expected heterozygosity (HE) were calculated using the software GenAlEx 6.5 [28, 29], and allelic richness was estimated by rarefaction (N = 30) using the software HPRARE [30].
To identify likely genetic clusters and possible origins for each cluster, we used a Bayesian clustering method implemented by the software STRUCTURE v. 2.3.4 [31]. STRUCTURE identifies K genetic clusters and estimates what proportion of each individual’s ancestry is attributable to each cluster, with no a priori location information about the individuals. Twenty independent runs were conducted at K = 1–15 for the full set of CA and North American reference populations and at K = 1–12 for the subset of just CA populations. We ran each for 600,000 generations with 100,000 discarded as burn-in, assuming an admixture model and correlated allele frequencies. The optimal number of K clusters was chosen using the guidelines from Prichard et al. [31] and the Delta K method [32, 33]. The results were visualized using the program DISTRUCT v.1.1 [34].
To further explore population structure, discriminant analyses of principle components (DAPC), Principle Component Analyses (PCA), and plots illustrating FST values were created using the Adegenet package v. 2.0.2. [35], available on R software v. 3.2.4 and RStudio v.0.99.893 [36]. DAPC optimizes variation between clusters while minimizing variation within them. Data are transformed using a PCA and then clusters are identified using discriminant analysis. We assessed genetic differentiation among population pairs by calculating FST values with GenoDive v. 2.0b27 [37]. We also ran isolation by distance (IBD) analyses for all populations and for all CA populations using Genodive v. 2.0b27.
While the SNP chip has 50,000 probes, only 27,674 passed the initial stringent testing requiring unambiguous genotyping, biallelic and polymorphic markers, and Mendelian inheritance [25]. Further filtering was done using PLINK v.1.9 to exclude alleles showing up in <1% of samples as these could be genotyping errors, as well as loci not conforming to Hardy-Weinberg expectations (threshold of 0.00001), and those that genotyped in <98% of the samples [38, 39]. These filtering parameters are standard for SNP chip data [40–43]. The dataset contained 15,698 SNPs after this filtering.
For SNP data, we ran four runs with the Bayesian program fastSTRUCTURE to estimate the number of genetic clusters and calculate ancestry fractions for each individual given K numbers of genetic clusters [44]. The results were visualized using DISTRUCT v.1.1. For comparison, we also used the maximum likelihood software Admixture 1.3.0 and the CV error method described in the software’s manual to estimate the number of genetic structures and visualize the ancestry fractions calculated for each individual [45]. PCA analyses were conducted in both Adegenet and PLINK and plotted in R v. 3.2.4.
We inferred demographic history and estimated relevant parameters using microsatellite data and Approximate Bayesian Computation methods [46] as implemented by the program DIYABC [47]. Four colonization scenarios were tested to determine if the current Californian populations are more likely to be the result of one or two introduction events (Fig 2). In the first scenario, Northern California populations originate from an invasion from South Central US, and Southern California populations from an invasion from the Southwest US. In the second scenario the origins are reversed; Northern California comes from Southwest and Southern California from South Central. The third scenario depicts just one invasion into California, and the fourth scenario is a neutral model in which all four populations branch from a common ancestor at the same time. We ran the analysis using two different datasets. First, to reduce the excess noise that can result from grouping disparate populations together, we ran the analyses with regional groups that were made up of 133 randomly chosen individuals from populations that were representative of the geographic regions, as identified in STRUCTURE (Fig 3C) and noted S1 Table. For example, San Mateo, Madera, and Fresno always clustered together, so they were chosen to represent Northern California (Fig 3). To assess whether we were biasing the analysis by excluding some populations in this first test, we ran the analysis again this time forming the four regional groups from 217 randomly selected individuals from all populations from the respective geographic region. In both cases, the number of individuals was chosen based on the number of individuals in the smallest of the four groups.
For the DIYABC analyses, we first simulated 1,000,000 datasets for each scenario, resulting in 4,000,000 total simulated datasets. To determine which scenario was most supported by the data, we evaluated the relative posterior probability using a logistic regression on the 4,000 (1%) simulated datasets closest to the observed dataset. To estimate demographic parameters, we chose scenario 1 and estimated posterior distributions of parameters taking the 1,000 (1%) closest simulated datasets, after applying a logit transformation of parameter values. To evaluate confidence in the posterior probability of scenarios (in the form of Type I and Type II errors), we used a logistic regression on 250 test datasets simulated for each scenario with the same values that produced the original dataset. Priors and parameters are provided in S3 and S4 Tables.
For microsatellites, 17 out of 2,241 (0.76%) loci pairs were found to be in linkage disequilibrium, and 3 out of 360 (0.83%) locus-population pairs were not in agreement with Hardy Weinberg equilibrium after Bonferroni correction for multiple comparisons. This is consistent with previous work indicating that these 12 microsatellites are single-copy and can be treated as independent loci for population genetic analyses [15, 16].
Genetic diversity was significantly lower within CA populations than among other North American populations (Table 2 and S1 Table). The mean observed heterozygosity +/- the standard deviation (SD) for CA sites was 0.45±0.088, and the mean of all other sites included in analysis was 0.57±0060 (Student’s t-Test, p = 0.0079). This trend was largely due to Southern California populations; 5 out of the 10 populations with the lowest heterozygosity in this analysis were located in Southern California. Similarly, the mean estimated allelic richness for California, 3.05±0.64, was lower than the mean of other populations, 3.95±0.51 (Student’s t-Test, p = 0.0013), due to Southern California. The mean allelic richness of the Southern California populations (2.62±0.41) was significantly lower than the mean from the Northern California populations (3.48±0.52) (Student’s t-Test, p<0.0001). Nine out of 10 of the analyzed sites with the lowest allelic richness were from California, and 8 of these were from Southern California.
Among California populations, pairwise FST values ranged from 0.0010 to 0.42 (S2 Table). Excluding Exeter as an outlier, pairs of Northern California sites had low mean FST values (0.060±0.051 SD). In contrast, Exeter and the Southern California sites had significantly higher mean FST values when paired with themselves (0.26±0.11 SD) or with all the CA populations (0.26±0.10) (Student’s t-Test, p<0.0001). S1 Fig illustrates this pattern graphically and in the context of all analyzed North American populations. The IBD analyses on all the populations did not show a correlation between distance and FST values (Spearman’s r = -0.19; p = 0.010; R2 = 0.036), and the IBD test on just the populations from CA was not significant (Spearman’s r = 0.11; p = 0.14; R2 = 0.011).
In regard to overwintering, the FST values between Madera 2015 and the two Madera 2013 populations is lower (0.01 and 0.047) than between Madera 2015 and any of the sites in the Central South populations (range = 0.091–0.15). The FST between San Mateo 2014 and San Mateo 2015 is higher (0.16) than between San Mateo 2014 and some of the Central South populations, such as New Orleans (0.13).
Bayesian clustering analysis for microsatellite data on all CA populations identified two primary clusters (K = 2), which divided the Northern California populations from the Southern California populations with the exception of Exeter and most of the Clovis individuals (Fig 3A). At higher K values, STRUCTURE showed a high level of population differentiation between each of the southern California populations, Exeter, and the two Clovis populations (Fig 3B and S2 Fig). In contrast, San Mateo, Madera, and Fresno populations always clustered together (Fig 3B and S2 Fig). At higher Ks, Bayesian clustering analysis that included Northern CA, the South Central, and the Southeast consistently showed that San Mateo 2013, San Mateo 2014, Madera 2013, and Madera 2014 formed a cluster that was separate from the South Central populations (eg. S3 Fig). The results from these higher K values were also found with the SNP data, as described below.
A PCA analysis using microsatellites from California populations found that the first Principal Component accounted for 13.17% of the variation and the second accounted for 9.71%. Using the same data for a DAPC analysis, the “find.clusters” command in Adegenet found the data could best be described by 12 clusters (S4 Fig). The first axis on the DAPC plot corresponds with the north-south gradient and explains 27.26% of the total variance; the second axis highlights the uniqueness of Exeter and explained 20.21% of the total variance (S4 Fig). In a DAPC plot using populations as priors, the first axis corresponds relatively well to the north-south gradient and explained 38.67% of the total variance, while the second axis explained 17.70% of the total variance (S5 Fig).
Analysis of the SNP data largely reinforced the microsatellite results. Seven genetic clusters were identified by fastSTRUCTURE and six by Admixture’s CV Error method of K selection on the final dataset of 15,698 SNPs (Fig 4 and S6 Fig). San Mateo, Madera, and Fresno clustered together, and each of the other populations formed its own genetic cluster (although Admixture did not distinguish between San Diego and Los Angeles) (Fig 4 and S6 Fig). The clusters observed in the PCA on the SNP data corresponded to those identified by Admixture (Fig 5). Using Adegenet, the first Principal Component explained 8.06% of the variation and was correlated with the north to south gradient, with the exception of Exeter (Fig 5). The second Principal Component accounted for 6.64% of the variation and highlighted the uniqueness of Garden Grove and Exeter.
Including microsatellite data from all analyzed populations from this study (Table 1), the Bayesian clustering method implemented by STRUCTURE identified two ancestral groups (K = 2) (Fig 3C). Most of the Northern California populations clustered with the South Central US and Southeast US populations, while most of the southern California populations clustered with the Southwest US populations (Fig 3C).
The Bayesian clustering method implemented by STRUCTURE showed Northern California clustering with the South Central and Southeast regions, and Southern California clustering with the Southwest region. Since Gloria-Soria et al. showed that Houston or New Orleans was the likely origin of the Northern Californian populations [20], we used a simulation to test the hypothesis that Northern California populations originated from South Central populations and that Southern California populations originated from Southwest populations. We ran the analysis first with individuals from populations that were representative of their regions (Table 1), and secondly without excluding any populations. Since the results are similar (S3 and S4 Tables) with one described exception, all results refer to the first analysis unless otherwise noted.
Evaluation of the relative posterior probability of each of the four competing scenarios in Fig 2 supported Scenario 1 as the most plausible invasion scenario (S3 Table), in which Northern California populations split from the Central South, and Southern California populations split from the Southwest populations (p = 0.990 CI95%: 0.983–0.996). Scenario 4, the neutral model in which the four populations diverge from the same ancestor simultaneously, had the next highest posterior probability (0.0079 CI95%: 0.0019–0.0138).
The type I error rate was 0.14, and the type II error rates under the three other scenarios were 0.16, 0.11, and 0.064. The estimated time of divergence of the Central South from the Southwest populations was 4,200 generations (approximately 420 years), and the 95% credible interval was 1,730–5,900 generations assuming 10 generations/year. The estimated time of divergence of Northern California populations from South Central populations was 292 generations or approximately 29 years (95% credible interval = 61–853 generations), and the estimated time of divergence of Southern California from the Southwest was 224 generations or approximately 22 years (95% credible interval = 39.9–774). Full details are provided in S3 Table.
The results from the second analysis (in which no populations were excluded) were mostly similar including a high support for Scenario 1 (p = 0.9995 CI95%: 0.9991–0.9999). The estimated time of divergence between Southern California and Southwest was approximately 35 years (mean = 348, 95% credible interval = 66.9–893), and the estimated time of divergence between Northern California and the South Central/Southeast was approximately 8 years (mean = 78.0, credible interval = 15.6–414). Full details are provided in S4 Table.
Our analyses suggest multiple introductions of A. aegypti into CA that came from at least two different regions in North America. As previously shown, the populations in Northern California likely originated from mosquitoes that were introduced from the South Central region of the US [20]. The results of this study also suggest that a second introduction event likely occurred from the Southwest/northern MX region, and that these mosquitoes gave rise to the current populations found in Southern California. Given the considerable distance between the Northern CA and Southern CA populations (>275km from Exeter [E] to Los Angeles [F]) and the recentness of the invasions, it is not surprising that the two groups maintain distinct signatures of their genetic ancestry. As more populations are discovered between Northern and Southern CA (for example, Kern County), it would be interesting to add them to this analysis to identify whether the north-south break is clean or if the transition occurs as a cline. Identifying the area where the two clusters meet could eventually help us understand the factors leading to this genetic break.
The DIYABC analysis using representative populations suggests that the Northern California lineage diverged from the South Central lineage 292 generations ago, estimated to be ~29 years, and that the Southern California lineage diverged from the Southwest/MX lineage ~22 years ago. These estimates have large credible intervals, so we cannot tell which invasion into California happened first. The lower bounds of the intervals containing 95% probability are both more than 3 years, so it seems likely the invasions occurred at least a year prior to the initial detection of A. aegypti in 2013. In our second run of DIYABC, which did not exclude individuals from potential outliers (Clovis and Exeter), a more recent time of divergence (~8yrs) was estimated between Northern California and South Central/Southeast. This is more consistent with personal communications from CA vector control professionals who think it is unlikely the invasion could have occurred more than a year or two prior to initial detection.
Invasion events are often accompanied by a bottleneck in population size and subsequent decrease in genetic diversity, especially allelic richness. Consistent with Gloria-Soria et al. [20], we found that Northern California populations have similar levels of genetic diversity and allelic richness as other US populations. However, Southern California populations are less genetically diverse. This relatively low diversity is a possible signature of bottleneck(s) caused either by relatively recent founder effects and/or vector control measures that reduced A. aegypti population size.
We observed significant population structure with genetic differentiation even among populations in close geographic proximity, particularly among Southern California populations. For example, Anaheim, Orange, Garden Grove, and Santa Ana are each less than 12km apart from the others, but they are genetically distinguishable (S2 Fig). The dense highway system and highly discontinuous human habitats in Southern California could well cause A. aegypti to be broken into almost entirely isolated small local populations given the evidence suggesting this mosquito avoids crossing major roads/highways [48]. Similarly, Clovis and Fresno are less than 12km apart but genetically distinct (Figs 3 and 4). On the other hand, San Mateo is genetically indistinguishable from Madera and Fresno, despite the >200 km between them. Likely because of patterns like these, isolation by distance analyses did not explain the population structure in CA or throughout the regions sampled.
The large pairwise FST values between Southern California sites may indicate limited gene flow among the populations, consistent with the relatively short active dispersal distance that has been found for A. aegypti [23]. The observations are consistent with the possibility that effective A. aegypti control in one locality may not be easily influenced by migration from neighboring localities. Further analysis taking into account both timing of invasion and connectivity through human transportation routes may help us understand the patterns we see in the structure of both Southern and Northern CA populations.
In two cases, we analyzed CA populations from the same location from two different years: Madera 2013 and 2015, and San Mateo 2013 and 2014. We found strong evidence for overwintering in Madera and mixed results for San Mateo. The FST values between Madera 2013 and 2015 were lower than any of the FST values between Madera and populations from South Central US, suggesting this is a stable population and not one recolonized each year. The genetic diversity also changed very little: for example, the allelic richness was 3.5 in 2013 and 3.6 in 2015. San Mateo 2013 and San Mateo 2014 did not follow these patterns, and the small sample size of San Mateo 2014 (N = 7) could be a factor. It is unlikely that this change in genetic diversity is due to recolonization of the area by another northern California population because the mosquitoes were collected from the same confined area in 2013 and 2014, and were not detected elsewhere in the county. In both cases the populations are always part of the same genetic cluster determined with Bayesian clustering (e.g. Fig 3), consistent with a previous study demonstrating that temporal differentiation does not obscure geographic structure in CA A. aegypti populations [16]. Strikingly, STRUCTURE analyses of North American populations at higher K values showed that San Mateo and Madera formed a genetic cluster separate from other Central South and Southeast populations (S3 Fig). This temporal stability indicates that some populations in Northern California are stable and likely continuously breeding in situ, rather than being recolonized by their original source. In five cases, we have included multiple populations from the same city and the same years. Except for a few anomalies (e.g. a subset of Clovis), these populations did not show genetic differentiation, suggesting no population structure within a city.
Our paper is the first to combine SNP data and microsatellites to address the population structure and origins of the CA A. aegypti. We found the same genetic structure using both types of markers, which speaks to the robustness of our methods and results. Our results suggest the microsatellite markers, which are more cost-efficient, are sufficient for these types of analyses. However, we expect the SNP chip will continue to provide essential information in other situations, for example, at finer-scales, when fewer individuals are available, or when building phylogenies.
That A. aegypti invaded CA multiple times, probably years before its first detection in 2013, has important implications for vector control. It implies CA and other regions with a temperate climate may be more vulnerable to invasion than previously thought. Additionally, we find that CA A. aegypti populations near one another are often genetically distinct. A challenge to providing effective vector control is the potential for reinvasion in targeted regions. At least in Southern CA it appears that there is little evidence for extensive migration and gene flow among populations. The disparate genetic backgrounds in these A. aegypti populations may represent populations capable of responding differently to control measures such as pesticides, or perhaps these populations may have different vector competence for infectious pathogens. CA has one of the most extensive mosquito-monitoring systems in the US, so the possibility that A. aegypti was in CA years before detection may mean mosquito invasions have occurred elsewhere in the US but escaped notice. Understanding and accounting for the invasion dynamics of A. aegypti will continue to be essential for detecting new invasions, monitoring vector presence, and preventing disease outbreaks in California and other regions.
|
10.1371/journal.pgen.1006529 | A Hidden Markov Model Approach for Simultaneously Estimating Local Ancestry and Admixture Time Using Next Generation Sequence Data in Samples of Arbitrary Ploidy | Admixture—the mixing of genomes from divergent populations—is increasingly appreciated as a central process in evolution. To characterize and quantify patterns of admixture across the genome, a number of methods have been developed for local ancestry inference. However, existing approaches have a number of shortcomings. First, all local ancestry inference methods require some prior assumption about the expected ancestry tract lengths. Second, existing methods generally require genotypes, which is not feasible to obtain for many next-generation sequencing projects. Third, many methods assume samples are diploid, however a wide variety of sequencing applications will fail to meet this assumption. To address these issues, we introduce a novel hidden Markov model for estimating local ancestry that models the read pileup data, rather than genotypes, is generalized to arbitrary ploidy, and can estimate the time since admixture during local ancestry inference. We demonstrate that our method can simultaneously estimate the time since admixture and local ancestry with good accuracy, and that it performs well on samples of high ploidy—i.e. 100 or more chromosomes. As this method is very general, we expect it will be useful for local ancestry inference in a wider variety of populations than what previously has been possible. We then applied our method to pooled sequencing data derived from populations of Drosophila melanogaster on an ancestry cline on the east coast of North America. We find that regions of local recombination rates are negatively correlated with the proportion of African ancestry, suggesting that selection against foreign ancestry is the least efficient in low recombination regions. Finally we show that clinal outlier loci are enriched for genes associated with gene regulatory functions, consistent with a role of regulatory evolution in ecological adaptation of admixed D. melanogaster populations. Our results illustrate the potential of local ancestry inference for elucidating fundamental evolutionary processes.
| When divergent populations hybridize, their offspring obtain portions of their genomes from each parent population. Although the average ancestry proportion in each descendant is equal to the proportion of ancestors from each of the ancestral populations, the contribution of each ancestry type is variable across the genome. Estimating local ancestry within admixed individuals is a fundamental goal for evolutionary genetics, and here we develop a method for doing this that circumvents many of the problems associated with existing methods. Briefly, our method can use short read data, rather than genotypes and can be applied to samples with any number of chromosomes. Furthermore, our method simultaneously estimates local ancestry and the number of generations since admixture—the time that the two ancestral populations first encountered each other. Finally, in applying our method to data from an admixture zone between ancestral populations of Drosophila melanogaster, we find many lines of evidence consistent with natural selection operating to against the introduction of foreign ancestry into populations of one predominant ancestry type. Because of the generality of this method, we expect that it will be useful for a wide variety of existing and ongoing research projects.
| Characterizing the biological consequences of admixture—the mixing of genomes from divergent ancestral populations—is a fundamental and important challenge in evolutionary genetics. Admixture has been reported in a variety of natural populations of animals [1,2], plants [3–5] and humans [6,7], and theoretical and empirical evidence suggests that admixture may affect a diverse suite of evolutionary processes. Individuals’ ancestry can affect disease susceptibility in admixed populations, and inferring and correcting for sample population ancestries is a common practice in human genome wide association studies [8–10]. More generally, admixture has the potential to influence patterns of genetic variation within populations [11,12], to introduce novel adaptive [13,14] and deleterious variants [7,15,16], as well as to disrupt epistatic gene networks [17,18]. Therefore, developing a comprehensive understanding of the extent of admixture in natural populations and resulting mosaic genome structures is essential to furthering our understanding of a variety of evolutionary processes.
Estimating genome-wide ancestry proportions has become a common practice in population genetic inference. For example, the program STRUCTURE [19], originally released in 2000, uses a Bayesian framework to model the ancestry proportions of individuals derived from any number of source populations based on genotype data at a set of unlinked genetic markers. More recently, this model for ancestry proportion estimation has been extended to cases where individual genotypes are not known, but can be studied probabilistically using low-coverage sequencing short read sequencing data [20], which is an important step towards accommodating modern sequencing practices. Additionally, Bergland et. al. [21] developed a method for estimating ancestry proportions in pooled population samples of relatively high ploidy (i.e. 40–250 distinct chromosomes) from short read sequencing data. In general, it is straightforward to estimate genome-wide ancestry proportions using a number of sequencing strategies and applications.
It is substantially more challenging to accurately estimate local ancestry (LA) at markers distributed along the genome of a sample. Nonetheless, analyses of LA have the potential to yield more nuanced insights into our understanding of the evolutionary processes affecting ancestry proportions across the genome. One of the first LA inference (LAI) methods was an extension of the STRUCTURE [19] framework that modeled the correlation in ancestry among markers due to linkage. Because the ancestry at each locus is not observed, Falush et al. [22] suggested that a hidden Markov model (HMM) is a straightforward means of inferring the ancestry states at each site in the genome (which are unobserved) based on observed genotype data distributed along a chromosome. Most subsequent LAI methods have also used an HMM framework, and the majority are geared towards estimating LA in admixed human populations (e.g. [23,24]). Consequently, most existing LAI methods are limited to diploid genomes with high quality genotype calls. Furthermore, many methods require phased reference panels [24,25], and require the user to provide an estimate of, or make implicit assumptions about, the number of generations since the initial admixture event [2,23–25]. This is straightforward with human population genomic samples, where abundant high quality genotyped samples are available and for which well-documented demographic histories are sometimes known. However for most other species, demographic histories are less well characterized, and assumptions about admixture times may bias the result of LAI methods.
A number of approaches exist to estimate the time since admixture based on well characterized ancestry tract length distributions [26–29] but in general, these parameters are unknown prior to LAI. Conversely, another class of methods can be used to estimate the time of admixture based on the decay of linkage disequilibrium without performing LAI [30–32]; however as with LAI procedure above, these approaches are also limited to diploid genotype data. We may therefore expect to improve LAI by simultaneously estimating LA and demographic parameters (e.g. admixture time). Furthermore, in the majority of sequencing applications, relatively low individual sequencing coverage is often used to probabilistically estimate individual and population allele frequencies (e.g. [33]) but these data are often not sufficient to determine high confidence genotypes that are required for existing LAI applications. Hence, there is a clear need for a general LAI method that can accommodate genotype uncertainty and requires less advanced knowledge of admixed populations’ demographic histories.
Here, we introduce a framework for simultaneously estimating LA using short read pileup data and the time of admixture within a population. Briefly, as with many previously proposed LAI methods, we model ancestry across the genome of a sample as a HMM. We estimate LA by explicitly modeling read counts as a function of sample allele frequencies within an admixed population. Our method is generalized to accommodate arbitrary sample ploidies, and is therefore applicable to haploid (or inbred), diploid, tetraploid, as well as pooled sequencing applications. We show that this approach accurately infers the time since admixture when data are simulated under the assumed model. Furthermore, our method yields accurate LA estimates for simulated datasets, including samples of high sample ploidy and including evolutionary scenarios that violate the assumptions of the neutral demographic model. In comparisons between ours and an existing LAI method, WINPOP [34], we find that our approach offers a significant improvement and is accurate over longer time scales. Furthermore, we demonstrate, using a published dataset, that even state-of-the-art LAI methods can be significantly impacted by assumptions about the time since admixture, and that our method provides a solution to this problem.
Finally, we apply this method to a Drosophila melanogaster ancestry cline on the east coast of North America. This species originated in sub-Saharan Africa, and approximately 10,000–15,000 years ago a subpopulation expanded out of the ancestral range. During this expansion, the derived subpopulation experienced a population bottleneck that resulted in decreased nucleotide polymorphism, extended linkage disequilibrium within the derived population and substantial genetic differentiation between ancestral and derived populations [2,35–39]. Hereafter, the ancestral population will be referred to as “African” and the derived population as “Cosmopolitan”. Following this bottleneck, descendant populations of African and Cosmopolitan D. melanogaster have admixed in numerous geographic regions [2,11,21]. Of particular relevance to this work, North America was colonized recently by a population descendent from African individuals from the South, and by a population descendent from cosmopolitan D. melanogaster in the North [11,21,38]. Where these populations encountered each other in eastern North America, they form an ancestry cline where southern populations have a greater contribution of African ancestry than northern populations [21].
Previous work on these ancestry clines has shown that ancestry proportions vary across populations with increasing proportions of cosmopolitan alleles in more temperate localities. Evidence suggests spatially varying selection affects the distribution of genetic variants [40–45]. Furthermore, strong epistatic reproductive isolation barriers partially isolate individuals from northern and southern populations along this ancestry cline [46,47]. This may be generally consistent with recent observations of ancestry-associated epistatic fitness interactions within a D. melanogaster population in North Carolina [17], and with the observation of widespread fitness epistasis between populations of this species more generally [48]. There is therefore good reason to believe that natural selection has acted to shape LA clines that are tightly linked to selected mutations in these D. melanogaster populations.
Here, we show that African ancestry in North American D. melanogaster populations is negatively correlated with recombination rates, consistent with more efficient selection against foreign ancestry in high recombination rate regions of the genome. We also find that the X chromosome displays a higher rate of LA outlier loci, potentially consistent with a greater role of the X chromosome in clinal adaptation. Clinal loci are disproportionately likely to be associated with high level gene regulatory protein complexes, and may play important roles in ecological divergence between African and Cosmopolitan D. melanogaster populations. Furthermore, we identify numerous loci with decreased African ancestry across all populations, which suggests that these alleles that are disfavored on predominantly cosmopolitan genetic backgrounds. This subset of loci is enriched for genes related to oogenesis, potentially consistent with epistatic interactions that affect female reproductive success in these populations.
Although admixed populations often are diploid, we derived a general model of ploidy in which the individual has n gene copies at each locus, i.e. for diploid species n = 2. In practice, sequences are often obtained from fully or partially inbred individuals (e.g. [39,49]), which represent only a single uniquely derived chromosome. It is also common to pool individuals prior to sequencing for allele frequency estimation, so called pool-seq (e.g. [21,40,42,50–53]). If the pooling fractions are exactly equal, such a sample of b diploid individuals can be treated as a sample from a single individual with ploidy n = 2b. Although that requirement is restrictive, pool-seq has been experimentally validated as a method for accurate allele frequency estimation—i.e. alleles are approximately binomially sampled from the sample allele frequencies [54]. We therefore aimed to derive a model that can accommodate arbitrary sample ploidies. In the model, we assumed that the focal population was founded following a single discrete admixture event between two ancestral subpopulations, labeled 0 and 1, with admixture proportions 1-m and m, respectively, at a time t generations in the past. We modeled emission probabilities such that the method can work directly on read pileup data, rather than high quality known genotypes. Briefly, in our model, we specify an HMM {Hv} with state space S = {0,1,…,n}, where Hv = i, i ∈ S, indicates that in the vth position i chromosomes are from population 0 and n–i chromosomes are from population 1. In other words, this HMM enables one to estimate what ancestry frequencies are present at a given site along a chromosome within a sample. Importantly, we designed this method to simultaneously estimate the time of admixture, which is related to the correlation between ancestry informative markers along a chromosome. See Methods for a complete description of the HMM including the emissions and transition probability calculations. The source code and manual are available at https://github.com/russcd/Ancestry_HMM. For this model, it is assumed that the number of chromosomes present in a sample, n, is known and that the global ancestry proportion, m, is known. As there are many methods for accurately estimating m in a wide variety of contexts implemented in standard population genetic analysis pipelines [19,20], we believe this assumption is not too restrictive.
In order to test our method with data of known provenance, we also developed an approach for simulating chromosomes sampled from admixed populations. Briefly, we first simulated genetic diversity consistent with ancestral populations using a coalescent simulation method [55]. We then generated ancestry tracts consistent with admixture models developed to test our inference method using the forward-time admixture simulation program, SELAM [56]. We retained a portion of each coalescent-generated population to serve as a reference panel for allele frequency and LD estimation. We then took the remaining chromosomes and placed them on the appropriate ancestry tracts in admixed chromosomes. Finally, we generated read counts for these chromosomes, or pools of chromosomes for samples with ploidy greater than one, via binomial sampling from the genotype frequencies of the sample. Implicitly, this procedure assumes that the allele frequencies in the reference panel and the admixed individuals whose ancestry is from a given reference panels are equivalent. For large, well-mixed populations such as those of D. melanogaster, this is likely to be a reasonable assumption. Nonetheless, below we assess the impact of differences in the ancestral allele frequencies for plausible demographic models in this species.
Within an admixed population, there are two sources of LD. LD that is induced due to the correlation of alleles from the same ancestry type (i.e. admixture LD), and LD that is present within each of the ancestral populations (ancestral LD). Admixture LD, is the signal of LA that we seek to detect using the HMM. The second type, ancestral LD, limits the independence of the ancestral information captured by each marker, and is expected to confound HMM-based analyses, particularly as we aimed to estimate the time since admixture within this framework. We therefore sought to quantify the effect of ancestral LD by discarding one of each pair of sites in LD within either ancestral population. We found that ancestral LD tends to increase admixture time estimates obtained using our method, and we decreased the cutoff of the LD parameter, |r|, by 0.1 until the time estimates obtained for single chromosomes were unbiased with respect to the true time since admixture. We found that |r| ≤ 0.4 fit this criterion, although for relatively ancient admixture events with highly skewed ancestry proportions—i.e. m < 0.1 or m > 0.9—some residual bias was apparent in the estimates of admixture time (Fig 1). This reflects the fact that the SMC’ ancestry tract distribution performs poorly with highly skewed ancestry proportions and especially for long times since admixture [57].
Fig 1 also reveals a striking difference between otherwise equivalently skewed admixture proportions. For example when m = 0.1, there was a much larger effect of ancestral LD than when m = 0.9. This is due to differences in the variability of LD within the ancestral populations. That is, due to the strong population bottleneck, cosmopolitan D. melanogaster populations have substantially more LD and fewer polymorphic sites than African D. melanogaster populations. Because the time estimation procedure appears to be sensitive to the amount of ancestral LD present in the data, simulations of the type we described here may be necessary to determine what |r| cutoffs are required to produce unbiased time estimates given the ancestral LD of the populations in a given analysis using this method.
We next sought to quantify the accuracy of our approach across varying sample ploidies and times since admixture (Fig 2). Especially for moderate and short admixture times (i.e. 0–500 generations), our method performed well for all ploidies considered and we were able to accurately recover the correct admixture time with relatively little bias. However, as true admixture time increases, the time estimates for pooled samples become significantly less reliable and show a clear negative bias. Nonetheless, across the range of times presented in Fig 2, samples of ploidy one and two showed little bias, and we therefore believe our method will produce sufficiently accurate admixture time estimates for a wide variety of applications.
All measures of accuracy decrease with increasing time since admixture (Fig 2). However, even for relatively long times since admixture—2000 generations—and for large sample ploidies, the mean posterior error remained relatively low for all ancestry proportions and for long times since admixture. This indicates that this approach may be sufficiently accurate for a wide variety of applications, sequencing depths, and sample ploidies. Nonetheless, the proportion of sites within the 95% credible interval decreased with larger pool sizes and it is clear that for larger pools the posterior credible interval tends to be too narrow. Therefore, correcting for this bias may be necessary for applications that are sensitive to the accuracy of the credible interval.
An important consideration is that estimates of t will be reliable only if the local recombination rates are known with reasonably high accuracy [58]. In many species, an accurate broad-scale map is available. However, fine-scale variation in recombination rates has only been documented for a few model species. Therefore, for relatively short to moderate times since admixture, error in the genetic map is expected to have a limited impact on date estimates. However, for longer times since admixture, this factor has the potential to bias estimates of t [58], particularly in species with large variance in local recombination rates (e.g. due to hotspots). Since D. melanogaster has one of the best recombination maps currently available in any species [59] and because we do not aim to estimate time in our applications, we do not believe this will heavily impact the analyses we present below. However, for most applications, it will be necessary to consider the impact of error in the assumed genetic map to accurately interpret estimates of t obtained using this method. We emphasize that this challenge is not unique to this application, but will impact virtually all ancestry estimation methods that rely on a genetic map for estimating the time since admixture.
As described above, estimates of the time of admixture demonstrate an apparent bias in pools of higher ploidy (Fig 2). Specifically, time tends to be slightly overestimated for relatively short admixture times and underestimated at relatively long admixture times. This is particularly apparent at highly skewed ancestry proportions. Given that this bias is primarily evident in pools of 10 to 20 individuals, we hypothesized that it might be due to the non-independence of ancestry tracts among chromosomes, which should tend to disproportionately affect samples of higher ploidy because all ancestry breakpoints are assumed to be independent in our model. To test this, we simulated genotype data from independent and identically distributed exponential tract lengths as is assumed by our model. When we ran our HMM on this dataset, we found that no bias is evident for simulations of up to 2000 generations (S1 Fig), indicating that the primary cause of this bias was violations in the real data of the independence of ancestry tracts that we assumed when computing the transition probabilities. However, it should be possible to quantify and correct for this bias in applications of this method that aim to estimate the time since admixture.
The transition probabilities of this HMM depend on knowledge of the population size. In practice, this parameter is unlikely to be known with certainty. Hence, to assess the impact of misspecification of the population size, we performed simulations using a range of population sizes that span three orders of magnitude (N = 100, 1000, 10000, and 100000). All analyses presented here were conducted by applying our HMM to haploid and diploid samples, but qualitatively similar results hold for samples of larger ploidy. We then analyzed these data assuming the default population size, 10000, is correct. For relatively short times since admixture, there was not a clear bias for any of the true population sizes considered. However, at longer true admixture times, estimated admixture times for both N = 100 and N = 1000 asymptote at a number of generations near to the population sizes. This result reflects the fact that smaller populations will tend to coalesce at a portion of the loci in the genome relatively quickly, and ancestry tracts cannot become smaller following coalescence. Nonetheless, the accuracy of LAI remained high even when time estimates were unreliable (S2 Fig) for the tested marker densities and patterns of LD. Furthermore, in some cases it should be straightforward to determine if a population has coalesced to either ancestry state at a large portion of the loci in the genome, potentially obviating this issue.
A more subtle departure from the expectation was evident for population sizes that are larger than we assumed in analyzing these data (S2 Fig). This likely reflects the fact that the probability of back coalescence to the previous marginal genealogy to the left after a recombination event is inversely related to the population size. Hence, the rate of transition between ancestry types is actually slightly higher in larger populations where back coalescence is less likely than we assumed during the LAI procedure. This produced a slight upward bias in the estimates of admixture time when the population was assumed to be smaller than it is in reality. However, this bias appears to be relatively minor, and we expect that time estimates obtained using this method will be useful so long as population sizes can be approximated to within an order of magnitude. Of course, this bias is not unique to our application, and it will affect methods that aim to estimate admixture time after LAI as well. That is, estimating the correct effective population size is an inherent problem for all admixture demographic inference methods.
Although it is clear that accurately estimating relatively ancient admixture times is challenging in higher ploidy samples, we sought to determine the limits of our approach for LAI and time estimation for longer admixture times for haploid sequence data. Because of rapid coalescence in smaller samples (see above), we performed admixture simulations with a diploid effective population size of 100,000. It is clear that there is a limit to the inferences that can be made directly using our method. Like the higher ploidy samples, time estimates for haploid samples departed from expectations shortly after 2,000 generations since admixture (S3 Fig). Nonetheless, the magnitude of this bias is slight, and it is likely that it could be corrected for when applying this method even for very ancient admixture events. For all admixture times considered, LAI remained acceptably accurate despite the slight bias in time estimates (S3 Fig).
One question is what effect varying the reference panel sizes will have on LAI inference using this method. We therefore compared results from reference panels of size 10 chromosomes with those from panels of size 100 chromosomes (S4 Fig). As with results obtained for reference panels of size 50, panels of size 100 were sufficient to accurately estimate admixture time and LA over many generations since admixture. Whereas, when panel sizes were just 10 chromosomes, time estimates were clearly biased and the result was variable across ancestry proportions (S4 Fig). However, since there was a strong correlation between true and estimated admixture times even with relatively small panel sizes, it may therefore be possible to infer the correct time by quantifying this bias through simulation and correcting for it. Furthermore, although LAI is clearly less reliable with smaller panels, these results are not altogether discouraging and this approach, in conjunction with modest reference panels may still be effective for some applications.
Ultimately, there are three reasons why allele frequencies in the reference panels and in the admixed population panel would be expected to differ beyond that expected from binomial samples with the same mean. First, some amount of genetic drift may have occurred in the ancestral population and in the admixed population in the time since the admixed population was founded. Second, in some cases, it is infeasible to sample the ancestral population of an admixed group, and a genetically divergent population must suffice as the reference panel if this method is to be used. Third, divergent selection may quickly modify allele frequencies between admixed and ancestral populations. Hence, genetic divergence between reference and admixed populations may be an important challenge for this method.
To address this, we simulated the second scenario, where increasingly divergent populations are used as the reference panels to study admixed populations. In order to make this relevant to the application to D. melanogaster populations, below, we selected times for divergence that might be consistent with differences across continental populations in Sub-Saharan Africa and in Cosmopolitan populations. Although time estimates obtained using this approach are weakly positively biased with increasing divergence between the ancestral population and reference panels, the accuracy of this LAI method is largely unaffected (S5 Fig). Hence, for biological scenarios potentially consistent with those of D. melanogaster ancestral populations, we do not expect this challenge to strongly bias our method. Nonetheless, in applications to other populations, with potentially differently structured ancestral populations, it would be necessary to examine the effects of this bias in detail.
In a wide variety of pool-seq applications, samples are pooled in larger groups than we have considered above (e.g. [40,50,52]). We are therefore interested in determining how our method will perform on pools of 100 individuals. Towards this, we performed simulations as before, but we designed our parameters to resemble those of the pooled sequencing data that we analyze in the application of this method below. Specifically, we simulated data with a mean sequencing depth of 25, a time since admixture of 1500 generations, and an ancestry proportion of 0.8. Consistent with results for ploidy 20, we found that time tends to be dramatically underestimated (i.e. the mean estimate of admixture time was 680 generations). However, when we provided the time since admixture, our method produced reasonably accurate LAIs for these samples. Although the posterior credible interval was again too narrow, the mean posterior error was just 0.053 when expressed as an ancestry frequency, indicating that this approach can produce LA estimates that are close to their true values for existing sequencing datasets (e.g. Fig 3). However, the HMM’s run time increases dramatically for higher ploidy samples and higher sequencing depths, a factor that may affect the utility of this program for some analyses. Nonetheless, for more than 36,000 markers, a sample ploidy of 100 and a mean sequencing depth of 25, the average runtime was approximately 42 hours. In contrast, for the same set of parameters, but where individuals are sequenced and analyzed as diploids, the mean runtime was just 8 minutes (See S1 Table for a comparison of run times across many parameter sets).
An important concern is that many biologically plausible admixture models would violate the assumptions of this inference method. In particular, continuous migration and selection acting on alleles from one parental population are two potential causes of deviation from the expected model in the true data. To assess the extent of this potential bias, we performed additional simulations. First, we considered continuous migration at a constant rate that began t generations prior to sampling. In simulations with continuous migration, additional non-recombinant migrants enter the population each generation. Relative to a single pulse admixture model, this indicates that the ancestry tract lengths will tend to be longer than those under a single pulse admixture model in which all individuals entered at time t. Indeed, we found that admixture times tended to be underestimated with models of continuous migration. However, the accuracy of LAI remained high across all situations considered here (Table 1), indicating that the LAI aspect of this approach may be robust to alternative demographic models.
In the second set of simulations, we considered additive selection on alleles that are perfectly correlated with local ancestry in a given region (i.e. selected sites with frequencies 0 in population 0 and frequency 1 in population 1), and experience relatively strong selection (selective coefficients were between 0.005 and 0.05). We placed selected sites at 2, 5, 10 and 20 loci distributed randomly across the simulated chromosome, where admixture occurred through a single pulse. Ancestry tracts tend to be longer immediately surrounding selected sites, and we therefore expected admixture time to be underestimated when selection is widespread. When the number of selected loci was small, time estimates were nearly unbiased (Table 2), suggesting that our approach can yield reliable admixture time estimates despite the presence of a small number of selected loci (i.e. 2 selected loci on a chromosome arm). However, with more widespread selection on alleles associated with local ancestry, time estimates showed a downward bias that increased with increasing numbers of selected loci. This is likely because selected loci will tend to be associated with longer ancestry tracts due to hitchhiking. However, the accuracy of the LAI remains high for all selection scenarios that we considered here, further indicating that our method can robustly delineate LA, even when the data violate assumptions of the inference method (Tables 1 and 2).
We next compared the results of our method to those of WinPop [34]. Because WinPop accepts only diploid genotypes, we provided this program diploid genotype data. However, for these comparisons, we still ran our method on simulated read pileups with the mean depth equal to 2. WinPop was originally designed for local ancestry inference in very recently admixed populations. As expected, WinPop performed acceptably for very short admixture times, but rapidly decreased in performance with increasing time (S6 Fig). However, by default, WinPop removes sites in strong LD within the admixed samples, which includes ancestral LD, but also admixture LD—the exact signal LAI methods use to identify ancestry tracts.
We therefore reran WinPop, but instead of pruning LD within the admixed population, we removed sites in strong LD within the ancestral populations as described above in our method. With this modification, WinPop performs nearly as well as our method, but remains slightly less accurate especially at longer admixture times (S6 Fig). This difference presumably reflects the windowed-based approach of WinPop. At longer times since admixture a given genomic window may overlap a breakpoint between ancestry tracts. Although the performance is nearly comparable with this modification, we emphasize that our method enables users to estimate the time since admixture, where this must be supplied for WinPop, and allows for LAI on read pileups, therefore incorporating genotype uncertainty into the LAI procedure. Indeed our method is more accurate at longer timescales even when supplied with considerably lower quality read data. However WinPop supports LAI with multiple ancestral populations, which our method currently does not (but see Conclusions). Furthermore many LAI algorithms utilize haplotype information, which may be particularly valuable in populations where LD extends across large distances as in e.g. human populations.
Given the strong interest in studying admixture and local ancestry in human populations (e.g. [22–25]), it is useful to ask if our method can be applied to data consistent with admixed populations of humans. Towards that goal, we simulated data similar to what would be observed in admixture between modern European and African lineages and applied our HMM to estimate admixture times and LA. We found that our method can accurately estimate admixture times for relatively short times since admixture, however, substantially more stringent LD pruning in the reference panels is necessary to produce unbiased estimates (Fig 4). This may be expected given that linkage disequilibrium extends across longer distances in human populations than it does in D. melanogaster. In other words, the scales of ancestral LD and admixture LD become similar rapidly in admixed human populations. Furthermore, this approach yields accurate time estimates for shorter times since admixture than with genetic data consistent with D. melanogaster populations. For a relatively short time since admixture, around 100 generations, it is possible to obtain accurate and approximately unbiased estimates of the admixture time over a wide range of ancestry proportions, indicating that this method may be applicable to recently admixed human populations as well (Fig 4). Nonetheless, this result underscores the need to examine biases associated with LD pruning in this approach prior to application to a given dataset.
To demonstrate that assumptions about the number of generations since admixture have the potential to bias LAI, we analyzed a SNP-array dataset from Greenlandic Inuits [60,61]. The authors had previously noted a significant impact of t on the LAI results produced using RFMix [24], which we were able to reproduce here for chromosome 10 (S7 Fig). Indeed, even for comparisons between t = 5 and t = 20, both of which may be biologically plausible for these populations, the mean difference in posterior probabilities between samples estimated using RFMix was 0.0903 (S7 Fig). However, when we applied our method to these data, a clear optimum from t was obtained at approximately 6–7 generations prior to the present, which is close to the plausible times of admixture for these populations (S7 Fig). This comparison therefore demonstrates that even relatively minor changes in assumptions of t have the potential to strongly impact LAI results, and underscores the importance of simultaneously performing LAI while estimating t.
However, these results also indicate that our method may not be robust in situations where the background LD is high and ancestry informative markers are neither common nor distributed evenly across the genome. When we compared the results of our method at t = 5 and at t = 20, we also obtained differences in the mean posterior among individuals as with RFMix. However, one notable difference is that the mean posterior difference using RFMix has a particularly high variance and therefore higher mean error (S7 Fig), but actually a lower median difference than we found using our method. There are likely two causes for differences observed in the mean ancestry posterior among individuals. First, the datasets considered were generated with a metabochip SNP-chip [62], which contains a highly non-uniform distribution of markers across the genome. Second, the ancestral LD in the Inuit population is extensive [61], and we could only retain a relatively small proportion of the markers after LD pruning in the reference panels. These results therefore also underscore the challenges of LAI when the signal to noise ratio is low as may be the case in some human populations, for which LD is extensive, and for some sequencing strategies.
Although in general it is straightforward to estimate m from genome-wide data, in some cases this parameter may be misestimated prior to LAI. We therefore sought to quantify this potential effect by performing LAI after supplying incorrect values of m. In general, we found that values close to the true range, i.e. within 0.05 of the true m, tend to yield reasonably accurate time estimates. However, increasingly incorrect values produce sharply downwardly biased time estimates and this effect is especially pronounced for highly skewed true m (S8 Fig). As could be expected given the robustness of LAI to many perturbations (above), when the incorrect t is supplied to the program, the LA results remain reasonable. However it is worth noting that the penalty appears to be greatest when t is too small rather than too large (S9 Fig).
Although this is not a primary focus for this work, for some users it may be of interest to construct confidence intervals for estimates of t. We recommend the block bootstrap as the preferred method for estimating confidence interval for t, and we have written a script that will produce these (available on the github page for this project: https://github.com/russcd/Ancestry_HMM). Simulations confirm that this can produce confidence intervals overlapping the true t (S10 Fig), but bias in t estimates for higher ploidy samples may still be apparent in some cases.
Given their effects suppressing recombination in large genomic regions, chromosomal inversions may be expected to strongly affect LAI [2,63]. Although we attempted to limit the impact of chromosomal inversions by eliminating known polymorphic arrangements from the reference panels (see methods), many known inversions are present within the pool-seq samples we aimed to analyze [64]. We therefore focused on known inverted haplotypes within the DGPR samples [63,65–67], which are comprised of inbred individuals, and therefore phase is known across the entire chromosome.
In comparing LA estimates between inverted and standard arrangements, it is clear that chromosomal inversions can substantially affect LA across the genomes (Fig 5). In general, the chromosomal inversions considered in this work originated in African populations of D. melanogaster [63], and consistent with this observation, most inversion bearing chromosomes showed evidence for elevated African ancestry. This was particularly evident in the regions surrounding breakpoints, where recombination with standard arrangement chromosomes is most strongly suppressed. Importantly, this pattern continued outside of inversion breakpoints as well, consistent with numerous observations that recombination is repressed in heterokaryotypes in regions well outside of the inversion breakpoints in Drosophila (e.g. [2,63,68]). In(3R)Mo is an exception to this general pattern of elevated African ancestry within inverted arrangements (Fig 5). This inversion originated within a cosmopolitan population [63], and has only rarely been observed within sub-Saharan Africa [69,70]. Consistent with these observations, In(3R)Mo displayed lower overall African ancestry than chromosome arm 3R than standard arrangement chromosomes.
Although chromosomal inversions may affect patterns of LA in the genome on this ancestry cline, we believed including chromosomal inversions in the pool-seq datasets would not heavily bias our analysis of LA clines. Inversions tend to be low frequency in most populations studied [64], and because they affect LA in broad swaths of the genome—sometimes entire chromosome arms—including inversions is unlikely to affect LA cline outlier identification which appears to affect much finer scale LA (below). Furthermore, inversion breakpoint regions were not enriched for LA cline outliers in our analysis (S2 Table), suggesting that inversions have a limited impact on overall patterns of local ancestry on this cline. Nonetheless, the LAI complications associated with chromosomal inversions should be considered when testing selective hypotheses for chromosomal inversions as genetic differentiation may be related to LA, rather than arrangement-specific selection in admixed populations such as those found in North America.
Finally, we applied our method to ancestry clines between cosmopolitan and African ancestry D. melanogaster. Genomic variation across two ancestry clines have been studied previously [21,38,40,52]. In particular, the cline on the east coast of North America has been sampled densely by sequencing large pools of individuals to estimate allele frequencies, and previous work has shown that the overall proportion of African ancestry is strongly correlated with latitude [21]. Consistent with this observation, we found a significant negative correlation for all chromosome arms between proportion of average African ancestry and latitude (rho = -0.891, -0.561, -0.912, -0.913, and -0.755, for 2L, 2R, 3L, 3R, and X respectively).
Although global ancestry proportions have previously been investigated in populations on this ancestry cline [21,38], these analyses neglected the potentially much richer information in patterns of LA across the genome. We therefore applied our method to these samples. Because of the relatively recent dual colonization history of these populations and subsequent mixing of genomes, a genome-wide ancestry cline is expected [21]. However, loci that depart significantly in clinality from the genome-wide background levels may indicate that natural selection is operating on a site linked to that locus.
Previously Pool (2015) found that regions of low recombination are disproportionately enriched for African ancestry in the Raleigh, NC population [17]. Here, we find a similar pattern and we further find that is replicated across all populations that were assayed on this ancestry cline. Specifically, in all populations studied the proportion of African ancestry is significantly negatively correlated with local recombination rates (Fig 6). Ultimately, this correlation may have two causes. First, if selection is more efficient at purging African alleles in high recombination regions, these loci will tend to be removed preferentially in those genomic regions. An alternative explanation is that introgressing African alleles that are favored by selection would tend to bring larger linkage blocks along with them in the predominantly low recombination regions. Regardless of the specific source of natural selection, a neutral admixture model would not predict this robust correlation between LA and recombination rates within all populations, indicating that natural selection has played an important role in shaping LA on this ancestry cline.
Previous studies have found that heterogeneity in the genome with respect to ancestry informative markers may impact the accuracy of LAI [71]. To assess this possibility, we computed the mean difference between posterior mean estimates for the two samples from Florida and between the two samples from Maine. Importantly, because these pooled samples were created using different isofemale lines [40], this is a conservative test of our method since there will be true biological differences as well as stochastic sequencing differences between replicates from each population. We found no correlation between the mean difference of the posterior means and local recombination rates (P = 0.2353 and P = 0.7529, Spearman’s rank correlation for Florida and Maine respectively), indicating that the correlation observed between local recombination rates and LA is unlikely to be an artifact of differential accuracy of LAI in different genomic regions. However, it should be acknowledged that in some genomic regions it maybe challenging to unambiguously infer LA [17,71].
Selection within admixed populations may take several distinct forms. On the one hand, loci that are favorable in the admixed population—either because they are favored on an admixed genetic background, enhance reproductive success in an admixed population, or are favorable in the local environment—will tend to achieve higher frequencies, and we would expect these sites to have a more positive correlation with latitude than the genome-wide average. Conversely, loci that are disfavored within the admixed population may be expected to skew towards a more negative correlation with latitude.
Although it is not possible to distinguish between these hypotheses directly, a majority of evidence suggests that selection has primarily acted to remove African ancestry from the largely Cosmopolitan genetic backgrounds found in the Northern portion of this ancestry cline. First, abundant evidence suggests pre-mating isolation barriers between some African and cosmopolitan populations [72–74]. Second, there is strong post-mating isolation between populations on the ends of this cline [46,47]. Third, we report here a strong negative correlation between LA frequency and local recombination rates (above). Finally, circumstantially, the local environment on the east coast of North America is perhaps most similar to the environment of Cosmopolitan compared to African ancestral populations, which further suggests that Cosmopolitan alleles are likely favored through locally adaptive mechanisms. For these reasons, we therefore examined loci that are outliers for a negative partial correlation with latitude, as this is the expected pattern for African alleles that are disfavored in more temperate populations. In other words, the outlier regions show a significantly stronger negative correlation between local African ancestry and latitude than the chromosome arm does on average.
There is an ongoing debate about the relative merits of an outlier approach versus more sophisticated models for detecting and quantifying selection in genome-wide scans. We believe that the difficulties of accurately estimating demographic parameters for this ancestry cline make the outlier approach most feasible for our purposes. Using our outlier approach, we identified 80 loci that showed the expected negative partial correlation with latitude (Fig 7). Although the specific statistical threshold that we employed is admittedly arbitrary, given the strength of evidence indicating widespread selection on local ancestry in this species (above), we expected that the tail of the LA cline distribution would be enriched for the genetic targets of selection.
Due to the differences in inheritance, evolutionary theory predicts that selection will operate differently on the X chromosome relative to autosomal loci. Of specific relevance to this work, the large-X effect [75,76] is the observation that loci on the X chromosome contribute to reproductive isolation at a disproportionately high rate. Additionally, and potentially the cause of the large-X effect, due to the hemizygosity of X-linked loci, the X chromosome is expected to play a larger role in adaptive evolution, the so-called faster-X effect [77]. There is therefore reason to believe that the X chromosome will play a significant role in genetically isolating Cosmopolitan and African D. melanogaster.
Consistent with a larger role for the sex chromosomes in generating reproductive isolation or selective differentiation between D. melanogaster ancestral populations, we found that that the X chromosome has a lower mean African ancestry proportion than the autosomes in all populations. Furthermore, the X displays a stronger correlation between local recombination rates and the frequency of African ancestry than the autosomes in all 14 populations samples, potentially indicating that selection has had a disproportionately strong effect shaping patterns of local ancestry on this chromosome than on the autosomes. In addition, the X has a significantly higher rate of outlier LA clinal loci than the autosomes (23 LA outliers on the X, 57 on the Autosomes, p = 0.0341, one-tailed exact Poisson test). Although consistent with evolutionary theory, differences between autosomal arms and the X chromosome may also be explained in part by differences in effective recombination rates on the X chromosome than the autosomes, differences in power to identify LA clines associated with chromosome arm specific patterns, or by the disproportionately larger number of chromosomal inversions on the autosomes than on the X chromosome in these populations [64,69]. Distinguishing between this hypothesis and confounding factors will be central to determining whether key results from speciation research are replicated in much more recently diverged populations.
We next applied gene ontology analysis to the set of outlier genes to identify common biological attributes that may suggest more specific organismal phenotypes underlying LA clinal outliers. In total, we identified seven GO terms that remained significant after applying a 5% FDR correction (S3 Table). These GO terms reflect the presence of two primary clusters of genes. The first, which corresponds broadly to histone acetylation, may be related to chromatin remodeling and therefore is expected to effect gene expression levels across a large number of loci. Previous work focused on this ancestry cline has identified chromatin remodeling genes as a potentially important component locally adaptive variation on this ancestry cline [78]. This may indicate that this previous efforts to identify spatially varying selection in these populations may have been detecting selection on local ancestry components associated with ecological adaptation in ancestral populations. The second GO cluster, eukaryotic translation initiation factor 2 complex, also appears to implicate a central role of clinal LA outliers on the regulation of gene expression. One plausible explanation of these observations is that gene expression, particularly high level regulation of gene expression, may be especially likely to contribute to epistatic interactions as these proteins will inherently interact with a diverse set of loci throughout the genome. Given that two distinct gene clusters related to gene expression are identified by this analysis, gene expression would appear to be a plausible candidate phenotype to investigate in future work on ecological divergence and isolating factors in admixed D. melanogaster populations. Testing this prediction empirically through expression profiling may therefore offer fruitful grounds of understanding the earliest stages of reproductive isolation.
Another subset of loci that we may wish to identify using these data are those that contribute to reproductive isolation between African and Cosmopolitan D. melanogaster populations and would therefore be removed by selection from most populations on this ancestry cline. Although it is possible that Cosmopolitan alleles would be disfavored in an admixed background as well, because these populations are predominantly Cosmopolitan, we expect that the majority of selection on negatively epistatically interacting loci would remove African alleles from populations. To identify these loci, we first computed the mean African ancestry across all populations, and we then identified the subset of loci that were in the lowest 5% tail. From those loci, we selected the loci minima from adjacent genomic windows (see Methods, Fig 8), and we obtained a total of 84 local ancestry outliers.
As with the clinal outlier analysis above, to identify commonalities in the types of loci identified by this analysis, we performed GO analysis on the set of loci that are outliers for the mean proportion of African ancestry. After a 5% FDR correction, there are again several gene clusters that are significantly enriched in this set of outlier loci (S4 Table). Of particular interest is the GO term oogenesis, which may indicate that female reproduction is affected during admixture between cosmopolitan and African populations of D. melanogaster. This finding is particularly interesting in light of the fact that female fertility is strongly affected when autosomal chromosomes from one end of this ancestry cline are made homozygous on a genetic background carrying the X chromosome from the other end of this ancestry cline [47]. Hence, the effects of combining divergence ancestry types on female fertility, and specifically the genetic basis of oogenesis, may be an appealing phenotype to characterize in detail in attempting to clarify the genetic effects that isolate African and Cosmopolitan D. melanogaster populations.
Given the abundance of evidence supporting a role for pre-mating isolation barriers between African and Cosmopolitan flies [72–74], we are interested in highlighting genes potentially related to behavioral isolation between ancestral populations of D. melanogaster. Consistent with this observation, one of the strongest LA cline outliers, egh, has been conclusively linked to strong effects on male courtship behavior using a variety of genetic techniques [79]. Additionally, gene knockouts of CG43759, another LA cline outlier locus, have strong effects on inter-male aggressive behavior [80], and may also contribute to behavioral differences between admixed individuals. These loci are therefore appealing candidate genes for functional follow-up analyses, and illustrate the power of this LAI approach for identifying candidate genes that are potentially associated with well characterized phenotypic differences between ancestral populations.
The Pennsylvania population included in this study has been sampled extensively, including several paired fall and spring samples across three consecutive years. Previously, Bergland et al. [40] identified numerous SNPs that showed recurrent and rapid seasonal frequency changes in the Pennsylvania populations included in this study. They concluded that these sites are experiencing recurrent selection associated with recurrent environmental seasonal changes. To determine if LA across the D. melanogaster genome might also experience selection associated with seasonal frequency shifts, we searched for loci that showed a strong recurrent seasonal shift in LA. However, we identified fewer significantly seasonal sites than we would expect to by chance (the proportion of significant site at the alpha = 0.05 level of significance is 0.041). Furthermore, after applying a false discovery rate correction [81], there are no sites that are significantly seasonal at the q = 0.1 level. Collectively, these results indicate that LA within the Pennsylvania populations of D. melanogaster remains remarkably stable during seasonal environmental cycles.
Although this observation may, to a first approximation, appear to be at odds with the results reported in Bergland et al. [40], we believe that it is consistent with the model proposed in that work. Specifically, the authors suggested that long term balancing selection may maintain these seasonally favorable polymorphisms in diverse D. melanogaster populations and even in the ancestors of D. melanogaster and D. simulants [40]. We therefore may expect that these polymorphisms will be maintained at similar frequencies in African and Cosmopolitan populations. Hence, although the seasonal SNPs change rapidly in frequency between spring and fall [40], the LA at these sites can remain stable during seasonal fluctuations.
A growing number of next-generation sequencing projects produce low coverage data that cannot be used to unambiguously assign individual genotypes, but which can be analyzed probabilistically to account for uncertainty in individual genotypes [82–84]. However, most existing LAI methods require genotype data derived from diploid individuals. Hence, there is an apparent disconnect between existing LAI approaches and the majority of ongoing sequencing efforts. In this work, we developed the first framework for applying LAI to pileup read data, rather than genotypes, and we have generalized this model to arbitrary sample ploidies. This method therefore has immediate applications to a wide variety of existing and ongoing sequencing projects, and we expect that this approach and extensions thereof will be valuable to a number of researchers. Although evaluating this application is beyond the scope of this work, one particularly enticing potential use of this method is LAI in ancient DNA samples for which sequencing depths often preclude accurate genotype calling. Importantly, it would be straightforward to model site-specific errors in this framework, which could be particularly important for ancient DNA applications [6].
For many applications, a parameter of central biological interest is the time since admixture began (t). A wide variety of approaches have been developed that aim to estimate t and related parameters in admixed populations [26,28–31,85,86]. Many of these methods are based on an inferred distribution of tract lengths, however, inference of the ancestry tract length distribution is associated with uncertainty that is typically not incorporated in currently available methods for estimating t. Furthermore, incorrect assumptions regarding t have the potential to introduce biases during LAI. Hence, it is preferable to estimate demographic parameters such as the admixture time during the LAI procedure. Nonetheless, as noted above, although LAI using our method is robust to many deviations from the assumed model, admixture time estimates are sensitive to a variety of potential confounding factors and examining the resulting ancestry tract distributions after LAI may be necessary to confirm that the assumed demographic model provides a reasonable fit to the data.
To our knowledge, this is the first method that attempts to simultaneously link LAI and population genetic parameter estimation directly, and we can envision many extensions of this approach that could expand the utility of this method to a broad variety of applications. For example, it is straightforward to accommodate additional reference populations (e.g. by assuming multinomial rather than binomial read sampling). Alternatively, any demographic or selective model that can be approximated as a Markov process could be incorporated—in particular, it is feasible to accommodate two-pulse admixture models and possibly models including ancestry tracts that are linked to positively selected sites. Such methods can be used to construct likelihood ratio tests of evolutionary models and for providing improved parameter estimates.
We model the ancestry using an HMM {Hv} with state space S = {0,1,…,n}, where Hv = i, i ∈ S, indicates that in the vth position i chromosomes are from population 0 and n–i chromosomes are from population 1. In the following, to simplify the notation and without loss of generality, we will omit the indicator for the position in the genome as the structure of the model is the same for all positions of equivalent ploidy. We assume each variant site is biallelic, with two alleles A and a, and the availability of reference panels from source populations 0 and 1 with total allelic counts C0a, C1a, C0A, and C1A, where the two subscripts refer to population identity and allele, respectively. Also, C0 = C0A + C0a and C1 = C1A + C1a. Finally, we also assume we observe a pileup of r reads from the focal population, with rA and ra reads for alleles A and a respectively (r = rA + ra). The emission probability of state i ∈ S of the process is then defined as ei = Pr(rA, C0A, C1A | r, C0, C1, H = i, ε), where ε is an error rate. This probability can be calculated by summing over all possible genotypes in the admixed sample and over all possible population identities of the reads, as explained in the following section.
The probability of obtaining r0 (= r–r1) reads, in the admixed population, from chromosomes of ancestry 0, given r and the hidden state H = i, and assuming no mapping or sequencing biases, is binomial,
r0|H=i,n,r∼Bin(r,i/n)
(1)
These probabilities are pre-computed in our implementation for all possible values of i ∈ S and r0, 0 ≤ r0 ≤ r. Similarly, for the reference populations, for j = 0,1,
CjA|Cj,fj∼Bin(Cj,fj)
(2)
where fj is the allele frequency of allele A in population j. The analogous allelic counts in the admixed population, denoted CM0a, CM1a, CM0A, and CM1A, are unobserved (only reads are observed for the admixed population), but are also conditionally binomially distributed, i.e.:
CM0A|H=i,f0∼Bin(i,f0)andCM1A|H=i,n,f1∼Bin(n−i,f1)
(3)
Finally, in the absence of errors, and assuming no sequencing or mapping biases, the conditional probability of obtaining r0A reads of allele A in the admixed population is
r0A|H=i,r0,CM0A∼Bin(r0,CM0A/i)
(4)
It should be noted that because we are explicitly modeling the process of sampling alleles from the population (Eq 3) and the process of sampling reads conditional on the sample allele frequencies (Eq 4), that this approach corrects for the increased variance associated with two rounds of binomial sampling in poolseq applications that has been reported previously (e.g., in [52]).
This probability can be expanded to include errors, in particular assuming a constant and symmetric error rate ε between major and minor allele, and assuming all reads with nucleotides that are not defined as major or minor are discarded, we have
r0A|H=i,ro,CM0A,ε∼Bin(r0,(1−ε)CM0A/i+ε(1−CM0A/i)).
(5)
Using these expressions, and integrating over allele frequencies in the source populations, we have
Pr(r0A,C0A,|r0,C0,n,H=i,ε)=∫01∑k=0iPr(r0A|H=i,r0,CM0A=k,ε)Pr(CM0A=k|H=i,f0)p(f0)df0=C0!i!(C0−C0A)!C0A!(C0+i+1)!∑k=0iPr(r0A|H=i,r0,CM0A=k,ε)(C0−C0A+i−k)!(C0A+k)!(i−k)!k!
(6)
assuming a uniform [0, 1] distribution for f0. A similar expression is obtained for Pr(r1A,C1A,|r1,C1,n,H = i,ε), assuming f1 ∼ U[0,1], and these expressions combine multiplicatively to give
Pr(rA,C1A,,C0A,|r0,C0,r1,C1,n,H=i,ε)=∑r0A=max{0,rA−r1}min{r0,rA}Pr(r0A,C0A,|r0,C0,n,H=i,ε)Pr(r1A=rA−r0A,C1A,|r1,C1,n,H=i,ε),
(7)
and the emission probabilities become
Pr(rA,C0A,C1A|r,C0,C1,H=i,ε)=∑r0=0rPr(r0|H=i,n,r)Pr(rA,C1A,,C0A,|r0,C0,r1=r−r0,C1,n,H=i,ε)
(8)
Alternatively, if the sample genotypes are known with high confidence, i.e. CMA = CM0A + CM1A is observed, the emission probabilities are the defined as
Pr(CMA,C0A,C1A|C0,C1,n,H=i)=(C0C0A)(C1C1A)∑k=max{CMA−i,0}min{n−i,CMA}∫01(n−ik)(f0)C0A+k(1−f0)C0+n−i−C0A−kdf0∫01(iCMA−k)(f1)CMA−k+C1A(1−f1)C1+i−C1A−CMA+kdf1=∑k=max{CMA−i,0}min{n−i,CMA}C0!C1!i!(n−i)!(CMA+C1A−k)!(C0A+k)!(C1−CMA−C1A+i+k)!(C0−C0A−i−k+n)!(C0−C0A)!C0A!(C1−C1A)!C1A!(CMA−k)!k!(k+i−CMA)!(n−k−i)!(n−i+C0+1)!(i+C1+1)!
(9)
These emissions probabilities are sometimes substantially faster to compute than those for short read pileups, especially when sequencing depths are high. However, the genotypes must be estimated with high accuracy for this approach to be valid. For applications with low read coverage, or with ploidy >2 for which many standard genotype callers are not applicable, it is usually preferable to use the pileup-based approach described above.
We assume an admixed population, of constant size, with N diploid individuals, in which a proportion m of the individuals in the population where replaced with migrants t generations before the time of sampling. Given these assumptions, and an SMC’ model of the ancestral recombination graph [87], the rate of transition from ancestry 0 to 1, along the length of a single chromosome, is
λ0=2Nm(1−e−t2N)
(10)
per Morgan [57]. Similarly, the rate of transition from ancestry 1 to 0 on a single chromosome is
λ1=2N(1−m)(1−e−t2N)
(11)
per Morgan. Importantly, because these expressions are based on a coalescence model, they account for the possibility that a recombination event occurs between two tracts of the same ancestry type and the probability that the novel marginal genealogy will back-coalesce with the previous genealogy [57]. Both events are expected to decrease the number of ancestry switches along a chromosome and ignoring their contribution will cause overestimation of the rate of change between ancestry types between adjacent markers.
The transition rates are in units per Morgan, but can be converted to rates per bp, by multiplying with the recombination rate in Morgans/bp, rbp within a segment. The transition probabilities of the HMM for a single chromosome, P(l) = {Pij(l)},i,j ∈ S, between two markers with a distance l between each other, is then approximately
P(l)=[1−λ0rbpλ0rbpλ1rbp1−λ1rbp]l
(12)
using discrete distances, or
P(l)=[λ1λ0+λ1+λ0λ0+λ1e−rbpl(λ0+λ1)λ0λ0+λ1−λ0λ0+λ1e−rbpl(λ0+λ1)λ0λ0+λ1+λ1λ0+λ1e−rbpl(λ0+λ1)λ1λ0+λ1−λ1λ0+λ1e−rbpl(λ0+λ1)]
(13)
using continuous distances along the chromosome. Here, we use the continuous representation for calculations. We emphasize that the assumption of a Markovian process is known to be incorrect [57], in fact admixture tracts tend to be more spatially correlated than predicted by a Markov model, and the degree and structure of the correlation depends on the demographic model [57]. Deviations from a Markovian process may cause biases in the estimation of parameters such as t.
The Markov process defined above is applicable to a single chromosome. We now want to approximate a similar process for a sample of n chromosomes from a single sequencing pool. The true process is quite complicated, and we choose for simplicity to approximate the process for n chromosomes sampled from one population, as the union of n independent chromosomal processes. We will later quantify biases arising due to this independence assumption using simulations. Under the independence assumption, the transition probability from i to j is simply the probability of l transitions from state 1 to state 0 in the marginal processes and j–i + l transitions from state 0 to state 1, summed over all admissible values of l, i.e.,
Pr(Hv+k=j|Hv=i)=∑l=max{0,i−j}min{n−j,i}(n−ij−i+l)(P01(k))j−i+l(1−P01(k))n−j+i−l(il)(P10(k))l(1−P10(k))i−l
(14)
Although this procedure can be computationally expensive when there are many markers, read depths are high, and especially when n is large, in our implementation, we reduce the compute time by pre-calculating and storing all binomial coefficients.
A parameter of central biological interest, that is often unknown in practice, is the time since the initial admixture event (t). We therefore use the HMM representation to provide maximum likelihood estimates of t using the forward algorithm to calculate the likelihood function. As this is a single parameter optimization problem for a likelihood function with a single mode, optimization can be performed using a simple golden section search [88]. Default settings for this optimization in our software, including the search range maxima defaults, tmax and tmin, are documented in the C++ HMM source code provided at https://github.com/russcd/Ancestry_HMM.
After either estimating or providing a fixed value of the time since admixture to the HMM, we obtained the posterior distribution for all variable sites considered in our analysis using the forward-backward algorithm, and we report the full posterior distribution for each marker along the chromosome.
To validate our HMM, we generated sequence data for each of two ancestral populations using the coalescent simulator MACS [55]. We sought to generate data that could be consistent with that observed in Cosmopolitan and African populations of D. melanogaster, which has been studied previously in a wide variety of contexts [2,11,35–37]. We used the command line “macs 400 10000000 -i 1 -h 1000 -t 0.0376 -r 0.171 -c 5 86.5 -I 2 200 200 0 -en 0 2 0.183 -en 0.0037281 2 0.000377 -en 0.00381 2 1 -ej 0.00382 2 1 -eN 0.0145 0.2” to generate genotype data. This will produce 200 samples of ancestry 0 and 200 samples of ancestry 1 on a 10mb chromosome—i.e. this should resemble genotype data for about half of an autosomal chromosome arm in D. melanogaster. Unless otherwise stated below, we then sampled the first 50 chromosomes from each ancestral population as the ancestral population reference panel, whose genotypes are assumed to be known with low error rates. The sample size was chosen because it is close to the size of the reference panel that we obtained in our application of this approach to D. melanogaster (below).
To evaluate the performance of our method on data consistent with human populations, we simulated data that could be consistent with that observed for modern European and African human populations. Specifically, we simulated the model of [89] using the command line “macs 200 1e8 -I 3 100 100 0 -n 1 1.682020 -n 2 3.736830 -n 3 7.292050 -eg 0 2 116.010723 -eg 1e-12 3 160.246047 -ma x 0.881098 0.561966 0.881098 x 2.797460 0.561966 2.797460 x -ej 0.028985 3 2 -en 0.028986 2 0.287184 -ema 0.028987 3 x 7.293140 x 7.293140 x x x x x -ej 0.197963 2 1 -en 0.303501 1 1 -t 0.00069372 -r 0.00069372”. Admixture between ancestral populations was then simulated as described below.
Although it is commonly assumed that admixture tract lengths can be modeled as independent and identically distributed exponential random variables (e.g. [26,29] and in this work, above), this assumption is known to be incorrect as ancestry tracts are neither exponentially distributed, independent across individuals, nor identically distributed along chromosomes [57]. We therefore aim to determine what bias violations of this assumption will have on inferences obtained from this model. Towards this, we used SELAM [56] to simulate admixed populations under the biological model described above. Because this program simulated admixture in forward time, it generates the full pedigree-based ancestral recombination graph, and is therefore a conservative test of our approach relative to the coalescent which is known to produce incorrect ancestry tract distributions for short times [57]. Briefly, we initialized each admixed population simulation with a proportion, m, of ancestry from ancestral population 1, and a proportion 1-m ancestry from ancestral population 0. Unless otherwise stated, all simulations were conducted with neutral admixture and a hermaphroditic diploid population of size 10,000.
We then assigned the additional, non-reference chromosomes from the coalescent simulations, to each ancestry tract produced in SELAM simulations according to their local ancestry along the chromosome. In this way, each chromosome is a mosaic of the two ancestral subpopulations. See, e.g. [2], for a related approach for simulating genotype data of admixed chromosomes.
Correlations induced by LD between markers within ancestral populations violates a central assumption of the Markov model framework. Although it may be feasible to explicitly model linkage within ancestral populations (e.g., [24,25]), when ancestral populations have relatively little LD, such as those of D. melanogaster, another effective approach is to discard sites that are in strong LD in the ancestral populations. Hence, to avoid this potential confounding aspect of the data, we first computed LD between all pairs of markers within each reference panel that are within 0.01 centimorgans of one another. We then discarded one of each pair of sites where |r| in either reference panel exceeded a particular threshold, and we decreased this threshold until we obtained an approximately unbiased estimate of the time since admixture estimates of the HMM. This approach differs from a previous method, WinPop [34], where LD is pruned from within admixed samples (see also below).
We first identified all sites where the allele frequencies of the ancestral populations differ by at least 20% within the reference panels. We excluded weakly differentiated sites to decrease runtime and because these markers carry relatively little information about the LA at a given site. Then, to generate data similar to what would be produced using Illumina sequencing platforms, we simulated allele counts for each sample, by first drawing the depth at a given site from a Poisson distribution. In most cases and unless otherwise stated, the mean of this distribution is set to be equal to the sample ploidy. We did this to ensure equivalent sequencing depth per chromosome regardless of pooling strategy, and because this depth is sufficiently low that high quality genotypes cannot be determined. We then generated set of simulated aligned bases via binomial sampling from the sample allele frequency and included a uniform error rate of 1% for both alleles at each site.
Unless otherwise stated, we simulated a total of 40 admixed chromosomes. The HMM can perform LAI on more than one sample at a time, and we therefore included all samples when running it. Hence, we used 40 haploid, 20 diploid, 4 pools of 10 chromosomes, and 2 pools of 20 chromosomes for most comparisons of accuracy reported below, unless otherwise stated. It is worth noting that it is possible to jointly analyze distinct samples from the same subpopulation that have been sequenced at different ploidies.
To investigate the effects of allele frequency differences between reference populations and admixed populations, we performed coalescent simulations using the software MACS [55], using the command line “./macs 500 10000000 -i 1 -h 1000 -t 0.0376 -r 0.171 -c 5 86.5 -I 8 100 100 50 50 50 50 50 50 0 -en 0 2 0.183 -en 0.0037281 2 0.000377 -en 0.00381 2 1 -ej 0.00382 2 1 -eN 0.0145 0.2 -ej 0.0005 3 2 -ej 0.000500001 4 1 -ej 0.001 5 2 -ej 0.001000001 6 1 -ej 0.002 7 2 -ej 0.002000001 8 1”. This might be expected to produce populations that are differentiated similarly to how populations of D. melanogaster would be across European populations or between populations in Central Africa. We then substituted the increasingly divergent populations for the reference panel. All allele frequency differences and LD pruning were performed as described above on each of the substitute reference panels.
To evaluate the performance of the HMM, we computed four statistics. First, we compute the proportion of sites where the true state is within the 95% posterior credible interval, where ideally, this proportion would be equal to or greater than 0.95. As this HMM has discrete states, there are many ways the 95% credible interval could be defined. In light of the fact that the credible interval tends to be narrow (Results), we defined the interval to include all states that are overlapped, by any amount, in the 95% confidence interval of the posterior distribution. Second, we compute the mean posterior error, the average distance between the posterior distribution of hidden states and the true state
E=∑v=0S∑i=0np(Hv=i|r)|i−Iv|Sn
Here S is the total number of sites, Iv is the true state at site v, and r is all the combined read data. Third, we also report the proportion of sites where the maximum likelihood estimate of the hidden state is equal to the true ancestry state. Finally, as an indicator of the specificity of our approach, we also report the average width of the 95% credible interval.
A potential issue with this framework is that the assumptions underlying the transition matrixes and related time of admixture estimation procedure is likely to be violated in a number of biologically relevant circumstances. We therefore simulated populations wherein individuals of ancestral population 1 began entering a population entirely composed of individuals from ancestral population 0, at a time t generations before the present, at a constant rate that is sustained across all subsequent generations until the time of sampling. That is, additional unadmixed individuals of ancestry 1 migrate each generation from t until the present.
Natural selection acting on admixed genetic regions has been inferred in a wide variety of systems (e.g. [5,7,13,17,18]), and is expected to have pronounced effects on the distribution of LA among individuals within admixed populations. Here again, this aspect of biologically realistic populations will tend to violate central underlying assumptions of the model assumed in this work. Towards this, we simulated admixed populations with a single pulse of admixture t generations prior to the time of sampling. We then incorporated selection at 2,5,10, and 20 loci at locations uniformly distributed along the length of the chromosome arm. All selected loci were assumed to be fixed within each ancestral population. Selection was additive and selective coefficients were assigned based on a uniform [0.005, 0.05] distribution to either ancestry 0 or 1 alleles with equal probability. As above, these simulations were conducted using SELAM [56].
For both selected and continuous migration simulations, we then performed the genotype and read data simulation procedure, and reran our HMM as described above. We performed 10 simulations for each treatment.
We next sought to compare our method to a commonly used local ancestry inference method, WinPop [34]. Towards this, we again simulated data using MACS and SELAM as described above. For these comparisons, the initial ancestry contribution was 0.5 and the number of generations since admixture varied between 5 and 1000. For comparison, we supplied WinPop and our program the correct time since admixture and ancestry proportions, as these are required parameters for WinPop. We also supplied the program with genotypes rather than read counts, another requirement of WinPop, whereas we supplied our HMM with read data simulated as described above. We then ran WinPop under default parameters, and we also reran WinPop using LD pruning within the reference panels, as we do in our method, instead of the default LD pruning implemented in WinPop.
To demonstrate that LAI methods can be biased by the arbitrary selection of the time since admixture, we analyzed a dataset of SNP-array genotype data from Greenlandic Inuits. These data are described in detail elsewhere [60,61]. This population has received some admixture from a European source population, and the authors had previously used RFMix [24] to perform LAI, and found some sensitivity to the assumed time since admixture (J. Crawford pers. Comm.). We analyzed data from chromosome 10 using RFMix v1.5.4 [24] as described in Moltke et al. [61] assuming admixture occurred either 5 or 20 generations ago. We subsequently analyzed chromosome 10 using our HMM including the genotype-analysis emissions probabilities and assuming a genotype error rate of 0.2%. For our analysis we identified the LD cutoff that is appropriate for these data as described above.
To generate reference panels, we used a subset of the high quality D. melenaogaster assemblies that have been described previously in Pool et al. (2012) and Lack et al. (2015). As in the local ancestry analysis of Pool (2015), we used the French population. For our African reference panel, we selected a subset of the Eastern and Western African populations (CO, RG, RC, NG, UG, GA, GU) and grouped them into a single population for the purposes of our analysis. We elected to combine populations so that we would have a larger reference panel of African populations for this analysis, this solution may be justified because these D. melanogaster populations are only weakly genetically differentiated [2,21,90], particularly after common inversion-bearing chromosomes are removed from analyses. Specific individuals were selected for inclusion in the African reference panel if previous work found they have relatively little cosmopolitan ancestry (i.e., below 0.2 genome-wide in [2]).
Because of their powerful effects on recombination, chromosomal inversions are known to have substantial impacts on the distribution of genetic variants on chromosomes containing chromosomal inversions in D. melanogaster [2,63]. For this reason, we removed all common inversion-bearing chromosome arms from the reference populations [91]. Nonetheless, it is clear that chromosomal inversions are present in the pool-seq samples [64]. Although the inversions certainly violate key assumptions of our model—particularly the transmission probabilities—given that our approach is robust to a many perturbations, we expect the LA within inverted haplotypes can be estimated with reasonable confidence, and the overall LAI procedure will still perform adequately with low frequencies of chromosomal-inversion bearing chromosomes present within these samples.
Although these reference populations are believed to have relatively little admixture, some admixture is likely to remain within these samples [2]. To mitigate this potential issue, we first applied our HMM to each reference population using the genotype-based emissions probabilities (above). Calculated across all individuals, we found that our maximum likelihood ancestry estimates were identical with those of Pool et al. (2012) at 96.2% of markers considered in our analysis. The differences between the results of these methods may reflect differences in the methodology of LAI or differences in the reference panels. Nonetheless, the broad concordance suggests the two methods are yielding similar overall results. We masked all sites where the posterior probability of non-native ancestry was greater than 0.5 within each reference individual’s genome. These masked sequences were then used as the reference panel for the analyses of pool-seq data below.
We acquired pooled sequencing data from six populations from the east coast of the United States. The generation of these samples, sequencing data, and accession numbers are described in detail in [21,40]. Briefly, the samples are comprised of individuals drawn from natural populations and sequenced in relatively large pools of 66–232 chromosomes. We aligned all data using BWA v0.7.9a-r786 [92] using the ‘MEM’ function and the default program parameters. For all alignments, we used version 5 of the D. melanogaster reference genome [93] in order to make our analysis and coordinates compatible with the Drosophila genome nexus [91]. We then realigned all reads using the indelrealigner tool within the GATK package [84], and we extracted the sequence pileup using samtools mpileup v1.1 [94] using the program’s default parameters.
We extracted sites at ancestry informative positions within the reference panels, where we required that the reference panel have a minimum of 50% of individuals with a high quality genotype call in both Cosmopolitan and African reference populations. As above, ancestry informative sites were defined as those with a minimum of 20% difference in allele frequencies between the reference panels used, and we retained only ancestry informative sites for our analyses. We then produced global ancestry estimates for each chromosome arm separately for each sample using the method of Bergland et al. (2016). We ran our HMM for each chromosome arm and each population, and we provided the program this estimate of the ancestry proportion and the time since admixture, 1593 generations [17]. We elected to provide the time since admixture because we have found that this parameter is difficult to estimate in relatively large pools (see Results). However, the program can accurately estimate LA in high ploidy samples even when the time since admixture cannot be estimated correctly (see Results).
To assess the correlation between local recombination rates and LA in the genome, we computed Spearman’s rank sum correlation between the proportions African ancestry and the local recombination rates in windows of 100 ancestry informative markers. As above, we used the recombination rate estimates of [59]. We estimated confidence intervals using 1000 block-bootstrap samples using window sizes of 100 SNPs.
To determine if there are systematic biases in LAI across the genome, we computed the mean difference in genomic windows between LA estimates for two samples form Maine and between two samples from Florida. We assessed evidence for systematic biases through the correlation between local recombination rates and differences in local ancestry inference using Spearman’s rank sum correlation.
To detect loci that show evidence for steeper ancestry clines than the genomic average, we first computed the Spearman’s rank correlation between mean ancestry proportions and latitude for each chromosome arm separately. Then, for each site for which we obtained a posterior ancestry distribution for all samples, we computed the partial Spearman’s rank correlation between the posterior ancestry mean and latitude while correcting for the correlation between latitude and the overall ancestry proportion. We then computed the probability of obtaining the observed partial correlation in R, which implements the approach of [95], and we retained those sites where the probability of the partial correlation between local ancestry and latitude was less than 0.005 as significant in our analysis. Although this cutoff is arbitrary, given the strong evidence for local adaptation and reproductive isolation in these populations [46,47,96], the tail of the LA cline distribution will likely be enriched for sites experiencing selection on this ancestry gradient. Due to linkage, adjacent sites show strong autocorrelation. We therefore selected the local optima for a given clinally significant LA segment (i.e. a tract where all positions are significantly correlated with latitude at our threshold) and retained these for analyses of outlier loci. Finally, to further reduce the effect of autocorrelation, we retained only those local optima for which no other optimum had a stronger correlation with latitude within 100,000bp on either side on the site.
To identify loci with a disproportionately low proportion of African ancestry across this ancestry cline, we computed the mean African ancestry across all populations. We then selected those sites in the lowest 5% tail on each chromosome arm and selected only the local minima within 100kb windows on either side of a selected locus.
We performed Gene-ontology (GO) analyses on outlier SNPs using Gowinda [97], where the background set of SNPs was all positions at which we obtained a posterior distribution in all samples (i.e. the set on which we obtained estimates of the posterior probability of African ancestry). We ran the program using default parameters, except that we included all genes within 10000bp of a focal SNP, and we performed 1e6 total GO simulations.
To identify recurrent seasonal changes in the local ancestry, we followed an approach similar to [40]. Specifically, we fit a generalized linear model of the form
MeanPosteriorAncestry∼Season+ε
We then recorded the estimated effect size, and probability of the observed correlation for each site in the genome at which we obtained a posterior ancestry distribution in all samples considered. To correct for multiple testing, we applied a false discovery rate correction [81] to the resulting p-value distribution.
|
10.1371/journal.pntd.0003862 | Snake Cathelicidin NA-CATH and Smaller Helical Antimicrobial Peptides Are Effective against Burkholderia thailandensis | Burkholderia thailandensis is a Gram-negative soil bacterium used as a model organism for B. pseudomallei, the causative agent of melioidosis and an organism classified category B priority pathogen and a Tier 1 select agent for its potential use as a biological weapon. Burkholderia species are reportedly “highly resistant” to antimicrobial agents, including cyclic peptide antibiotics, due to multiple resistance systems, a hypothesis we decided to test using antimicrobial (host defense) peptides. In this study, a number of cationic antimicrobial peptides (CAMPs) were tested in vitro against B. thailandensis for both antimicrobial activity and inhibition of biofilm formation. Here, we report that the Chinese cobra (Naja atra) cathelicidin NA-CATH was significantly antimicrobial against B. thailandensis. Additional cathelicidins, including the human cathelicidin LL-37, a sheep cathelicidin SMAP-29, and some smaller ATRA peptide derivatives of NA-CATH were also effective. The D-enantiomer of one small peptide (ATRA-1A) was found to be antimicrobial as well, with EC50 in the range of the L-enantiomer. Our results also demonstrate that human alpha-defensins (HNP-1 & -2) and a short beta-defensin-derived peptide (Peptide 4 of hBD-3) were not bactericidal against B. thailandensis. We also found that the cathelicidin peptides, including LL-37, NA-CATH, and SMAP-29, possessed significant ability to prevent biofilm formation of B. thailandensis. Additionally, we show that LL-37 and its D-enantiomer D-LL-37 can disperse pre-formed biofilms. These results demonstrate that although B. thailandensis is highly resistant to many antibiotics, cyclic peptide antibiotics such as polymyxin B, and defensing peptides, some antimicrobial peptides including the elapid snake cathelicidin NA-CATH exert significant antimicrobial and antibiofilm activity towards B. thailandensis.
| Burkholderia species such as B. pseudomallei, which causes melioidosis, and the model organism B. thailandensis are extremely resistant to antibiotics, including cyclic peptide antibiotics such as polymyxin B. Treatment for Burkholderia infections is impeded by this resistance, and new approaches are needed. We hypothesized that the cathelicidin NA-CATH from the Chinese cobra, Naja atra, and smaller derivative peptides (ATRA peptides) may have antimicrobial activity against Burkholderia. We therefore tested the bactericidal effects of the cathelicidin and its derivative peptides. We also wanted to determine whether the antimicrobial peptides exert anti-biofilm activity, although the role of biofilm as a critical virulence factor of Burkholderia has not yet been established. We found that the peptide ATRA-1A, as well as the stereo-isomer D-ATRA-1A, were able to kill B. thailandensis, and the full-length snake cathelicidin NA-CATH was able to both kill B. thailandensis and inhibit its biofilm formation, unlike the human-alpha defensin peptides HNP-1 and HNP-2, and the small peptide derived from hBD3. These results show that the NA-CATH antimicrobial peptide possess bactericidal and anti-biofilm activity against B. thailandensis, and suggest that these compounds should be tested for their effect against the more virulent strains of Burkholderia.
| Burkholderia pseudomallei is a Gram-negative soil bacterium which acts as a facultative intracellular pathogen that can infect both humans and animals, causing melioidosis. Melioidosis is endemic to Southeast Asia and Northern Australia, where the mortality rates are 50% and 19% respectively [1–3]. In addition, B. pseudomallei is of interest because it is considered a class B priority pathogen and a Tier 1 Select Agent and has potential for aerosol delivery. In this study, Burkholderia thailandensis is used as a model for B. pseudomallei [4]. B. thailandensis is a BSL-2 organism closely related to B. pseudomallei with an LD50 in mice 1000-fold higher than that of B. pseudomallei, making it an easier and safer model organism with which to work [5]. B. thailandensis has been successfully demonstrated to be a useful BSL-2 surrogate for B. pseudomallei [4,6–8] for both in vitro and in vivo experiments. Thus, B. thailandensis may be a good model in which to study the molecular actions of full length cathelicidins such as LL37 both as antibacterial and antibiofilm peptides against Burkholderia strains.
In B. pseudomallei and B. thailandensis the significant resistance towards several categories of antibiotics, including chloramphenicol, quinolones, tetracyclines, and trimethoprim, is mediated by the overexpression of efflux pumps [9,10]. B. pseudomallei and B. thailandensis are typically grown in the laboratory in the presence of >100 mg/ml polymyxin B [11]; such ready growth indicates their high level of resistance to cyclic peptide antibiotics. In fact, the genus Burkholderia is said to have “extreme antimicrobial peptide and polymyxin B resistance” [12]. Therefore, the discovery of novel therapeutic alternatives is urgently required.
We have previously studied the cathelicidin peptide from the elapid snake Naja atra and designed smaller peptide derivatives called ATRA peptides; we reported that these peptides were highly active against both Gram-positive and Gram-negative bacteria, such as Gram-positive Staphylococcus aureus and Gram-negative Pseudomonas aeruginosa [13–16]. We were very interested to know whether B. thailandensis would be susceptible to other antimicrobial peptides, and particularly to the very effective cathelicidin peptide (NA-CATH) and smaller peptide derivatives from elapid snakes that we had been studying.
Cationic antimicrobial peptides (CAMPs) are produced as part of the innate immune system by higher-order organisms. These peptides are also referred to as host-defense peptides (HDPs). CAMPs are low-molecular-weight, cationic, and often amphipathic peptides, and their overall positive charge enables association with the negatively charged bacterial outer membrane [17]. In this study, we tested two types of CAMPs: the cathelicidin type and the defensin type.
It has been previously reported that Burkholderia species, specifically B. cepacia, are very resistant to beta-defensins, a category of defensin CAMPs [18]. Defensins function by replacing Ca2+ and Mg2+ ions in the bacterial membrane, disrupting membrane stability and leading to loss of electric potential and eventual cell lysis [19,20]. Defensins are important to consider because they are found in human skin under inflammatory conditions [21] and could potentially play a role during a wound infection by B. pseudomallei. In this study, we tested the antibacterial activity of two alpha-defensins and a small peptide from beta-defensin against B. thailandensis.
Cathelicidins are a class of antimicrobial peptide characterized by a highly conserved cathelin domain [22] and a sequence-variable active cathelicidin domain. The majority of cathelicidin peptides form amphipathic alpha helices when in contact with a membrane, and these helices are believed to play a crucial role in their function [23,24]. Recently a cathelicidin, designated NA-CATH, has been discovered in Naja atra, the Chinese cobra [25], an elapid snake found in Southeast Asia [26]. This cathelicidin contains an imperfect repeated 11-amino-acid motif named the ATRA motif (Table 1) [13]. The first repeat is called ATRA-1 and the second repeat ATRA-2 [13]. A derivative, ATRA-1A, was created by replacing the 3rd residue of ATRA-1 with an alanine [13]. In previous work we demonstrated that the full-length cathelicidin (NA-CATH) and peptides based on the first repeat (ATRA-1 and ATRA-1A) were effective broad-spectrum antimicrobial agents against Francisella novicida, Aggregatibacter actinomycetemcomitans, Pseudomonas aeruginosa, and Staphylococcus aureus [13–16]. Therefore, the Naja atra cathelicidin NA-CATH and its ATRA derivatives were chosen for studies against Burkholderia species [10]. It is of note that we previously demonstrated that the cathelicidins LL-37, D-LL-37, NA-CATH, and NA-CATH derivatives cause no hemolysis at the antimicrobial concentrations used in our study [13,14]. This leads us to suggest that these peptides may be very useful as a potential new therapeutic approach, perhaps in a topical application, by virtue of their demonstrated antimicrobial action and minimal host-cell cytotoxicity, with the D-peptides having the added advantage of less susceptibility to protease digestion.
B. thailandensis (E264) was obtained from the American Type Culture Collection (Manassas, VA), ATCC 700388, and grown in nutrient broth overnight in a shaking incubator at 37°C. Cultures of B. thailandensis were grown up and the stocks were aliquotted, frozen in 20% glycerol, and stored at -80°C. Cultures were enumerated by serial dilution on nutrient agar.
The antimicrobial activity of various antimicrobial peptides against B. thailandensis was determined as previously described [16]. Briefly, in a sterile 96-well plate, 1x105 CFU per well of bacteria were incubated with serial dilutions of antibiotic (control) and peptide in 10 mM phosphate buffer (3 h, 37°C). Bacterial survival was then determined by serial dilution at each peptide concentration in sterile PBS. Dilutions were plated in triplicate on nutrient agar and incubated at 37°C for 24 h; colonies were then counted to determine survival. Bacterial survival was calculated by the ratio of the number of colonies on each experimental plate to the average number of colonies in the control plates lacking any antimicrobial peptide.
The antimicrobial peptide concentration required to kill 50% of B. thailandensis (EC50) was determined by graphing percent survival versus log of peptide concentration (log μg/ml). Data were plotted using GraphPad Prism 5 (GraphPad Software Inc., San Diego, CA, USA). Survival was determined as the ratio of colonies from experimental plates relative to the average number of colonies from plates lacking peptide. EC50 was determined by fitting the data to a standard sigmoidal dose-response curve. Each experiment was performed three times with three replicates per experiment for n = 9. Error is reported as 95% confidence intervals (CI) for each antimicrobial peptide.
Biofilm was grown and measured as previously described [27–29]. Modified Vogel and Bonner’s medium (MVBM) [30] was inoculated from an overnight culture of B. thailandensis and allowed to incubate for 18 h in a shaking incubator at 37°C. The optical density at 540 nm (OD540) was adjusted to 0.8 OD540. Bacterial suspension (100 μl) was added to wells of a sterile tissue-culture-treated 96-well plate along with various concentrations of peptide and fresh MVBM (final volume 200 μl). Wells containing only medium or no peptide served as negative and positive controls, respectively. Plates were then incubated aerobically at 37°C for 3 h. Following aerobic adhesion, supernatant fluid was removed from wells (to remove planktonic bacteria), fresh MVBM/peptide was added to each well (200 μL final volume), and plates were incubated for 21 h at 37°C. After incubation, supernatant was removed and replaced with 200 μL of fresh MVBM/peptide, then incubated at 37°C for an additional 24 h. This is described as a 48 h biofilm. After final incubation, the plate was read at OD600nm to measure bacterial growth, then washed, fixed, and stained with crystal violet as previously described [31]. Each assay was performed in triplicate and the experiment repeated three times for n = 9.
Biofilm dispersion assay was performed using B. thailandensis E264 (ATCC 700388) in 100 μL MVBM and was incubated 24h, 37°C. After allowing biofilm to form for 24h, the biofilm was treated with 10 μg peptide or 0 peptide and then incubated at 37°C for an additional 24h. The optical density was measured prior to staining to measure bacterial growth after 48h incubation. Eight wells were used for each peptide (n = 8). Production of biofilm was measured using crystal violet staining as described previously [31].
Protein ID numbers were obtained from the UniProt protein database. The protein ID number of LL-37, the human cathelicidin, is P49913. The protein number for SMAP-29, the sheep cathelicidin, is P49928. The protein number for NA-CATH, the cathelicidin from the Chinese King Cobra, is B6S2X0.
In this study, we demonstrated the effectiveness of various snake-derived cathelicidin peptides against B. thailandensis, including NA-CATH. Control peptides included SMAP-29 and LL-37 (Table 2). Ceftazidime is the first-line antibiotic against B. pseudomallei [32]. Therefore, it was used as a positive control for observing a bactericidal effect in the antimicrobial plating assay. We determined the first-line antibiotic ceftazidime to have an EC50 value of 0.328 μM (95% CI of 0.20–0.548 μM) (Fig 1).
We found that the Naja atra peptide NA-CATH had EC50 values of 0.877 μM (95% CI of 0.61–1.26 μM) against B. thailandensis. This compared favorably to the sheep peptide SMAP-29, which was observed to have an EC50 value of 0.628 μM (95% CI of 0.194–2.03 μM). The human cathelicidin LL-37 was also found to have a good antimicrobial effect against B. thailandensis, with an EC50 value of 1.87 μM (95% CI of 1.20–2.94 μM). Data are presented in μM to reflect the number of peptide molecules, thereby compensating for differing molecular weights. These results are consistent with the published values for the effect of LL-37 against Burkholderia [33–35]. These three cathelicidin peptides (NA-CATH, SMAP-29, LL-37) were not statistically different in their anti-B. thailandensis performance. The activity of LL-37 is similar to that shown for B. pseudomallei [34]. In previous work, these same cathelicidin peptides were tested against P. aeruginosa, S. aureus, and F. novicida [14–16]. We had expected Burkholderia species to have a higher EC50 than those organisms because of their wide range of mechanisms to evade destruction by antibiotics and antimicrobial peptides [10]. Surprisingly, the EC50 results against B. thailandensis were similar to cathelicidin EC50 values against other Gram-negative bacteria [14–16].
Naja atra cathelicidin peptide derivatives were tested for antimicrobial activity against B. thailandensis. Each imperfect repeat from NA-CATH (ATRA-1 and ATRA-2) was tested, as well as the synthetic peptide ATRA-1A, in which amino acid 3 of ATRA-1 was switched from phenylalanine to alanine [13] [26]. The EC50 of ATRA-1 against B. thailandensis was determined to be 6.94 μM (95% CI of 4.26–11.3 μM) (Fig 2). EC50 plating assays determined that ATRA-2 was not an effective antimicrobial peptide, correlating with our previous ATRA-2 results with other bacteria [14–16]. This leads to the conclusion that the first imperfect repeat of NA-CATH contributes to most of the observed antimicrobial activity of NA-CATH. We then looked at a synthetic peptide, named ATRA-1A, which contains a single amino acid change at position 3 (F->A). This synthetic peptide exhibited an EC50 of 9.83 μM (95% CI of 6.52–14.8 μM).
In previous work we demonstrated that peptides produced with each amino-acid in the D-form (enantiomer) can be antimicrobial [14,15,36]. In addition, peptides in the D-form are resistant to proteases such as trypsin [14,15,37,38]. Dean et al. demonstrated that while the L-form of the LL-37 peptide is digested by trypsin, the D-form shows no degradation after 1 h trypsin digestion [15]. Thus, we synthesized all-D-enantiomers of LL-37 and ATRA-1A to compare the antimicrobial activity of these protease-resistant enantiomers.
We found the antimicrobial effect of the D-enantiomer to be comparable to that of the L-enantiomer for both ATRA-1A and LL-37 (Fig 3). LL-37 had an EC50 value of 1.87 μM (95% CI of 1.20–2.94 μM), while D-LL-37 had a statistically similar EC50 of 3.64 μM (95% CI of 2.04–6.53 μM). D-ATRA-1A had an EC50 value of 4.82 μM (95% CI of 3.20–7.27 μM, as compared to 9.83 μM (95% CI of 6.52–14.8 μM) for ATRA-1A. For both enantiomeric conversions, the 95% confidence intervals of the D-peptide results overlapped those from the normal L version of the peptide. These data demonstrate that converting each peptide to an all-D-enantiomer did not statistically alter its antimicrobial effect.
We examined a second category of CAMPs, the defensins, for anti-Burkholderia activity. Under conditions of B. pseudomallei respiratory infection in mice, neutrophil granules were observed to be the predominant cell type seen in association with B. pseudomallei infection [39]. Neutrophil granules are known to be a significant source of cathelicidins and human neutrophil peptides (alpha-defensins) [40]. Therefore, to further explore the effect of defensins upon B. thailandensis, human alpha defensin-1 (aka human neutrophil peptide 1, HNP-1) and human alpha defensin-2 (HNP-2) were chosen as candidates to test for antimicrobial killing against B. thailandensis (Fig 4). For HNP-1, at the highest concentration tested (1000 μg/ml peptide), only 65% killing could be achieved for B. thailandensis, suggesting that this is a highly ineffective peptide. HNP2 was even less effective than HNP1 in killing B. thailandensis at every concentration tested.
Sahly et al. demonstrated that the LD50 of human beta-defensin-3 (hBD-3) against multiple Burkholderia species was >100 μg/ml [18]. However, other reports demonstrated that regions of cationic peptides in the C-terminus of hBD-3 possessed antimicrobial activity against E. coli and P. aeruginosa [41,42]. Based on our previous work, we tested a small fragment (Peptide 4) of human beta-defensin 3 (hBD-3), which was previously shown to have significant activity against another Gram-negative bacterium, E. coli [42]. Incubation of B. thailandensis with Peptide 4 of hBD-3 at the highest concentration tested (1000 μg/ml) resulted in only 65% killing for B. thailandensis, suggesting that this is a highly ineffective peptide. These findings confirmed published reports [18,43] that the beta-defensin CAMPs are ineffective against Burkholderia species, and demonstrated that the alpha-defensins are also ineffective against B. thailandensis.
As B. pseudomallei has been reported to form biofilm [44,45], we sought to demonstrate the ability of various cathelicidins to inhibit biofilm formation in the model organism B. thailandensis. In addition, we and others have demonstrated that the cathelicidin LL-37 inhibits biofilm formation in P. aeruginosa [15,46], an important gram-negative pathogen. Therefore, a panel of cathelicidins was tested in biofilm inhibition assays. We first had to establish conditions under which the biofilm of B. thailandensis could be reliably formed and measured. To do this, we grew the bacteria in modified Vogel-Bonner medium (MVBM) [30] overnight and then adjusted OD540 to 0.8. Following growth and measurement, a biofilm inhibition assay was performed. This assay incubates the test compound with the bacteria and determines if the compound can inhibit biofilm formation. The growth of bacteria is also measured to control for bactericidal effects of the test compound although bactericidal effects are reduced due to the high salt concentration of the media used in these assays. We demonstrated that the antibiotic ceftazidime did not inhibit biofilm production but simply killed the B. thailandensis (Fig 5), as we would expect. Interestingly, the cathelicidins LL-37, SMAP-29, and NA-CATH all showed at least 50% biofilm inhibition at peptide concentrations at or above 3 μg/ml. The negative control peptide, which was a scrambled LL-37 (same amino acid composition and net charge, different sequence of amino acids), did not inhibit biofilm, as we previously reported [14].
In addition, the D-enantiomers of the peptides were tested for their effect on biofilm inhibition. D-LL-37 produced results similar to those of its L-enantiomer. Both inhibited at least 50% of biofilm at concentrations as low as 3 μg/ml. Thus both the L- and D- form of LL-37 exhibit anti-biofilm activity against B. thailandensis. When the ATRA-1A enantiomers were compared, the results differed slightly (S1 Fig). ATRA-1A did not inhibit biofilm formation, whereas D-ATRA-1A did slightly, but only at the highest concentration of peptide tested (300 μg/ml). At these levels, this is unlikely to be a significant activity of the D-ATRA-1A peptide. ATRA-1 and ATRA-2, the imperfect repeats from NA-CATH, were also tested for biofilm inhibition and did not inhibit biofilm formation (S1 Fig). Thus, the full-length cathelicidin peptides including the snake cathelicidin NA-CATH were able to inhibit B. thailandensis biofilm formation.
We have demonstrated that some of our cathelicidins can inhibit biofilm formation in B thailandensis. We wanted to know if these cathelicidin peptides could also disperse pre-formed biofilms. Therefore, a number of cathelicidins were chosen for the pre-formed biofilm dispersion assay as described in the methods. We demonstrated that LL-37 and D-LL-37 were able to disperse at least 50% of the pre-formed biofilm when 10 μg (11μM) peptide was added to 24h pre-formed biofilms (Fig 6), showing statistically equivalent activity. Other cathelicidins, SMAP-29 and NA-CATH (which demonstrated biofilm inhibition) did not demonstrate the ability to disperse pre-formed biofilms. Also, as expected, small NA-CATH derivative peptides (ATRA-1A, D-ATRA-1A, ATRA-2) showed no biofilm dispersion activity. The ability of D-LL-37 to disperse preformed Burkholeria biofilm has not been previously reported. These results suggest that LL-37 and D-LL-37 have a unique property that enables dispersion of preformed biofilm in this organism.
B. thailandensis and B. pseudomallei have a wide range of mechanisms for evading antibiotics and antimicrobial peptides. These mechanisms include, but are not limited to, having a more impermeable membrane, multi-drug-resistant efflux pumps, inactivation of host proteins, and modification of drug targets [10]. In this study, we demonstrate that certain peptides can evade these mechanisms and exhibit antimicrobial activity against B. thailandensis, despite its reported extreme peptide resistance. We also demonstrate that D-amino acid peptides exhibit comparable antimicrobial activity [15]. Finally, we describe the anti-biofilm activity of some of these peptides against B. thailandensis.
The modes of action of CAMPs against bacteria are varied and require further elucidation; however, two potentially co-existing mechanisms exist. The first model is that these CAMPs cause the formation of transmembrane pores, causing dissolution of the membrane potential and eventual destruction of the bacterial cell [47]. A second proposed mechanism includes the internalization of the CAMP which can then bind to internal targets and interfere with cell wall synthesis [48]. In humans, only one cathelicidin has been identified: LL-37. This cathelicidin is released by proteolysis from the C-terminus of CAP18 protein [49]. The LL-37 cathelicidin, as well as other cathelicidins with similar alpha-helical structures, has been demonstrated to associate with the bacterial membrane and cause bacterial death [50,51]. The effective killing concentration of these alpha-helical cathelicidin-type CAMPs has driven the search for new cathelicidins. Recently, a cathelicidin has been discovered in the species Naja atra, the Chinese cobra [25,52] which is effective against multiple bacteria [13–16]
Previous reports indicated that LL-37 is antimicrobial against B. thailandensis and B. pseudomallei. It has been reported that 15 μM LL-37 in 1 mM potassium phosphate buffer (PPB) kills 105 cfu/mL of B. thailandensis [35]. Another study with B. thailandensis demonstrated that 5 strains of B. thailandensis were >90% killed at concentrations of 12.5 mg/L or greater in 1 mM PPB [33]. We obtained the same result, but in this study we have reported our results in terms of EC50 rather than lethal concentration. Research has also shown that LL-37 is effective at killing B. pseudomallei. One study reports that LL-37 effectively killed 24 strains of B. pseudomallei at a peptide concentration of 100 μM in 1 mM PPB [34], while another demonstrated >90% killing of 9 strains at concentrations of 6.25 mg/L (1.39 μM) or greater [33]. Under conditions of B. pseudomallei respiratory infection in mice, neutrophil granules were observed to be the predominant cell type seen in association with B. pseudomallei infection [39]. Since neutrophil granules are known to be a significant source of cathelicidins [40], our data and published results suggest a significant potential role of LL-37 expression during the infection of humans by Burkholderia species. We were able to demonstrate that LL-37, SMAP-29, NA-CATH, and small NA-CATH-derived ATRA peptides exert strong antimicrobial activity against B. thailandensis. Another group reported that the bovine cathelicidin BMAP-18 was antimicrobial against B. pseudomallei at 20 μM [53]. Together, these studies suggest that B. thailandensis and B. pseudomallei may be quite susceptible to cathelicidins as a class of peptides. A recent study also found that additional peptides, including the 12-aa peptide bactenecin, the hybrid peptide CA-MA, and RTA3, were antimicrobial against B. pseudomallei [53], suggesting that there are peptides that appear to be effective against this organism, particularly those with predominantly helical properties.
In addition, we addressed the potential issue of proteolytic degradation of AMPs in vivo by bacterial proteases. Sieprawska-Lupa et al. demonstrated that mammalian hosts and numerous bacteria express proteases capable of degrading and inactivating LL-37 [54]. Therefore, we tested D-enantiomers, which we previously showed to be resistant to trypsin digestion [15], to compare their antimicrobial activity to that of the natural L-enantiomer. Our results show that for both LL-37 and ATRA-1A, the D-enantiomer exhibited antimicrobial activity comparable to that of the L-enantiomer, suggesting that bacterial proteases were not active against this peptide.
Our results also demonstrate that defensin peptides have at best a weak antimicrobial effect against Burkholderia. Human beta-defensins had previously been shown to be ineffective against B. cepacia or B. pseudomallei [18,43]. The alpha-defensins HNP-1 and HNP-2 both demonstrated poor antimicrobial performance against B. thailandensis in this work. For HNP-1, this is consistent with the literature on B. pseudomallei [55,56]. (No previous work was found on HNP-2’s effect on Burkholderia pseudomallei.) In addition, we tested HNP-3 and HNP-4 (Fig 4B) and found a similar lack of antimicrobial activity. This leads us to conclude that alpha-defensins in general do not exert strong antimicrobial activity against B. thailandensis. A fragment of hBD3 (Peptide 4 from hBD3) was also ineffective against B. thailandensis. Thus, Burkholderia does seem to be highly resistant to both classes of defensins.
It has also been demonstrated that both B. thailandensis and B. pseudomallei form biofilms in vivo [44,45], which may be a virulence factor [45]. We were able to demonstrate that the cathelicidins LL-37, D-LL-37, NA-CATH, and SMAP-29 are capable of biofilm inhibition in B. thailandensis each at similar extents of ~50% inhibition at 3 μg/ml. This is in agreement with the published capability of LL-37 to inhibit biofilm formation in Pseudomonas [46]. In addition, we demonstrated the new result of the ability of D-LL-37 to disperse pre-formed biofilms. Thus, the biofilm inhibition we demonstrate in this work may be a crucial component of the activity of cathelicidin-derived peptides as possible therapeutics.
Novel approaches to treatment for Burkholderia infections are critically needed, especially for treatment of melioidosis. A novel peptide-based treatment for melioidosis would ideally include both antimicrobial activity and biofilm inhibition, and may take the form of a topical application. In this work, we have demonstrated the effects of LL-37, SMAP-29, and NA-CATH as both antimicrobial and anti-biofilm peptides, and showed promising results of short, synthetic peptides, such as ATRA1. We have also extended previous studies [35] showing here that an all-D-enantiomer of LL-37, which is resistant to proteolytic degradation, maintains antimicrobial activity as well as significant anti-biofilm properties against B. thailandesnsis. The results of this study illustrating the susceptibility of B. thailandensis to cathelicidin-like peptide killing, resistance to defensins, and the ability of D- and L-LL-37 peptides to inhibit biofilm formation may provide a new understanding of the potential use for peptides, perhaps as topical applications, in melioidosis infection.
|
10.1371/journal.ppat.1006311 | CD8+ T cell evasion mandates CD4+ T cell control of chronic gamma-herpesvirus infection | Gamma-herpesvirus infections are regulated by both CD4+ and CD8+ T cells. However clinical disease occurs mainly in CD4+ T cell-deficient hosts. In CD4+ T cell-deficient mice, CD8+ T cells control acute but not chronic lung infection by Murid Herpesvirus-4 (MuHV-4). We show that acute and chronic lung infections differ in distribution: most acute infection was epithelial, whereas most chronic infection was in myeloid cells. CD8+ T cells controlled epithelial infection, but CD4+ T cells and IFNγ were required to control myeloid cell infection. Disrupting the MuHV-4 K3, which degrades MHC class I heavy chains, increased viral epitope presentation by infected lung alveolar macrophages and allowed CD8+ T cells to prevent disease. Thus, viral CD8+ T cell evasion led to niche-specific immune control, and an essential role for CD4+ T cells in limiting chronic infection.
| Gamma-herpesviruses chronically infect most people. While infection is usually asymptomatic, disease occurs if the immune system is weakened. Understanding how immune control normally works should provide a basis for preventing disease. In mice, CD8+ T cells can control acute gamma-herpesvirus infection but not chronic infection. We show that acute and chronic infections involve different cell types. CD8+ T cells controlled epithelial cell infection, which predominated acutely, but they could not control chronic macrophage infection unless viral immune evasion was disabled. Instead CD4+ T cells were required. Thus, viral evasion made host defence cell type-specific: CD8+ T cells controlled epithelial cell infection; CD4+ T cells controlled macrophage infection; and comprehensive control required both T cell subsets.
| Herpesviruses chronically infect immunocompetent hosts. CD4+ and CD8+ T cells both help to contain these infections, but disease occurs mainly when CD4+ T cells are lacking [1], implying that they have particular importance. Among the gamma-herpesviruses, CD4+ T cell deficiency leads Epstein-Barr virus (EBV) to cause lymphoproliferative disease and oral hairy leukoplakia, a virus-productive epithelial lesion [2]; it leads the Kaposi's Sarcoma-associated Herpesvirus (KSHV) to cause endothelial cell proliferation with inflammation and viral lytic gene expression [3]; and it leads MuHV-4 to replicate chronically in the lungs [4]. Thus the pathologies of CD4+ T cell-deficient hosts vary, but increased lytic infection is a common theme.
Gamma-herpesviruses characteristically persist in lymphocytes. EBV, KSHV and MuHV-4 all persist in B cells. However to reach B cells then re-emerge to reach new hosts they must also infect other cell types. EBV emerging from plasma cells [5] reaches the saliva via epithelial cells [6]. The normal association of plasma cells with mucosal epithelial cells provides a basis for virus transfer. How EBV reaches naive B cells is less well understood, as they have little direct communication with mucosal epithelia. Antigen presentation by myeloid cells provides a potential route to naive B cells. KSHV can infect many cell types [7], including myeloid cells [8]; EBV colonization of NK cell and T cell cancers [9] suggests a broader tropism than is usually evident in vitro; and MuHV-4, after epithelial host entry [10, 11], reaches B cells via dendritic cells [12]. Myeloid cell infection also features prominently in acute MuHV-4 colonization of splenic B cells [13–15]. Thus while only modestly efficient in vitro [16] and hard to detect in the long-term [13], myeloid cell infection plays a key role in MuHV-4 tropism.
Acute MuHV-4 lung infection is controlled mainly by CD8+ T cells [17]. They also help to control splenic B cell infection [18], and macrophage infection after peritoneal challenge [19]. β2-microglobulin-deficient BALB/c mice show a 3-fold increase in lymphoma incidence after MuHV-4 infection [20]. However β2-microglobulin deficiency impairs more than just than CD8+ T cell function, for example it reduces serum IgG [21]. Therefore the increased lymphoma incidence was not just CD8+ T cell dependent. Moreover few if any lymphoma cells showed evidence of MuHV-4 infection [20], and no lymphomas were seen in MuHV-4-infected β2M-/- C57BL/6 [22] or 129 mice [20]. Inbred mice are prone to lymphomagenesis by strain-polymorphic endogenous retroviruses [23], and gamma-herpesviruses can transactivate retroviruses [24]. Therefore the ontogeny of the lymphomas remains unclear. The most obvious consequence of CD8+ T cell deficiency for MuHV-4 is increased lytic infection [22]. While cancers are the most harmful outcome of EBV infection, T cell deficiency again mainly increases lytic infection [25].
CD4+ T cell-deficient mice also show more lytic infection. However unlike CD8+ T cell-deficient mice, and despite maintaining strong anti-viral CD8+ T cell responses [26–28], they suffer a wasting disease [4]. Anti-MuHV-4 antibody responses help to contain infection [29] and depend on CD4+ T cells [30], but a lack of antibody alone does not explain the disease of CD4+ T cell-deficient mice, as B cell-deficient mice survive [31]. Acutely CD4+ T cells suppress MuHV-4 replication independently of B cells, with an important role for interferon-γ (IFNγ) [32, 33], so their effector function may also be important for long-term MuHV-4 control.
Why CD8+ T cells alone fail to control MuHV-4 is important to understand because they are a therapeutic focus for EBV. CD8+ T cell evasion is a near universal characteristic of mammalian herpesviruses. While its molecular mechanisms have been studied extensively, its impact on infection is less well understood. The MuHV-4 K3 degrades MHC class I (MHC I) [34] and the transporter associated with antigen processing (TAP) [35]. K3 disruption impairs virus-driven lymphoproliferation [36]. We show that in chronic infection, K3 protects lung macrophages against CD8+ T cells and so makes CD4+ T cells essential to prevent disease.
MuHV-4 replicates chronically in MHC class II (MHC II)-deficient (IA-/-) C57BL/6 mice, which lack classical CD4+ T cells [4]. Intranasal (i.n.) BAC-derived MuHV-4 reached similar peak titers in the lungs of IA-/- and wild-type control mice (WT, IA+/-) at day (d) 5 post-infection, but then maintained higher titers in IA-/- mice (Fig 1A). The main cell populations of the lung alveoli are type 1 epithelial cells (AEC1), which have a characteristically flattened shape with a large surface area for gas exchange, and express podoplanin (PDP); type 2 AEC (AEC2), which express surfactant proteins; and alveolar macrophages (AM), which phagocytose inhaled debris and express CD68. MuHV-4 entering the lungs binds to AEC1 and is then captured by AM [11]. Subsequent replication in AM allows spread back to AEC1 and makes them the main site of acute virus production. At d5, immunostaining for viral lytic antigens showed mainly AEC1 infection in both WT and IA-/- mouse lungs (Fig 1B and 1C; S1 Fig; S2 Fig). By d9, WT lungs contained few infected cells. IA-/- lungs contained significantly more (p<0.01), and most of these were AM. AM infection remained detectable at d30 in IA-/- (Fig 1B and 1D; S3 Fig) but not WT mice. Few AEC2 or lung B cells expressed viral lytic antigens. Thus WT mice resolved acute AEC1 infection, while in IA-/- mice it evolved into a chronic infection of AM.
CD4+ T cell-depleted C57BL/6 mice showed a similar picture (Fig 1E and 1F). The virus titers in depleted mouse lungs were equivalent to those of undepleted controls at d5, then higher at d12 (Fig 1E). Lung sections showed mainly AEC1 infection at d5, and mainly AM infection at d12 (Fig 1F). We did not see evidence of chronic epithelial cell infection, as reported for B cell-deficient mice [37]. Ongoing myeloid cell infection may seeds epithelial infection in some settings, but the main cell type supporting chronic lung infection in CD4+ T cell-deficient mice was myeloid.
MuHV-4 causes disease more readily in BALB/c than in C57BL/6 mice [20], with acute protection being CD8+ T cell-dependent [17]. We tested whether BALB/c mice also showed CD4+ T cell-dependent myeloid infection control (Fig 2). Live imaging of i.n. luciferase+ MuHV-4 showed CD4+ T cell depletion significantly increasing lung and nose infections at d7 and d9 (Fig 2A). CD8+ T cell depletion had significantly more effect, and dual depletion had more effect still. CD8+ T cell depletion also increased colonization of the superficial cervical lymph nodes (SCLN), which drain the upper respiratory tract, while CD4+ T cell depletion reduced SCLN colonization, consistent with the amplification of B cell infection in lymphoid tissue being CD4+ T cell-dependent [38]. D9 virus titers in lungs and noses (Fig 2B) matched the luciferase signals, with dual depleted > CD8+ T cell depleted > CD4+ T cell depleted > undepleted controls. Thus acutely, when epithelial infection predominated, CD8+ T cells contributed more than CD4+ T cells to controlling virus replication in both the upper and lower respiratory tract.
Immunostaining infected BALB/c lungs at d9 (Fig 2C–2E) showed significantly more AEC1 than AM infection in all groups except that depleted of CD4+ T cells, which showed significantly more AM infection. Epithelial and fibroblast infections were consistently more virus-productive than myeloid cell infection in vitro (Fig 2F), and the higher virus titers of mice with more AEC1 infection were consistent with AEC1 producing more virus acutely than AM. Thus, CD8+ T cells appeared to be more important than CD4+ T cells for acute infection control because they targeted a more immediately virus-productive cell type—AEC1—while a lack of CD4+ T cells increased AM infection.
In mice lacking B cells and CD8+ T cells, IFNγ is required for acute infection control [32, 33], suggesting that it mediates the anti-viral effect of CD4+ T cells. It also inhibits ex vivo MuHV-4 reactivation from peritoneal macrophages [39]. In otherwise immunocompetent BALB/c mice, IFNγ neutralization increased d9 lung virus titers significantly more than did CD4+ T cell depletion (Fig 3A), and increased both AM and AEC1 infections (Fig 3B and 3C), implying that it also mediated other anti-viral effects. Again CD4+ T cell depletion decreased MuHV-4 colonization of lymphoid tissue, whereas IFNγ neutralization increased it (Fig 3A).
CD4+ T cells, CD8+ T cells and NK cells all produce IFNγ. When CD8+ T cells were eliminated, CD4+ T cell depletion increased virus titers and AM infection significantly more than did IFNγ neutralization (Fig 3D and 3E). Therefore while IFNγ was an important CD4+ T cell-mediated defence, it was not the only one and it contributed also to CD8+ T cell-mediated defence. NK cell depletion does not affect the course of MuHV-4 lung infection in otherwise immunocompetent mice [40], but increases LN infection by MuHV-4 inoculated into footpads [41]. In C57BL/6 mice NK cells did not make a significant contribution to infection control in lungs at d10, making it unlikely that they were a significant source of IFNγ in this setting (Fig 4A). They did contribute to infection control in noses (Fig 4B). Here NK cell depletion increased virus titers regardless of whether CD4+ T cells were depleted. Therefore CD4+ T cells and NK cells functioned as independent defences.
CD4+ and CD8+ T cells differ in both target cell recognition and predominant effector functions: CD8+ act mostly via perforin and granzymes, while IFNγ is a key effector for CD4+ T cells [42]. Thus, CD4+ T cell-dependent myeloid infection control could have reflected either that only CD4+ T cells efficiently recognized infected myeloid cells (via MHC II), or that only IFNγ was able to control their infection. To explore these possibilities we tracked the infection of MHC II+ lung cells (Fig 5). In naive lungs, 1/3 of MHC II+ cells were CD11c+ AM or dendritic cells, and 2/3 were surfactant protein C precursor (SPC)+ AEC2 (Fig 5A). After MuHV-4 infection, most MHC II+ cells (>70%) were SPC-, presumably reflecting myeloid cell recruitment and MHC II up-regulation. Almost all MuHV-4+ cells (>95%) were SPC-, that is myeloid rather than AEC2 (Fig 5B).
Again we depleted CD8+ T cells as a source of IFNγ, then compared additional CD4+ T cell depletion with IFNγ neutralization. CD4+ T cell depletion increased the number of infected MHC II+ lung cells, while IFNγ neutralization gave only a non-significant increase (Fig 5C and 5D). Most AM express CD11c [43, 44]. Both IFNγ neutralization and CD4+ T cell depletion increased significantly the number of CD11c+MHC II+ infected cells. CD4+ T cell depletion but not IFNγ neutralization significantly increased the number of CD11c-MHC II+ infected cells (Fig 5E and 5F). SPC+ infection remained rare (Fig 5G), so CD11c-MHC II+MHV+ cells were presumably infected CD11c- AM or infiltrating monocytes. Thus, CD4+ T cell depletion increased MuHV-4 infection of MHC II+ lung myeloid cells, and IFNγ neutralization reproduced much of this effect, consistent with IFNγ production being an important CD4+ T cell effector function. However the greater effect of CD4+ T cell depletion than IFNγ neutralization on lung myeloid cell infection implied that target cell recognition was the key parameter, rather than susceptibility to IFNγ.
The importance of CD4+ T cell recognition for AM infection control implied poor CD8+ T cell recognition. Virus-specific CD8+ T cells were evidently functional in IA-/- mice, as they controlled AEC1 infection; and myeloid cells are normally good CD8+ T cell targets [45]. However the MuHV-4 K3 degrades MHC I and TAP. To test whether K3 compromised CD8+ T cell recognition of AM, we exposed AM to K3+ or K3- viruses, then measured epitope presentation to a MuHV-4-specific CD8+ T cell hybridoma (Fig 6A). K3 disruption significantly increased hybridoma stimulation by both WT and IA-/- AM.
To establish whether better CD8+ T cell recognition of infected AM translated into better infection control, we infected IA-/- mice with K3+ or K3- MuHV-4. K3+ viruses caused significantly more disease (weight loss and general ill health requiring euthanasia) (Fig 6B) and reached higher titers in both lungs and spleens (Fig 6C). When CD8+ T cells were depleted, K3+ and K3- viruses reached equivalent titers (Fig 6D). Immunostaining IA-/- lung sections at infection d10 (Fig 6E and 6F) showed that K3- viruses lacked the chronic AM infection of WT MuHV-4. Infectious center assays of AM recovered by lung washout (Fig 6G) confirmed greater infection by K3+ viruses. Therefore K3 limited CD8+ T cell-mediated control of MuHV-4 replication in lung myeloid cells.
Gamma-herpesviruses establish chronic, low-level transmission with generally few symptoms. Immunodeficiencies shift this equilibrium towards greater viral replication and disease. The key parameters of disease control in humans have been hard to define. Thus, anti-viral therapies have remained largely empirical. Our analysis of murine infection showed cell type-specific immune control. In mice lacking CD4+ T cells, CD8+ T cells still controlled epithelial infection. We did not see extensive B cell infection, presumably due to a lack of CD4+ T cell-dependent B cell proliferation [38]. However myeloid cell colonization, which is normally transient, became chronic and caused disease. This disease depended on CD8+ T cell evasion by the MuHV-4 K3: when K3 was disrupted, CD8+ T cells achieved long-term infection control and CD4+ T cells were not required.
Why did K3 protect infected macrophages and not epithelial cells against CD8+ T cells? K3 stabilization by tapasin [46] titrates its expression to the cellular capacity for antigen presentation, for example overcoming induction by IFNγ [35]. However immune evasion only raises the threshold for epitope presentation: peptides competing strongly for the few remaining MHC I complexes can still be recognized. Infected epithelial cells produced more virus than infected myeloid cells, implying that they produced more viral peptides, making break-through viral epitope presentation more likely. Cell type differences in susceptibility to CD8+ T cell attack are also possible. The faster clearance of pro-lytic MuHV-4 mutants from the lungs despite faster spread in vitro [47, 48] suggests that more indolent gamma-herpesvirus infections generally constitute more difficult immune targets. Myeloid cell infection is a common characteristic of lymphotropic viruses [49, 50], and MuHV-4 myeloid cell infection caused chronic disease despite limited virus production. Therefore poorly lytic infection should not exclude myeloid cell infection from consideration as a source of human gamma-herpesvirus-driven disease.
While IA-/- mice make large CD8+ T cell responses to MuHV-4 [26], the non-uniformity of in vivo infection means that large immune responses are not always the most effective responses. For example MuHV-4-infected mice normally mount a large CD8+ T cell response to viral reactivation from latency in B cells [51]; yet if MHC I epitope presentation is enforced during viral episome maintenance [52], a relatively small CD8+ T cell response essentially abolishes latent B cell infection and hence also reactivation. The large acute CD8+ T cell responses to EBV lytic antigens [53] analogously imply a failure to suppress virus-driven lymphoproliferation. In IA-/- mice CD8+ T cells kept AEC1 infection in check, but they did not shut down virus production by K3-protected myeloid cells. This required CD4+ T cells. Thus without CD4+ T cells, myeloid infection could constantly re-seed epithelial infection.
CD4+ T cells may be difficult for MuHV-4 to evade because it needs them to drive infected B cell proliferation. Also MHC II presents mainly exogenous antigens, so the presenting cells need not be infected, making them difficult to target. While CD4+ T cells have limited cytotoxic capacity, they can trigger apoptosis via tumor necrosis factor receptors and fas, activate myeloid cells to reduce their susceptibility to infection [54], and through cytokine signalling repress viral lytic gene expression directly [55]. Thus, there are abundant opportunities for CD4+ effector T cells to restrict MuHV-4 replication.
Most studies of anti-viral immunity have averaged outcomes across whole organs. Direct visualization is revealing additional complexity. For example CD8+ T cells combat cutaneous vaccinia virus by killing infected monocytes rather than keratinocytes [56]. Direct visualization revealed that immune evasion makes MuHV-4 control cell type-specific: CD8+ T cells controlled epithelial infection, and CD4+ T cells controlled myeloid infection. Thus CD4+ and CD8+ T cells co-operated, but less through classical help than through recognizing distinct components of a complex infection. Such niche-specific immune function suggests that single component vaccines eliciting mainly one effector class might only ever have partial efficacy against complex viruses; multi-component vaccines that prime complementary defences may be necessary for full protection.
Adult C57BL/6, BALB/c, and C57BL/6 back-crossed IA-/- mice [57] were infected i.n. with MuHV-4 (104 p.f.u.) under isoflurane anaesthesia. Luciferase+ infection was imaged by peritoneal (i.p.) injection of D-luciferin (2mg/mouse, Pure Science) and charge-coupled camera scanning (IVIS spectrum, Xenogen). IFNγ was neutralized by i.p. mAb XMG1.2 (200μg/mouse/48h). CD4+ and CD8+ T cells were depleted by i.p. mAbs GK1.5 and 2.43 (200μg/mouse/48h, from 96h before infection). NK cells were depleted with NK1.1-specific mAb PK-136 (200μg/mouse/48h, from 48h before infection). Antibodies were from Bio X Cell. T cell subset depletion, measured by flow cytometry of spleen cells with antibodies to a distinct CD4 epitope (rat mAb RM4-4), and to CD8β (rat mAb H35-17.2) and was >95% complete. CD4+ T cell-depleted mice further lacked detectable MuHV-4-specific serum IgG by ELISA (S4 Fig). NK cell depletion, monitored by flow cytometric staining of spleen cells for CD49d with mAb DX5, was >90% complete. Statistical comparisons were by Student's 2-tailed unpaired t test unless otherwise stated.
All experiments were approved by the University of Queensland Animal Ethics Committee in accordance with the Australian code for the care and use of animals for scientific purposes, from the Australian National Health and Medical Research Council (project 301/13).
Peritoneal macrophages were recovered by peritoneal lavage. After discarding non-adherent cells, the remainder were >80% F4/80+ by flow cytometry. Lung macrophages were recovered by bronchio-alveolar lavage and were >70% CD11c+ by flow cytometry. These cells, BHK-21 fibroblasts (American Type Culture Collection (ATCC) CCL-10), RAW-264 monocytes (ATCC TIB-71), NMuMG epithelial cells (ATCC CRL-1636), the 49100.2 T cell hybridoma [58], and mouse embryo fibroblasts were grown in Dulbecco’s Modified Eagle’s Medium with 2mM glutamine, 100IU/ml penicillin, 100μg/ml streptomycin, and 10% fetal calf serum (complete medium).
Luciferase+ [59] and GFP+ [60] MuHV-4, a K3- mutant and its revertant [36] and an independent K3 mutant (K3-I) [54] were propagated in BHK-21 cells. Infected cell supernatants were cleared of debris by low speed centrifugation (200 x g, 5 min). Cell-free virions were then concentrated by ultracentrifugation (35,000 x g, 1.5h). To titer infectious virus, freeze-thawed samples were plated on BHK-21 cells (plaque assay); to titer total reactivatable MuHV-4, live cells were plated (infectious center assay) [60]. After 2h the cells were overlaid with complete medium plus 0.3% carboxymethylcellulose, cultured for 4d (37°C in complete medium) then fixed with 1% formaldehyde and stained with 0.1% toluidine blue.
Organs were fixed in 1% formaldehyde / 10mM sodium periodate / 75mM L-lysine (18h, 4°C), equilibrated in 30% sucrose (24h, 4°C), then frozen in OCT. 6μm sections were air-dried, washed 3x in PBS, blocked with 0.3% Triton X-100 / 5% donkey serum (1h, 23°C), then incubated (18h, 4°C) with combinations of antibodies to GFP (rabbit or goat pAb), CD68 (rat mAb, FA-11) (AbCam), B220 (rat mAb RA3-6B2), CD4 (rat mAb GK1.5), CD11c (hamster mAb N418, Abcam), MHC II (rat mAb M5/114, Serotec), SPC (goat pAb; Santa Cruz Biotechnology), podoplanin (goat pAb, R&D Systems), and MuHV-4 (rabbit pAb). The MuHV-4-immune serum was raised by 2x subcutaneous inoculation of rabbits with MuHV-4 virions (109 p.f.u.). Like previously described immune sera [61] it recognizes a range of virion proteins by Western blot, including the products of ORF4 (gp70), M7 (gp150) and ORF65 (p20). Sections were washed 3×, incubated (1h, 23°C) with combinations of Alexa568-donkey anti-rat IgG pAb, Alexa488 or Alexa647-donkey anti rabbit IgG pAb, Alexa 647-goat anti-hamster IgG pAb (AbCam), and Alexa488-donkey anti-goat IgG pAb (Life Technologies), then washed 3×, mounted in Prolong Gold with DAPI (Life Technologies), and imaged with a Zeiss LCM510 confocal microscope.
Lung macrophages were infected or not with MuHV-4 (3 p.f.u./cell, 4h), washed, then incubated (18h, 37°C) in complete medium with 49100.2 T cells, which recognize an immunodominant, H2Db-restricted MuHV-4 epitope and express β-galactosidase (β-gal) from an NFAT-responsive promoter [58]. To assay β-gal the cells were washed in PBS and lysed in PBS / 5mM MgCl2 / 1% NP-40 / 0.15μM chlorophenol-red-beta-D-galactoside (CPRG, Merck Biosciences). 595nm absorbance was read after 2–4h.
MuHV-4 virions in 0.05% Triton-X100 / 50mM sodium carbonate pH = 8.5, were coated (18h, 4°C) onto Maxisorp ELISA plates (Nalge Nunc). The plates were washed x3 in PBS / 0.1% Tween-20, blocked with 1% bovine serum albumin / PBS / 0.1% Tween-20, incubated with 3-fold serum dilutions (1h, 23°C), washed x4 in PBS / 0.1% Tween-20, incubated (1h, 23°C) with alkaline phosphatase-conjugated goat anti-mouse IgG-Fc pAb (Sigma Chemical Co.), washed x5, and developed with nitrophenylphosphate substrate (Sigma). Absorbance was read at 405nm (Biorad).
|
10.1371/journal.pntd.0001594 | Taenia solium Infections in a Rural Area of Eastern Zambia-A Community Based Study | Taenia solium taeniosis/cysticercosis is a parasitic infection occurring in many developing countries. Data on the status of human infections in Zambia is largely lacking. We conducted a community-based study in Eastern Zambia to determine the prevalence of human taeniosis and cysticercosis in a rural community.
Stool and serum samples were collected from willing participants. Geographical references of the participants' households were determined and household questionnaires administered. Taeniosis was diagnosed in stool samples by coprology and by the polyclonal antibody-based copro-antigen enzyme-linked immunosorbent assay (copro-Ag ELISA), while cysticercosis was diagnosed in serum by the B158/B60 monoclonal antibody-based antigen ELISA (sero-Ag ELISA). Identification of the collected tapeworm after niclosamide treatment and purgation was done using polymerase chain reaction-restriction fragment length polymorphism (PCR-RFLP). A total of 255 households from 20 villages participated in the study, 718 stool and 708 serum samples were collected and examined. Forty-five faecal samples (6.3%) were found positive for taeniosis on copro-Ag ELISA while circulating cysticercus antigen was detected in 5.8% (41/708) individuals. The tapeworm recovered from one of the cases was confirmed to be T. solium on PCR-RFLP. Seropositivity (cysticercosis) was significantly positively related to age (p = 0.00) and to copro-Ag positivity (taeniosis) (p = 0.03) but not to gender. Change point analysis revealed that the frequency of cysticercus antigens increased significantly in individuals above the age of 30. Copro-Ag positivity was not related to age or gender. The following risk factors were noted to be present in the study community: free-range pig husbandry system and poor sanitation with 47.8% of the households visited lacking latrines.
This study has recorded high taeniosis and cysticercosis prevalences and identified the need for further studies on transmission dynamics and impact of the disease on the local people.
| Taenia solium taeniosis/cysticercosis is a zoonotic infection endemic in many developing countries, with humans as the definitive host (taeniosis) and pigs and humans as the intermediate hosts (cysticercosis). When humans act as the intermediate host, the result can be neurocysticercosis, which is associated with acquired epilepsy, considerable morbidity and even mortality. In Africa, most studies have been carried out in pigs with little or no data in humans available. In this human study, conducted in a rural community in Eastern Zambia, prevalences for taeniosis and cysticercosis were determined at 6.3% and 5.8% respectively, indicating the hyperendemicity of the area. Cysticercosis infection was strongly related with age, with a significant increase in prevalence occurring in individuals from the age of 30 onward. A collected tapeworm was confirmed to be T. solium. Risk factors associated with the transmission and maintenance of the parasite such as free roaming pigs, households without latrines, backyard slaughter of pigs without inspection and consumption of undercooked pork were also present. The findings of this work have identified the need for further research in the transmission dynamics and the burden that this infection has on the resources of poor local people.
| Taenia solium taeniosis/cysticercosis is a neglected parasitic zoonosis, affecting mostly developing countries [1]. Its occurrence is strongly associated with poverty, poor hygiene and sanitation, free-range pig management and lack of meat inspection [2]–[4]. Adult intestinal tapeworm infection (taeniosis) in man, who is the sole natural definitive host [5], is acquired by eating undercooked infected pork. Infective eggs are shed (with proglottids) via the stool and contaminate the environment. Pigs are the intermediate host and are infected by ingestion of these infective eggs (or proglottids), which develop into cysticerci (porcine cysticercosis). Humans can act as intermediate hosts as well; cysts can develop subcutaneously, intramuscularly, but often in the central nervous system causing neurocysticercosis (NCC). NCC has been described as the most frequently reported helminthic infection of the central nervous system [6] and is a major cause of acquired epilepsy in cysticercosis endemic regions associated with considerable morbidity [7].
In the last decade, many studies have been carried out in sub-Saharan Africa to determine the presence/absence of T. solium. Until now, most studies have been carried out on porcine cysticercosis reporting endemicity in countries like Tanzania, Zambia, Mozambique and Kenya [8]–[12]. Prevalence of human cysticercosis, which has been less studied, ranges from 7.4% in South Africa to 20.5% in Mozambique (both based on specific antibody detection) and 21.6% in the Democratic Republic of Congo (based on circulating antigen detection) [13]–[16]. In studies conducted in Kenya, taeniosis prevalences were estimated from 4 to 10% [17]. These data emphasize the need for more studies in humans to gather information on the epidemiology of the parasite and to estimate the burden on affected communities.
In Zambia, pig keeping and pork consumption have increased significantly during the past decade with Eastern and Southern provinces accounting for a greater part of this increase [4]. Pigs are mostly kept by smallholder producers, under free-range management. Several studies carried out in Zambia have indicated high prevalences of porcine cysticercosis. A study at a slaughter slab in Lusaka receiving village pigs, indicated a prevalence of up to 64.2% [18] while field studies in Southern and Eastern province estimated sero-prevalences between 10.8–20.8% and 9.3–23.3% respectively [19], [20]. These studies clearly showed that T. solium was present in rural areas of Zambia. However, except for preliminary unpublished data, no information was available on human cysticercosis/taeniosis.
The main objective of the current study was to determine the prevalence of human taeniosis and cysticercosis in a rural community of Petauke district in the Eastern province of Zambia.
Ethical clearance (approval) for both human and animal parts of the study was obtained from the University of Zambia Biomedical Research Ethics Committee (IRB0001131). For the human part of the study, further approval was sought from the Ministry of Health of Zambia and also from the local District health authorities. A meeting was held with the community leaders (headmen) to explain the purpose of the study and request their permission to conduct the study in their area. Finally consent was also sought from the individual subjects to take part in the study. Subjects were not forced to participate and participation was requested of individuals of all ages after written informed consent. For individuals below the age of 16, permission was sought from their parents or guardians by way of written informed consent. All participants found positive for taeniosis and other helminths were provided with treatment, namely niclosamide and mebendazole respectively. Those positive for cysticercosis were referred to the District hospital for follow-up and the recommended standard of care provided to them if required.
Collection of reference samples from pigs at a local abattoir was carried out by professional veterinarians adhering to the Zambian regulations and guidelines on animal husbandry.
The study was conducted in a rural community (Kakwiya) in Petauke district of the Eastern province of Zambia (Figure 1). The community is serviced by a Rural Health Centre (RHC) whose catchment population is 11,344 (Clinic headcount records). The people in this community practice subsistence farming growing mostly crops like maize and groundnuts primarily for home consumption and cotton and bananas grown for household income generation. They also keep animals, mostly pigs with a few keeping cattle, goats and chickens. A preliminary visit to the area indicated that, as reported by Sikasunge et al. [21], there was a high number of free roaming pigs and reports of cysts observed in pigs slaughtered in backyards.
A community-based cross sectional survey was conducted in the dry season between July and August 2009. The Kakwiya community was selected because it is a pig keeping community without any active ongoing sanitation programmes and cysticerci were observed in slaughtered pigs. The willingness of the community to participate in the study and the RHC to collaborate was also taken into account together with the availability of an adequate working space in the clinic and staff for the collection of samples. Only villages (n = 20) within a radius of 7 km from Kakwiya RHC were selected. The selected villages were visited and all persons invited to participate in the study. Each willing participant was, after written informed consent, given two plastic sample bottles and requested to submit a stool and a urine sample at the RHC. Upon submission of these samples, a blood sample was also taken by qualified health personnel. A questionnaire targeting household characteristics (such as number of inhabitants in a household, main household income, highest level of education attained, main source of drinking water), presence of household level risk factors (such as keeping of pigs and how they are kept, backyard slaughter of pigs, inspection of slaughtered pigs, consumption of pork by at least one member of the household, presence of a pit latrine, consumption or resell of clearly infected pork) and knowledge of the parasite (such as observation of tapeworm in human faeces, how people acquired a tapeworm, observation of cysts in pork meat, what the cysts were and how pigs acquired them) was administered to each participating household. At the same time geographic co-ordinates of each participating household were obtained using a Global Positioning System (GPS) receiver (eTrex Legend® Cx, Garmin).
Submitted stool samples were divided into two aliquots; one placed in 10% formalin and the other in 70% ethanol and these were stored at 4°C until use. Urine samples were aliquoted in 1.8 ml vials and stored at −20°C. Analyses and results for the urine specimens were discussed in our earlier report [22]. About 5 ml of blood were collected into sterile plain blood collecting tubes and allowed to clot. To obtain the maximum amount of serum, the blood tubes were allowed to stand at 4°C overnight and then centrifuged at 3000 g for 15 minutes. The supernatant (serum) was aliquoted into 1.8 ml vials and stored at −20°C until use. All the samples were transported to Lusaka where they were stored at −20°C until analysis.
Presence of helminth ova in stool was examined microscopically using the formalin-ether concentration technique [23]. Presence of a taeniid egg on a slide was recorded as being positive for taeniosis. The presence of other helminth eggs was also noted during the examination.
An in house copro-antigen detection ELISA (copro-Ag ELISA) as described by Allan et al. [24], with slight modifications, was performed on the stool samples. Briefly, the stool samples stored in 10% formalin were processed by mixing an equal amount of Phosphate Buffered Saline (PBS) and stool sample. This was allowed to soak for one hour with intermediate shaking and centrifuged at 2000 g for 30 minutes. The supernatant was then used in the ELISA. The assay involved coating polystyrene ELISA plates (Nunc® Maxisorp) with the capturing hyper immune rabbit anti-Taenia IgG polyclonal antibody diluted at 2.5 µg/ml in carbonate-bicarbonate buffer (0.06 M, pH 9.6). After coating, the plates were incubated for 1 hour at 37°C, washed once with PBS in 0.05% Tween 20 (PBS-T20) and all wells blocked by adding blocking buffer (PBS-T20+ 2% New Born Calf Serum). After incubating at 37°C for 1 hour and without washing, 100 µl of the stool supernatant was added and plates were incubated for 1 hour at 37°C followed by washing five times with PBS-T20. A biotinylated hyper immune rabbit IgG polyclonal antibody diluted at 2.5 µg/ml in blocking buffer was used as the detector antibody. One hundred microlitre was added and the plates were incubated for 1 hour at 37°C followed by washing 5 times. One hundred microlitre of Streptavidin-horseradish peroxidase (Jackson Immunoresearch Lab, Inc.) diluted at 1/10,000 in blocking buffer was added as conjugate. After 1 hour incubation at 37°C and washing 5 times, 100 µl of ortho phenylenediamine (OPD) substrate, prepared by dissolving one tablet in 6 ml of distilled water and adding 2.5 µl of hydrogen peroxide, was added. The plates were incubated in the dark for 15 minutes at room temperature before stopping the reaction by adding 50 µl of sulphuric acid (4 N) to each well. The plates were read using an automated spectrophotometer at 490 nm with a reference of 655 nm. To determine the test result, the optical density (OD) of each stool sample was compared with the mean of a series of 8 reference Taenia negative stool samples plus 3 standard deviations (cutoff).
Presence of circulating cysticercus antigens was measured by the monoclonal antibody based B158/B60 Ag-ELISA (sero-Ag ELISA) [25], [26]. Sera from two known highly positive pigs (obtained from a local pig market and confirmed by dissection) were used as positive controls. The OD of each serum sample was compared with a sample of negative serum samples (N = 8) at a probability level of P = 0.001 to determine the result in the test [26].
Differentiation of the Taenia spp. was done using molecular methods. Taeniosis positive individuals were treated with niclosamide (2 g single dose) followed by a purgative (Magnesium sulphate, 30 g). The collected tapeworm segments were stored in 70% ethanol until use. DNA was extracted from the parasitic material using the Boom extraction method slightly modified as described by Rodriguez-Hidalgo et al. [27] and PCR used to amplify the mitochondrial 12 s rDNA gene fragment. Restriction fragment length polymorphism (RFLP) was then used to differentiate the Taenia spp. [27].
All collected data was entered into an Excel (Microsoft Office Excel 2007®) spreadsheet and analyses were conducted in Stata 10 (http://www.stata.com). Chi square test was used to check for differences between disease positivity and gender. Uni- and multivariate logistic regressions were used to investigate the relationship between taeniosis and cysticercosis positivity and individual gender and age. The age variable was first used as a continuous variable and then categorized into 10 categories of 10 years each, in order to identify changes in positivity frequencies as a function of the age of individuals. A change point analysis was used to simplify the observed relations into antigen patterns as a function of age [28], [29]. The level of significance was set at p<0.05 for all statistical analyses.
The geo-reference data collected was used for spatial analysis using ArcView Gis 3.2 (http://www.esri.com). Analysis of the possibility of geographical clustering of households with latrines or those that kept pigs and also cases of taeniosis and cysticercosis was done by means of the risk-adjusted nearest neighbour hierarchical spatial clustering (Rnnh) using Crimestat® III [30]. Given the limited number of individuals infected with taeniosis and cysticercosis, the minimum number of cases per cluster was set at 3 while the minimum number of households with a latrine or that kept pigs was set at 20. Monte Carlo simulations were run in this software to determine the significance of the clusters. Significance level of a cluster in the simulation was set at 95% and a cluster was determined significant if the density of the points was higher than that obtained at the 95th percentile after 1000 simulations.
A total of 720 willing participants from 20 villages belonging to 255 households participated in the study. Of these, 428 (59.4%) were females and 292 (40.6%) were males and the age ranged from 1 to 96 years with a median age of 12 years. The age distribution, with a majority of the younger age group, was typical of rural areas in developing countries [31]. The number of people living in a household ranged from 1 to 13 with a median of 7. Seven hundred and eighteen of the participants gave a stool sample and 708 gave a blood sample. At least one participant from each participating household gave a sample depending on the willingness of the household members. The number of individuals sampled from each household ranged from 1 to 11. Some household characteristics recorded from the questionnaire administered to the 255 households included; 32.6% kept pigs with 99.6% of these rearing on free-range, 47.8% of the households did not have latrines (Figure 2) and 94.5% of the households had at least one individual who consumed pork. Three clusters each of households with latrines (14.13881S, 31.19501E, density = 748.97; 14.14338S, 31.20369E, density = 151.95 and 14.09718S, 31.17940E, density = 117.15; 95th percentile density = 0.02) and those that kept pigs (14.13891S, 31.19493N, density = 299.89; 14.14390S, 31.20374E, density = 134.33 and 14.09773S, 31.17961E, density = 86.79; 95th percentile density = 0.01) were identified in the study area (Figure 2). About 44% of the households reported to have slaughtered a pig in their backyards. None of them had the meat inspected before either home consumption or resell to members of the community. Pork was reported to be consumed in a variety of ways including boiling, frying and roasting. The data obtained in the questionnaire on risk factors is described in more detail in another report (Mwape et al., submitted article).
The results for both the coproscopic examination and copro-Ag ELISA are shown in Table 1. Two (0.3%) individuals were positive for taeniosis on coproscopic examination while copro-Ag ELISA detected 45 (6.3%) positives. The two coproscopic positives were also positive on copro-Ag ELISA. Figure 3 shows the copro-Ag ELISA results in function of 10 age groups of 10 years each. The highest prevalence was determined in the 80–89 years age group, though this was not significantly different from the other age groups. A univariate logistic regression analysis did not indicate any relationship between copro-Ag ELISA positivity and sex (p = 0.548) or age (p = 0.311). This finding was the same for the multivariate analysis with p values of 0.139 and 0.645 for sex and age respectively. One cluster of taeniosis cases (14.13868S, 31.19509E, density 116.55; 95th percentile density = 49.33) was identified in the study community (Figure 4). At household level, the number of positives per household ranged from 0 to 3. All taeniosis positive individuals were treated with 2 g niclosamide per os and given a purgative (Magnesium sulphate) two hours later. One tapeworm was collected and confirmed to be T. solium by PCR-RFLP.
The results for the sero-Ag ELISA are shown in Figure 5 in function of 10 age groups of 10 years each. Circulating cysticercus antigens were detected in 41 (n = 708) participants giving an apparent prevalence of 5.8% (95% CI, 4.1–7.5). Uni- and multivariate logistic regression analysis revealed a very strong relationship between sero-Ag positivity and age (p<0.001). Figure 6 indicates that the prevalence of cysticercosis is initially low and a change point analysis indicated a significant increase in positivity frequencies at 30 years of age. The logistic regression model indicated that the proportion of sero-Ag ELISA positive individuals remains at a constant level until the age of 30, and from this age onwards a significantly higher level is observed (p<0.001).
A relationship was observed between copro-Ag positivity and sero-Ag positivity (p = 0.03) indicating that a copro-Ag positive individual was at an almost three-fold higher risk of being sero-Ag positive than the one who was not (OR = 2.9, p = 0.029).
There was no statistically significant difference in prevalence between males and females (χ2 = 0.034, p = 0.854). Two clusters of cysticercosis cases (14.14048S, 31.19692E, density 31.31; 14.08460S, 31.22085E, density 195.17; 95th percentile density = 23.16) were identified in the study community with the larger cluster spatially related to the taeniosis cluster (Figure 4).
Other intestinal parasites detected on coproscopic examination included hookworms, Schistosoma mansoni, Trichuris trichiuria and Ascaris lumbricoides. Table 2 shows the prevalence rates for these parasites with their respective 95% confidence intervals.
The objective of this study was to determine the prevalence of taeniosis and cysticercosis in a rural community in the Eastern Province of Zambia, where risk factors for the transmission of T. solium are present.
T. solium taeniosis tends to have a low prevalence, typically less than 1%, even in endemic communities [32], a higher prevalence is considered hyper-endemic [33]. In this study a prevalence of 6.3%, based on copro-Ag ELISA, was determined, indicating a hyper-endemicity in the study community. As in a number of other studies, no significant association between age/sex and taeniosis positivity could be determined [31]–[35].
Even though similar high taeniosis prevalences have been recorded in Kenya (4–10%) [17], the 6.3% prevalence determined in this study should be looked at critically. The sensitivity and specificity of the copro-Ag ELISA are estimated at 96–98% and 98–100%; respectively [5], [36]. However, the possibility of false positive test results due to cross-reactions with other pathogens present in the community should be considered. The assay has been reported not to cross-react with other parasite species including A. lumbricoides, T. trichiuria, Hymenolepis nana, H. diminuta or parasitic protozoa [36]. Also in our laboratory, stool samples with known H. nana, Schistosoma spp., T. trichiuria and A. lumbricoides infections were analyzed, and all results remained under the cut off level (Unpublished data). As the assay is not species specific [24], the possibility of the high taeniosis prevalence to be partially due to T. saginata infections cannot be ruled out. However, bovine cysticercosis in Zambia has so far only been reported in the Central and Southern provinces [37] and Western province (I.K. Phiri, personal communication).
Interviews with local people in the study area revealed that pig slaughter and pork consumption increases in the dry season as it is time for harvest and residents have then the means to buy either an entire pig or pieces of pork for home consumption. During this period, pig owners not only slaughter more pigs for the market but also for their own home consumption. Higher pork consumption could have led to new (still immature) taeniosis infections at the time of sampling, which will be detected by copro-Ag ELISA but not yet by coprology [5].
Only one tapeworm (from a participant positive on both copro-Ag and coproscopic examination) could be recovered after treatment of the 45 copro-Ag positive participants. The low recovery rate of tapeworms can be explained by: (1) stools were obtained only over one day and not over 3 days post treatment [38] due to logistical constraints, (2) usually after antiparasitic treatment, small and unrecognisable fragments are expelled by most patients [38] and these are easily missed, and (3) treatment of copro-Ag positive individuals was conducted over six months after collection of stool samples; natural expulsion may have occurred in this period.
The sero-Ag ELISA assay detected an apparent cysticercosis prevalence of 5.8% indicating the presence of viable cysts and as such active infections in these individuals. The prevalence of cysticercosis recorded in our study is comparable with that recorded in other endemic areas such as in the Andean region of Ecuador and in north Vietnam [39], [40], higher than in west Cameroon (0.4 to 4.0%) and southern Ecuador (2.3%) [41], [42] but lower than that reported in the Democratic Republic of Congo (21.6%) [14]. Other studies that have recorded higher seroprevalences include those that used antibody detection techniques such as in Mozambique (12.1%), South Africa (7.4%) and Peru (13.9%) [15], [16], [31]. However, antibody detection indicates exposure to the parasite and not necessarily established infection and hence is likely to detect more positives than the antigen detecting assay used in the current study [43].
Change point analysis of the association of antigen seropositivity and age revealed that the number of individuals in which circulating antigens were detected was significantly higher in people older than 30 years, indicating that viable cysts were more frequently present in individuals above this age. Studies have shown that a higher proportion of vesicular stage cysticercii is found in older (60 years and above) NCC patients [44], [45] and this has been attributed to immunosenescence since a weaker immunity in the elderly would facilitate the establishment and maintenance of viable cysticercii unlike in fully immunocompetent younger individuals [46]. The significant increase in sero-antigen positive individuals in the elderly was also observed in Ecuador where the number of positive individuals was higher in people order than 60 years [28]. However, in our study we see an increase already in young people (from 30 years onwards) who are supposed to be immunocompetent.
Establishment and development of infection is influenced by a range of complex factors; among which are parasite factors (e.g. parasite virulence/pathogenicity influenced by genetic differences, number, stage, location), host factors (e.g. age, gender, genetics influencing the immunological responses of the host when exposed to infection) and environmental factors (e.g. presence of risk factors, level of exposure, presence of other infections) [47]. It is very difficult to pinpoint exactly those factors present in the study area/population/age groups; that can explain this early increase in establishment of viable infection.
The high taeniosis prevalence recorded in the study community entails a possible very high exposure risk to infective eggs. In a study in India, higher infection rates (as indicated by sero Ag detection) were noted in areas with higher taeniosis prevalences [48]. Also in our study a significant positive relationship between copro-Ag positivity (presence of tapeworm) and sero-Ag positivity (cysticercosis) was established (logistic regression and cluster analysis). Level of exposure/infection with which the host is confronted can have an important effect on the immunological response of the host [44], and as such on the establishment of viable infection.
General factors that lower immunity in groups of individuals in the population could be at play making them more susceptible to infection. According to the United Nations Human Development Indices of 2008, about 64% of Zambia's population lived on less than $1 per day as compared to only 20% for Ecuador [49]. Poverty is an indication for poor nutritional status, which has an impact on the immune system [50]. Also the presence of other diseases such as HIV-AIDS, malaria and tuberculosis, other helminthic infections and physical environmental conditions [51] can influence greatly the host's reaction to other infections. In 2008, Zambia's HIV prevalence stood at 14.3% with the age group between 20 and 40 years being the most affected [52]. The country is also endemic for malaria [53] and helminthic infections are widely reported in rural areas, as reported in this study.
Genetic polymorphism of the parasite is another important determining factor for the establishment and development of infection. Nakao et al. [54] have described a cluster of isolates from Asia, and another cluster from isolates from Latin America and Africa. However, genetic differences within a cluster (within a continent/country/region) need to be evaluated as well. Several Zambian isolates are currently being examined, and preliminary results indicate a high genetic variability (Unpublished results, K. Kanobana), which might explain differences in development of infection between regions.
We have, in this study, shown that T. solium taeniosis and cysticercosis are present in the study community. Many issues remain unclear and obviously more work is required to understand the many factors that contribute to the transmission dynamics of the parasite and disease development in endemic rural areas. Also the economic impact and burden of disease in rural pig keeping communities of Zambia needs to be determined.
|
10.1371/journal.pntd.0005229 | Overexpression of a Mycobacterium ulcerans Ag85B-EsxH Fusion Protein in Recombinant BCG Improves Experimental Buruli Ulcer Vaccine Efficacy | Buruli ulcer (BU) vaccine design faces similar challenges to those observed during development of prophylactic tuberculosis treatments. Multiple BU vaccine candidates, based upon Mycobacterium bovis BCG, altered Mycobacterium ulcerans (MU) cells, recombinant MU DNA, or MU protein prime-boosts, have shown promise by conferring transient protection to mice against the pathology of MU challenge. Recently, we have shown that a recombinant BCG vaccine expressing MU-Ag85A (BCG MU-Ag85A) displayed the highest level of protection to date, by significantly extending the survival time of MU challenged mice compared to BCG vaccination alone. Here we describe the generation, immunogenicity testing, and evaluation of protection conferred by a recombinant BCG strain which overexpresses a fusion of two alternative MU antigens, Ag85B and the MU ortholog of tuberculosis TB10.4, EsxH. Vaccination with BCG MU-Ag85B-EsxH induces proliferation of Ag85 specific CD4+ T cells in greater numbers than BCG or BCG MU-Ag85A and produces IFNγ+ splenocytes responsive to whole MU and recombinant antigens. In addition, anti-Ag85A and Ag85B IgG humoral responses are significantly enhanced after administration of the fusion vaccine compared to BCG or BCG MU-Ag85A. Finally, mice challenged with MU following a single subcutaneous vaccination with BCG MU-Ag85B-EsxH display significantly less bacterial burden at 6 and 12 weeks post-infection, reduced histopathological tissue damage, and significantly longer survival times compared to vaccination with either BCG or BCG MU-Ag85A. These results further support the potential of BCG as a foundation for BU vaccine design, whereby discovery and recombinant expression of novel immunogenic antigens could lead to greater anti-MU efficacy using this highly safe and ubiquitous vaccine.
| Mycobacterium ulcerans (MU) infection causes a highly disfiguring, necrotic skin disease known as Buruli ulcer (BU). Antibiotic treatments have low efficacy if the infection is diagnosed after ulceration begins, leading to frequent dependence on surgical removal of infected tissues. A prophylactic vaccine for BU does not exist and several attempts to create an effective vaccine have shown limited success. We recently demonstrated that a recombinant strain of M. bovis BCG expressing the immunodominant MU-Ag85A conferred significantly enhanced protection against experimental BU compared to the standard BCG vaccine. Here we show that BCG expression of a fusion between two alternative MU antigens, Ag85B and EsxH, can promote antigen-specific T cell and humoral immune response capable of significantly improving survival and protection against BU pathology, compared to BCG MU-Ag85A alone. These results support the potential for using the highly safe and ubiquitous BCG vaccine as a platform for further BU vaccine development.
| Subcutaneous skin infection by Mycobacterium ulcerans (MU) leads to a potentially disfiguring, necrotic condition known as Buruli ulcer (BU) [1]. What often begins as an indolent skin nodule or small lesion can ultimately progress to expanding ulcerations, body-wide scarring, loss of limbs or eyes, and osteomyelitis [2]. These infections disproportionately affect children and are largely endemic to Sub-Saharan Africa, Australia, and Japan, where the unconfirmed mode of transmission is thought to be dependent on exposure to contaminated wetland areas and insect vectors [3, 4]. Treatment regimens include lengthy combination anti-mycobacterial therapies, however, lack of medical access, absence of rapid and accurate diagnostics, and the often misleading symptoms of BU frequently lead to significant delays in therapeutic action [5, 6]. At the point of extensive tissue damage, surgical debridement and skin grafting is required, resulting in significant morbidity and social stigmatization [7, 8]. Antibiotics can be effective against MU if administered at an early time point prior to ulceration, and side effects of treatment can include nephrotoxicity and hearing loss [9]. While there is increasing promise for less toxic antibiotic therapies, currently no prophylactic vaccine is available to prevent BU in the areas with greatest prevalence [10].
BU vaccine research strategies have largely focused on prime-boost regimens using recombinant DNA and MU proteins, however, the efficacy of these approaches has not surpassed the transient, cross-reactive protection observed during experimental vaccination with tuberculosis vaccine strain, Mycobacterium bovis bacillus Calmette- Guérin (BCG) [11–18]. BCG, the most ubiquitous World Health Organization-approved vaccine administered across the world, possesses a promising safety profile but low efficacy against pulmonary tuberculosis afflicting millions of people [19, 20]. Experimental BCG vaccination has been studied using BU animal models and has been shown to confer protection by delaying ulceration after murine footpad challenge with MU [11]. While BCG vaccination extends the time to appreciable footpad swelling, protection ultimately wanes and animal euthanasia is required. Retrospective studies in humans also provide support for the potential use of BCG as a foundation for an effective BU vaccine. Patients previously vaccinated with BCG were shown to have delayed onset to ulceration after infection with MU, as well as significant protection against developing complications of MU infection, such as osteomyelitis [21–23]. These lines of evidence further support the potential of BCG as a foundation for BU vaccine design, whereby improvement of BCG immunogenicity could lead to greater efficacy using this highly safe and ubiquitous vaccine.
BCG has previously been engineered to express various recombinant immunogenic antigens and protein fusions for use in TB vaccine development, with numerous observed in vivo effects [24–26]. Recombinant BCG vaccine strains which have been engineered to overexpress major antigenic secretory proteins, such as ESAT-6, TB10.4, CFP10, heat shock proteins, and members of the mycolyl transferase complex Ag85A, Ag85B, and Ag85C, have displayed the greatest promise by increasing both humoral IgG antibody production and CD4 mediated Th1 responses against M. tuberculosis challenge [27–33]. Similar strategies have been investigated in application to BU vaccine design as well, with varying degrees of success. Priming with a DNA-based vaccine encoding multiple MU polyketide synthase modules and boosting with recombinant protein by Roupie et al. yielded differential levels of antigen-specific IgG responses, as well as IFNγ and IL-2 secretion upon recombinant MU antigen stimulation [12]. However, no improvement in protection over the level conferred by BCG vaccination was observed. Alternatively, an investigation by Tanghe et al. used a similar strategy that employed the plasmid-based expression of MU Ag85A as a prime followed by a recombinant protein boost [13, 14]. This vaccination regimen produced appreciable antigen-specific immunogenicity which correlated with a level of protection against MU challenge that was similar to that achieved by BCG vaccination alone.
We recently utilized a combination of strategies from the TB vaccine field, as well as those used by previous attempts to design anti-MU vaccines by engineering a quality controlled recombinant strain of BCG that overexpressed MU-Ag85A [15]. In this study, we showed that not only could this vaccine strain significantly induce proliferation of antigen-specific CD4+ T cells and increase IFNγ+ Th1 splenocytes responsive to whole MU and subcellular fractions, but subcutaneous priming also decreased MU burden, protected against mycolactone-induced pathology, and extended the lifespan of MU-challenged mice to significantly greater levels compared with BCG vaccination alone. Knowing that overexpression of one MU antigen by BCG could have these effects, we were subsequently interested in determining if alternative antigens or combinations of antigens could yield improvements on vaccine immunogenicity and efficacy.
The immunodominant antigens, Ag85B and TB10.4, are two such antigens that have been successfully used to augment the protective qualities of BCG against experimental tuberculosis in animal models [29, 31, 32, 34–37]. These individual antigens, as well as fusion proteins combining various small antigens or important T cell epitopes from multiple antigens, have also been successfully expressed heterologously in BCG. Importantly, these constructs have been shown to initiate production of antigen-specific CD4+ T cell populations known to be vital in generating the same anti-mycobacterial Th1 responses hypothesized to play a role in containment of MU in humans. Due to the encouraging results from our previous study involving BCG expression of Ag85A and the large body of evidence supporting the usefulness of Ag85B and TB10.4 antigens against other mycobacterial diseases, we generated a vaccine strain of BCG which expressed a fusion protein combining MU-Ag85B and the TB10.4 homolog from M. ulcerans, MU-EsxH (BCG MU-Ag85B-EsxH).
Here we will show that, compared to BCG vaccination, a single subcutaneous dose of BCG MU-Ag85B-EsxH induced significantly enhanced antigen-specific humoral responses, CD4+ T cell proliferation, and Th1 splenocyte responses in mice. In addition, a single, un-boosted, subcutaneous dose of BCG MU-Ag85B-EsxH conferred significantly greater protection compared to BCG by reducing bacterial burden in MU challenged footpads, resisting pathologies associated with MU infection, and significantly lengthened the lifespan of MU-challenged mice. Importantly, these effects were statistically improved over those conferred by BCG MU-Ag85A, which was the first and only vaccine strain superior to BCG vaccination against BU in mice.
Female, 6–8 week old, C57BL/6 mice were obtained from Jackson Laboratories. These mice were 14–16 weeks old by time of footpad challenge with MU. Animal work was approved by the Duke University Institutional Animal Care and Use Committee (IACUCU protocol A065-13-03). IACUC protocols performed at Duke University adhered to the AAALAC, USDA, Guide for Care and Use of Laboratory Animals and Public Health Service Policy on Humane Care and Use of Laboratory Animals and Animal Welfare Act.
All strains of Mycobacterium bovis BCG-Danish (BCGD) were cultured at 37°C on solid Difco Middlebrook 7H10 agar or in liquid Difco Middlebrook 7H9 media supplemented with 0.5% glycerol, oleic-albumin-dextrose-catalase (OADC), and 0.05% tyloxapol. Selection of BCG transformants expressing MU Ag85B-EsxH was accomplished by adding 25 or 50 μg/ml hygromycin to liquid or solid media, respectively. Liquid cultures consisting of volumes less than 50 ml were shaken at 120 rpm and larger volumes were expanded to 250 ml or less in one liter bottles rotated at 6 rpm. High-volume vaccine accession lots were aliquoted into 1 ml cryovials and frozen at a concentration of OD600 1 (~108 CFU/ml). Virulent M. ulcerans 1615 was kindly provided by Dr. Pamela Small (University of Tennessee) and was cultured at 32°C in Middlebrook media as prepared for BCG. For purification of plasmid DNA and sequencing, DH5α Escherichia coli (E. coli) was grown on lysogeny broth (LB) agar plates or in LB supplemented with 50 μg/ml hygromycin.
The M. ulcerans Ag85B open reading frame and endogenous secretion signal were amplified from MU1615 genomic DNA and cloned into the mycobacterial vector pMV261 [38], where antigen expression was mediated by the constitutive mycobacterial hsp60 promoter (henceforth, pSL402). Selection of plasmids was controlled through hygromycin resistance. The influenza hemagglutinin (HA) epitope was added to the C-terminus of the EsxH fusion. Electrocompetent BCG cells were prepared by centrifuging log phase culture (OD600 0.6–0.8) at 3000 rpm for 10 minutes, followed by washing in a buffer of 10% glycerol and 0.05% tyloxapol. Mycobacteria electroporated with 0.5 μg plasmid DNA were recovered by shaking at 37°C in 1 ml Middlebrook 7H9 media overnight.
A series of quality control characterizations was performed on vaccine accession lots by assessing expression of recombinant antigen, contamination, and plasmid DNA retention as previously described [39]. Bacterial lysates for immunoblot were prepared by centrifuging 10 ml of log-phase liquid culture at 3000 rpm for five minutes. After washing in phosphate buffer saline + 0.05% tyloxapol (PBST), the final cell pellet was resuspended in 200 μl lysis buffer with glass beads and vortexed for three minutes. Lysates were clarified by collecting the supernatant after centrifugation for five minutes at 3000 rpm. Pre-cast SDS PAGE gels were loaded with 15 μl clarified lysate boiled in Laemmli buffer and run for one hour at 130V. Protein was then transferred to PVDF membranes for one hour at 30V. Blocking of membranes was performed by shaking in 5% fat free milk in TBS with 0.1% tween (TBST) for one hour at room temperature. For detection of the HA epitope, mouse anti-HA-HRP (clone 3F10, Roche) was diluted in 5% milk-TBST (1:1000) and incubated room temperature for one hour. After washing in TBST, detection of proteins was performed using chemiluminescence (Lumi-light, Roche). Plasmid DNA was purified using a modified Qiagen Miniprep protocol [39]. Isolated plasmids were heat-shocked into E. coli DH5α and plasmid DNA was re-purified from the resulting transformants. Correctly sized plasmid inserts were assessed by reaction with NdeI/EcoRV and analysis using gel electrophoresis. Plasmid inserts from 10 E. coli clones were sequenced and analyzed using Clone Manager software (Sci-Ed). Vaccine contamination was assessed by spread plating 100 μl of thawed accession lot material on chocolate agar and examining for growth following a two week incubation at 37°C.
For protection studies, C57BL/6 mice were subcutaneously vaccinated by injection of 100 μl (~107 cells) from an accession lot vial of empty-vector BCG (pHA), recombinant BCG MU-Ag85A, or BCG MU-Ag85B-EsxH into the scruff of the neck. Eight weeks post-vaccination, the mice were challenged intradermally with 105 M. ulcerans 1615 (MU1615) via the footpad. MU1615 challenge inocula were consistently accessed from the same accession lot frozen at -80°C and were tested for virulence by pathology in mouse footpad models. The width and height of footpad swelling from infected mice were measured at two to three week intervals using digital calipers. To reduce animal suffering and in compliance with IACUC protocol, infected mice were euthanized once footpad swelling height exceeded 4.5 mm, prior to any visible ulceration.
Quantitation of Ag85-specific CD4+ T cell levels was performed by flow cytometric analysis of MHCII tetramer staining. Mice were inoculated subcutaneously or intravenously (via the retro-orbital route) with 100 μl (~107 cells) of a freshly thawed vaccine accession lot vial. Weekly blood samples were collected retro-orbitally followed by isolation of peripheral blood mononuclear cells (PBMCs) by gradient centrifugation using Lympholyte M (Cedarlane). PBMC buffy coats were collected, washed in 10 ml phosphate buffered saline (PBS), and were resuspended in 2 ml of ACK lysis buffer to remove erythrocytes. After centrifugation at 2000 rpm, PBMC pellets were washed with PBS and stained with APC-conjugated M. tuberculosis Ag85B-MHCII tetramer (1:500, NIH Tetramer Core Facility) in flow buffer (2% fetal bovine serum in PBS). The 15 amino acid epitope, FQDAYNAAGGHNAVF, was recognized by this tetramer and shares high sequence homology with MU Ag85. PBMCs were stained with tetramer for 30 minutes at 37°C and then stained with FITC-conjugated anti-mouse CD4 (1:500, clone GK-1.5, Biolegend) and PE-Cy5 anti-mouse CD8 (1:200, clone 53–6.7, Biolegend) for 30 minutes on ice. The PBMCs were then washed in with flow buffer, centrifuged for 2000 rpm for five minutes, and were then resuspended in 4% paraformaldehyde. Following fixation cells underwent flow cytometric analysis using a Becton Dickinson (BD) LSRII and FlowJo software (Tree Star Inc.). To stain effector and central memory cells, the tetramer protocol was followed by additional antibody incubation steps with APC-Cy7 anti-mouse CD4 (clone GK-1.5), FITC anti-mouse CD62L (clone MEL-14, BD Pharmingen), and PE-Cy7 anti-mouse CD44 (clone IM7, BD Pharmingen).
MU bacilli present in infected footpads were quantified using a method previously described [15]. At 6 and 12 weeks post-challenge, footpads infected with MU1615 were removed from euthanized mice and for disinfection and dissection. The footpads were disinfected by five-minute contact with 70% ethanol, followed by three washes in PBST. The challenged footpads were then minced and crushed by mortar and pestle. Resulting homogenates were subjected to N-acetyl-L-cysteine (NALC)/NaOH treatment by adding a mixture (50:50) of 4% NaOH and 2.9% sodium citrate + 1% NALC for 20 minutes. Homogenates were pelleted at 3000 rpm for five minutes, washed with PBST, and larger particulates were removed by passage through a 40 μm mesh. 5 μl of filtrate was then evenly distributed within a 0.8 cm2 circle upon glass microscope slides. Heat-fixed smears were stained with auramine-rhodamine (BD Biosciences), destained with acid alcohol, and counterstained with potassium permanganate. Stained slides were viewed under 100x oil immersion using a Nikon X microscope. Acid-fast bacilli (AFB) were counted in four random fields of view (FOV) per animal (16 images total per vaccinated group). Total AFB calculations were performed by multiplying cell counts by the number of 0.038 mm2 FOVs in each marked smear area per microliter of applied filtrate.
ELISPOT plates (PVDF, 96 well) were equilibrated with 70% ethanol, washed with PBS, and were coated overnight with anti-mouse IFNγ antibody (1 μg/ml, clone AN18, Mabtech). Necropsies were performed on mice eight weeks following intravenous vaccination to harvest splenocytes for growth in RPMI complete media (RPMI with L-glutamine and 10% fetal bovine serum). 96 well plates were blocked using RPMI media and 6 x 105 splenocytes were combined in each well with varied agonists: MU-Ag85A peptide (100 μg/ml, FQAAYNAAGGHNAVWNFDDN), MU-Ag85B peptide (100 μg/ml, FQDAYNAAGGHNAVFNFNDN), heat killed MU (HKMU, 1 mg/ml), or whole cell lysate prepared from log phase MU liquid culture. Splenocytes were stimulated for 16 hours at 37°C. The plates were then washed with PBS + 0.05% tween 20 followed by a two hour incubation with secondary anti-mouse IFNγ antibody (1:1000, clone R46A2, Mabtech) at 37°C. After a wash and three hour room temperature incubation with VectaStain avidin peroxidase complex (Vector Labs), plates were incubated with 3-amino-9-ethylcarbazole substrate for five minutes. The reaction was ceased by submersion of plates in deionized water. Spots were visualized and quantified using a CTL Immunospot plate reader.
High-binding, 384-well plates (Corning) were coated overnight at 4°C with 100 ng/ml recombinant M. tuberculosis Ag85A, Ag85B, or purified Ag85 complex (BEI resources, NIAID, NIH) diluted in 0.1 M sodium bicarbonate. Wells were then washed once, blocked with 40 μl blocking buffer (4% whey protein, 15% goat serum, 0.5% tween 20, and 0.05% sodium azide in PBS) for one hour at room temperature, and then washed again. After blocking, of a 1:100 dilution of intravenously vaccinated mouse serum in blocking buffer was added for 2 hours at room temperature. Plates were washed four more times and 15 μl of a 1:1000 dilution of goat anti-mouse IgG (Southern Biotech 1030–05) was added for one hour at room temperature. After four further washes, 20 μl of tetramethylbenzidine substrate was added per well for up to 15 minutes. The reaction was stopped by addition of 20 μl 0.33 N HCL solution and absorbance was read at 450 nm.
We have previously described the generation of quality controlled accession lots of recombinant BCG expressing antigens of interest, particularly MU-Ag85A [39]. In order to generate a vaccine strain of BCG which expressed an in-frame fusion between MU-Ag85B and the M. tuberculosis TB10.4 homolog, MU-EsxH, electrocompetent BCG was transformed with pSL402 (Fig 1A). The pSL402 replicating plasmid controls transcription of MU-Ag85B-EsxH containing a C-terminal fusion to the influenza hemagglutinin epitope using the constitutive mycobacterial hsp60 promoter. Plasmid replication was regulated by the mycobacterial oriM and by oriE in E. coli shuttle strains. Selection of bacterial transformants utilized plasmid-encoded resistance to hygromycin. Fig 1B displays expression patterns of the MU-Ag85B-EsxH fusion protein in whole cell lysates from 3 randomly picked BCG transformants. A single transformant which strongly expressed the fusion antigen was selected to produce a large-volume vaccine accession lot upon which a quality control panel was employed to further characterize expression of recombinant antigen, purity of the vaccine lot, and integrity of the recombinant plasmid sequence [39].
Anti-mycobacterial immunity is largely governed by responses from CD4+ T helper cells [40, 41]. In order to determine if antigen-specific adaptive immune responses could be generated by vaccination with BCG-MU-Ag85B-EsxH, C57BL/6 mice were either subcutaneously or intravenously or primed with 107 bacilli from thawed quality controlled vaccine lots prepared as previously described [39]. At weekly intervals, retro-orbital blood samples were collected to isolate peripheral blood mononuclear cells. Flow cytometric analysis of staining by MHCII tetramer was subsequently used to quantify the percentage of CD4+ T cells which recognized the Ag85 epitope, FQDAYNAAGGHNAVF.
Fig 2A and 2B show the levels of antigen-specific T helper cells induced by intravenous or subcutaneous vaccination, respectively. Responses from BCG MU-Ag85B-EsxH were compared to those from mice vaccinated with BCG containing an empty expression vector, BCG MU-Ag85A, and unprimed mice. While low levels of Ag85-specific T cells were induced in response to endogenously expressed antigen in BCG, a significantly greater number of T cells was produced upon vaccination with BCG overexpressing Ag85A or Ag85B-EsxH. The vaccination route did differentially affect the speed and amplitude with which peak responses were reached. Intravenous inoculation induced quicker, larger, and more prolonged T cell responses compared to the subcutaneous route; an 8% Ag85-specific T helper cell population was reached at 3 weeks after the intravenous injection with BCG MU-Ag85B-EsxH compared to a 1.75% peak response post-subcutaneous vaccination. Greater statistical significance was also achieved between BCG Ag85B-EsxH versus empty-vector BCG compared to the enhancement of BCG MU-Ag85A responses during the intravenous injection at weeks 2 and 3 post-vaccination (p<0.05, and p<0.001, respectively). Interestingly, the peak response to BCG MU-Ag85A reached a maximum of 5% compared to the 8% peak following vaccination with BCG MU-Ag85B-EsxH. However, subcutaneous vaccination resulted in highly similar T cell proliferative responses between BCG MU-Ag85A and BCG MU-Ag85B-EsxH, both of which were significantly higher than those induced by empty-vector BCG (p<0.01, p<0.002).
Additionally, we attempted to use MHCII tetramer staining to detect T cell populations capable of recognizing the M. tuberculosis TB10.4 epitope, SSTHEANTMAMMARDT (data not shown). However, no tetramer positive populations were detected, which may be due to the presence of two amino acid substitutions at this tetramer peptide position within the MU-EsxH sequence. Together, these data suggest that like BCG MU-Ag85A, BCG MU-Ag85B-EsxH is capable of generating helper T cell populations responsive to Ag85, an immunodominant antigen previously demonstrated to play a role in vaccine-mediated protection against MU.
Evidence from previous studies has highlighted the importance of CD4+ memory T cell populations in establishing greater efficacy for mycobacterial vaccines [42]. To quantify the levels of memory T cells produced by vaccination with the recombinant BCG strains, C57BL/6 mice were intravenously primed with 107 bacilli and, four weeks later, peripheral lymphocyte staining was performed for CD4 and the CD62L and CD44 memory T cell markers. While the absolute numbers of CD4+ T cells rose in all vaccinated groups regardless of BCG strain, vaccination with BCG MU-Ag85B-EsxH induced significantly higher levels compared to BCG (Fig 3A, p<0.05). Of those populations, naïve T cells were significantly reduced upon vaccination with either BCG MU-Ag85A or BCG MU-Ag85B-EsxH (Fig 3B, p<0.05). Interestingly, both Ag85-specific CD4+ effector memory and central memory T cell populations were significantly higher upon vaccination with either BCG MU-Ag85A or BCG MU-Ag85B-EsxH compared to BCG alone (Fig 3C and 3D, p<0.05). These data suggest that BCG MU-Ag85B-EsxH could also be useful in establishing T cell memory reservoirs capable of recognizing an MU antigen known to be immunoprotective and possibly representing sources of anti-mycobacterial IFNγ and IL-2 [43].
The requirement of IFNγ-yielding Th1 responses for generating efficacious anti-mycobacterial immunity has been characterized by several studies [44]. To determine if production of antigen-specific T cells following vaccination with recombinant BCG strains could generate such responses, C57BL/6 mice were primed with empty-vector BCG, BCG MU-Ag85A, or BCG MU-Ag85B-EsxH. Eight weeks post-vaccination, mice were euthanized and harvested splenocytes were stimulated with various MU antigens: MU-Ag85A peptide, MU-Ag85B peptide, heat-killed MU1615 (HKMU), or MU whole cell lysate. Quantification of IFNγ-producing splenocytes was performed by enzyme-linked immunospot (ELISPOT) and the numbers of IFNγ+ spot-forming units (SFU) detected after 24 hours of agonist stimulation were calculated (Fig 4).
Vaccination with both BCG MU-Ag85A and BCG MU-Ag85B-EsxH yielded significantly increased IFNγ+ splenocytes compared to BCG alone when stimulated with all MU antigens (p<0.05). The greatest responses were detected upon stimulation with whole heat-killed MU, whereby priming with BCG MU-Ag85A and BCG MU-Ag85B-EsxH increased SFU over BCG pHA by 2.8-fold and 3.3-fold, respectively. Notably, cytokine-secreting splenocyte numbers trended higher in the BCG MU-Ag85B-EsxH-vaccinated mice compared to BCG MU-Ag85A following stimulation with MU-Ag85B peptide (>2-fold increase) and whole heat killed MU (1.2-fold increase). Four separate M. tuberculosis TB10.4 peptides were also used to stimulate splenocytes, but no appreciable IFNγ+ populations were detected. Interestingly, the response of BCG MU-Ag85B-EsxH vaccinated mice to HKMU was over double that of the SFU generated by stimulation with MU-Ag85B peptide alone. These data suggest that the BCG MU-Ag85-EsxH vaccine may increase responsiveness of functional Th1 cells to additional Ag85B peptides or to other antigens expressed by M. ulcerans cells.
Humoral responses to mycobacterial infection are becoming increasingly recognized in adaptive defense and as potential therapeutics [45–47]. To determine if use of recombinant BCG could induce antigen specific antibody responses in vivo, C57Bl/6 mice were vaccinated with empty-vector BCG, BCG MU-Ag85A, or BCG MU-Ag85B-EsxH and peripheral blood was collected biweekly for 6 weeks. Sera were subsequently isolated and tested for antibodies specific to immunogenic Ag85 proteins by IgG ELISA. Fig 5A, 5B and 5C display time course antibody responses to recombinant M. tuberculosis Ag85A, recombinant Ag85B, and purified Ag85 complex, respectively. At 2 weeks post-vaccination, all IgG responses were low in mice vaccinated with BCG or BCG MU-Ag85A; however, BCG MU-Ag85B-EsxH induced high levels of anti-Ag85A, Ag85B, and Ag85 complex IgG. Over the course of 6 weeks, antibody kinetics varied depending on antigen and vaccine group. Antibodies produced by BCG alone were consistently lower than those elicited by recombinant strains but did increase over time. Ag85A IgG responses from BCG MU-Ag85A-vaccinated mice continued to rise over time compared to the BCG MU-Ag85B-EsxH response which began high and slightly decreased over 6 weeks. A similar trend was observed for BCG MU-Ag85A induced anti-Ag85B responses, however, BCG MU-Ag85B-EsxH vaccination yielded a very high response which did not decline over this time course. Finally, IgG antibody induction against purified Ag85 complex peaked for BCG MU-Ag85A-vaccinated mice at 4 weeks but began to decline by 6 weeks. Conversely, anti-Ag85 complex responses induced by BCG MU-Ag85B-EsxH immediately began high and continued to climb by week 6. Compared to empty-vector BCG, BCG MU-Ag85A vaccination did not generate statistically significantly higher IgG responses for any antigens except to rAg85A at week 6. However, BCG MU-Ag85B antibody induction was statistically significantly higher (p<0.05) over empty-vector BCG for both rAg85A and rAg85B for all time points except anti-rAg85A at week 6. Additionally, recombinant M. tuberculosis TB10.4 was plated in the same ELISA format, however no IgG could be detected by ELISA. Together these data highlight the ability of BCG MU-Ag85B-EsxH to induce high levels of IgG reactive to multiple immunogenic Ag85 antigens, with a more rapid initial response compared to BCG or BCG MU-Ag85A.
We previously demonstrated that priming with BCG MU-Ag85A could significantly extend the survival time of MU challenged mice compared to BCG vaccination alone [15]. Upon characterizing the immunogenic properties associated with BCG MU-Ag85B-EsxH, some of which displayed enhancement over BCG MU-Ag85A, we were further interested in determining the protection profiles in similarly challenged mice. C57Bl/6 mice were subcutaneously primed with 107 empty-vector BCG, BCG MU-Ag85A, or BCG MU-Ag85B-EsxH and, 8 weeks later, were intradermally challenged with 105 virulent MU1615 via the footpad. The width and height of challenged footpads were measured with digital calipers during the period of infection. If footpad swelling surpassed 4.5 mm in height, mice were euthanized to reduce suffering. Fig 6A shows the time to euthanasia for unprimed and vaccinated mice. As previously demonstrated, while BCG vaccination increased the mean survival time from 6.3 weeks for unprimed mice to 8 weeks, subcutaneous vaccination with BCG MU-A85gA significantly increased survival time over BCG alone to 17.4 weeks (p<0.01). Markedly however, a single subcutaneous dose of BCG MU-Ag85B-EsxH further significantly increased the survival time of MU-challenged mice over that of BCG MU-Ag85A to a mean of 29.4 weeks (p<0.001).
Previous BU studies in mice have demonstrated a correlation between the degree to which infected footpads swell and the MU bacterial load present within challenged tissue [15, 48]. To determine if the observed enhancement of survival associated with BCG MU-Ag85B-EsxH vaccination correlated with a reduction in MU burden, mice which had received vaccinations and challenged as above were euthanized for isolation of MU1615. Infected footpads were dissected at 6 and 12 weeks post-challenge and persisting acid-fast MU in filtered footpad homogenates were stained with fluorescent auramine-rhodamine. The evaluation of bacterial load using microscopy was previously assessed to confirm similar colony forming unit (CFU) results were achieved compared to plate counting [15]. Fig 6B shows the mean acid-fast burden for unprimed mice or those primed with empty-vector BCG, BCG MU-Ag85A, or BCG MU-Ag85B-EsxH. At both 6 and 12 weeks post-challenge, all subcutaneous vaccinations resulted in a significant reduction of footpad bacterial burden compared to unprimed mice. However, priming with BCG MU-Ag85B-EsxH consistently achieved the greatest protection at both time points, conferring a significantly greater reduction in footpad bacterial replication by 1.5 log and 2.56 log at 6 and 12 weeks post-infection, respectively (p<0.001, p<0.05). Importantly, protection conferred by BCG MU-Ag85B-EsxH was also superior when compared to burdens present in empty-vector BCG (a 0.78 log and 1.7 log reduction at 6 and 12 weeks, respectively) and BCG MU-Ag85A-vaccinated groups (0.84 log reduction at 12 weeks post-challenge). This potent inhibition of bacterial growth by BCG MU-Ag85B-EsxH vaccination, as well as the intermediate inhibition by BCG MU-Ag85A correlated well with the distinct ability of each vaccine to extend survival times in the footpad challenge model.
Production of the cytotoxin, mycolactone, is known to contribute to the tissue destruction histologically observed at the foci of MU infection [49, 50]. To determine if the reduction of in vivo bacterial burden by vaccination correlated with protection against tissue damage, footpads were collected for histopathological analysis 12 weeks post-infection. Fig 6C displays representative images of hematoxylin and eosin (H&E) stained tissue sections from unprimed, empty-vector BCG, BCG MU-Ag85A, or BCG MU-Ag85B-EsxH primed mice. Consistent losses of epidermal layers, as well as extensive areas of internal necrosis and infiltrates of inflammatory cells were observed in footpads of unprimed mice. While BCG vaccination had a reduction in epidermal loss, substantial edema replaced necrotic lesions found in unprimed animals. However these features were rare in BCG MU-Ag85A and BCG MU-Ag85B-EsxH primed mice where, at 12 weeks post-infection, full tissue integrity remained in most animals. To visualize the organization of persistent MU bacilli in vivo, Ziehl-Neelsen (ZN) staining was performed on tissue sections from challenged footpads. Fig 6D displays ZN staining of the contiguous tissue sections of the above-mentioned H&E tissue sections. Large masses of pink acid-fast bacilli (AFB) present in the extracellular milieu as well as granulomatous lesions could readily be observed in tissue sections from unprimed MU-challenged mice (Fig 6D and 6E), while fewer and smaller groups of AFB were detected in empty-vector BCG groups (Fig 6E). However, AFB could not readily be observed in footpad sections from the BCG MU-Ag85 and BCG MU-Ag85B-EsxH vaccinated groups, correlating with the lower overall bacterial burden present in these tissues. Overall these data suggest that the protective BCG MU-Ag85B-EsxH immune responses, characterized by enhanced proliferation of antigen-specific Th1 CD4+ T cell populations and potent antigen-specific humoral IgG induction, can contribute to a reduction in bacterial burden, inhibition of tissue destruction, and an overall greater lifespan for MU1615-infected mice, compared to BCG and BCG MU-Ag85A.
Buruli ulcer is a neglected tropical disease whose persistence continues to inflict severe patient morbidity in the absence of rapid diagnosis and treatment. Today’s diagnostics are limited to molecular techniques, microbial culture, or histopathological analyses not readily available in the most afflicted regions. Currently, the standard of care involves lengthy medical regimens, including rifampin and streptomycin treatment, both of which may confer issues of toxicity. Generation of a prophylactic vaccine against M. ulcerans infection could be incredibly invaluable, especially due to the disproportionately high incidence of BU in children, the threat of long-term disfigurement, and associated social stigma. Despite many efforts to develop an efficacious BU vaccine, new candidates have either not displayed immunity or have conferred limited and short-lived protection against experimental MU infection.
However, we recently have shown that a single subcutaneous dose of quality controlled BCG-based vaccine overexpressing the MU mycolyl transferase antigen 85A could significantly decrease bacterial burden, pathology, and increase survival time following MU1615 footpad challenge [15]. These protective effects were significantly greater than those conferred by the previously most protective vaccine strain to date, M. bovis BCG. Interestingly, a subsequent study highlighted MU-Ag85A as conferring protection when expressed in the M. marinum genetic background as well [51]. With the successful progress of recombinant BCG and M. marinum expressing MU-Ag85A as BU vaccine candidates, we decided to further improve a candidate vaccine strain by expressing other known immunodominant antigens. To this end, a novel recombinant BCG strain expressing the MU-Ag85B-EsxH fusion protein was generated, frozen into quality controlled accession lots, and evaluated for immunogenicity and protection against experimental BU in the mouse model.
Several aspects of immunogenicity were investigated in the present study, including the induction of MU antigen-specific CD4+ T cell and memory populations, the production of IFNγ-secreting cells responsive to stimulation by whole MU and cellular components, and the production of the antigen-specific humoral responses. Interestingly, both BCG MU-Ag85A and BCG MU-Ag85B-EsxH were equally more significantly immunogenic compared to BCG when examining adaptive T cell responses and memory populations. In contrast, BCG MU-Ag85B-EsxH vaccination was capable of inducing significantly greater IgG responses to both rAg85A and rAg85B compared to empty-vector BCG at multiple time points post-vaccination, while inoculation with BCG MU-Ag85A was not. Although the importance of humoral responses for immunity against MU is not well characterized, antibody mediated protection might be of major relevance against advanced stages of MU infection, where bacilli are predominantly found as extracellular clusters. Furthermore, previous studies have demonstrated the potential efficacy of antibody-based therapies against experimental M. tuberculosis infection [45–47].
In addition to an evaluation of immunogenicity for the BCG MU-Ag85B-EsxH vaccine, we were interested in characterizing any improvement in protection during in vivo MU challenge compared to empty-vector BCG or BCG MU-Ag85A. Vaccination with a single subcutaneous dose of BCG MU-Ag85B-EsxH significantly increased the lifespan of infected mice over that of unprimed mice or empty-vector BCG by 4.7-fold and 3.7-fold, respectively. Strikingly, expression of Ag85B-EsxH in BCG also conferred significantly greater protection compared to BCG MU-Ag85A by increasing the mean survival time 1.7-fold compared to this strain. This result was also associated with a significant decrease in MU burden at 6 and 12 weeks post-challenge, with BCG MU-Ag85B-EsxH reducing bacterial burden compared to empty-vector BCG and BCG MU-Ag85A by 55-fold and 7-fold at 12 weeks, respectively.
Of note, we were unable to demonstrate antigen-specific immune responses to the MU-EsxH protein. Proper molecular mass and plasmid sequencing of the fusion protein were confirmed in the BCG MU-Ag85B-EsxH accession lot according to our previously published quality control protocol, suggesting that the assays chosen were either not sensitive enough to detect EsxH immune responses or could not detect responses to EsxH as a fusion to Ag85B. Alternatively, the immunogenicity reagents used may have not represented a high enough homology to the MU-EsxH protein sequence. The TB10.4 MHCII tetramer, TB10.4 ELISPOT peptides, as well as recombinant TB10.4 protein used in the IgG ELISA were of M. tuberculosis origin, which shares 84% amino acid sequence identity with MU-EsxH. These reagents were chosen because of their ready availability from BEI resources; however, further characterization of antigen-specific immunity to the TB10.4 homolog will require the use of EsxH-specific reagents. This does insert a degree of complexity when determining the degree of contribution to protection each antigen provided, or if either antigen alone could singularly be responsible for the increase in protection observed in the present study. This would require a comparative analysis with BCG MU-Ag85B or BCG-EsxH, which is planned for future investigations.
Previous vaccine studies which utilized Ag85 members and TB10.4 in tuberculosis models may shed light on their relative contributions towards immunity against MU. A recombinant strain of BCG which overexpressed the M. tuberculosis Ag85B was previously shown by Horwitz et al. to significantly reduce bacterial burdens in both the lungs and spleens of challenged guinea pigs compared to either recombinant Ag85B protein or BCG vaccination alone ([31, 37]). Mu et al. generated recombinant adenovirus vaccines which expressed M. tuberculosis Ag85A or the fusion between Ag85A and TB10.4. Following vaccination and subsequent challenge with H37Rv, bacterial burden in the lungs of the antigen fusion strain was significantly reduced compared to the monovalent adenovirus-Ag85A and about 10 fold reduced compared to BCG alone [52]. Similarly, purified recombinant M. tuberculosis Ag85B or TB10.4 used as monovalent protein vaccinations was shown by Dietrich et al. to decrease bacterial burden in the mouse model of tuberculosis; furthermore, vaccination using either a mixture of the two or a fusion of the antigens significantly further reduced bacterial loads [34]. Interestingly, the fusion strategy for two MU proteins may itself confer an immunological advantage over separately expressed antigens, as shown by Palendira et al. [53]. In this study, vaccination with BCG separately overexpressing both M. tuberculosis Ag85B and the immunodominant M. tuberculosis antigen, ESAT-6, was not as efficacious as a BCG strain encoding a fusion between the two antigens in reducing bacterial burden.
Future improvements in strategies for BU vaccine design could be made to increase the protective qualities demonstrated by rBCG. We have shown that expression of MU-Ag85A or MU-Ag85B-EsxH both confer protection against the effects of MU infection, suggesting that future studies may have success in designing BCG strains which express a combination of these antigens or novel immunodominant MU proteins. Indeed, Sun et al. previously demonstrated that a triple antigen encoding strain of BCG (AERAS-402) which expressed M. tuberculosis Ag85A, Ag85B, and TB10.4 generated stronger immune responses and conferred significantly improved survival compared to BCG vaccinated mice when challenged with a hypervirulent M. tuberculosis strain [36]. In addition to improvements in priming strategies, boosting with recombinant protein or other recombinant mycobacterial or viral vectors may also enhance vaccine efficacy. We previously demonstrated that boosting rBCG with recombinant M. smegmatis significantly improved protection in the mouse model of BU [15]. Furthermore, human clinical trials have effectively used recombinant modified vaccinia virus Ankara (MVA) expressing Ag85A to boost BCG by inducing potent and durable Th1-type responses as well as high levels of antigen specific T cell populations [30]. This approach has also recently been successful in boosting the triple antigen AERAS-402 strain of BCG to generate antigen-specific CD8+ T cells which produced high levels of anti-mycobacterial IFN-γ, TNF-α and IL-2 [35].
Use of mycolactone negative mutants or other attenuated MU strains may represent an alternative approach as well. There is experimental evidence to support this strategy, whereby subcutaneous vaccination with the mycolactone negative strain, MU5114, yielded a delay to footpad swelling similar to that induced by BCG vaccination [41]. Furthermore, species of mycobacteria possessing greater genetic homology to MU than BCG may represent a richer source of protective antigens and lack the potential pathological features of an MU-based vaccine. Indeed, we and others have previously demonstrated that subcutaneous vaccination with M. marinum was significantly better able to delay the pathology of MU infection versus BCG alone [51, 54]. Further novel attenuated and immunogenic MU strains, or strains with high genetic similarity to MU, may hold potential as priming or boosting agents to recombinant BCG.
The well-supported safety profile, an established global administration infrastructure, reasonable cost of production, and previously demonstrated protection support BCG as an ideal vehicle for BU vaccine design. Further discovery of MU-specific antigens could be an invaluable source of protective immunogens and, upon combined expression by BCG, may achieve complete protection against experimental MU infection. This, in addition to identification of cross-reactive antigens which confer the intrinsic protection observed with BCG, will be essential to the application of this ubiquitous vehicle as an efficacious and safe Buruli ulcer vaccine.
|
10.1371/journal.pbio.1000619 | Tsetse Immune System Maturation Requires the Presence of Obligate
Symbionts in Larvae | Beneficial microbial symbionts serve important functions within their hosts,
including dietary supplementation and maintenance of immune system homeostasis.
Little is known about the mechanisms that enable these bacteria to induce specific
host phenotypes during development and into adulthood. Here we used the tsetse fly,
Glossina morsitans, and its obligate mutualist,
Wigglesworthia glossinidia, to investigate the co-evolutionary
adaptations that influence the development of host physiological processes.
Wigglesworthia is maternally transmitted to tsetse's
intrauterine larvae through milk gland secretions. We can produce flies that lack
Wigglesworthia
(GmmWgm−) yet retain their other
symbiotic microbes. Such offspring give rise to adults that exhibit a largely normal
phenotype, with the exception being that they are reproductively sterile. Our results
indicate that when reared under normal environmental conditions
GmmWgm− adults are also
immuno-compromised and highly susceptible to hemocoelic E. coli
infections while age-matched wild-type individuals are refractory. Adults that lack
Wigglesworthia during larval development exhibit exceptionally
compromised cellular and humoral immune responses following microbial challenge,
including reduced expression of genes that encode antimicrobial peptides
(cecropin and attacin), hemocyte-mediated
processes (thioester-containing proteins 2 and 4
and prophenoloxidase), and signal-mediating molecules
(inducible nitric oxide synthase). Furthermore,
GmmWgm− adults harbor a reduced
population of sessile and circulating hemocytes, a phenomenon that likely results
from a significant decrease in larval expression of serpent and
lozenge, both of which are associated with the process of early
hemocyte differentiation. Our results demonstrate that
Wigglesworthia must be present during the development of immature
progeny in order for the immune system to function properly in adult tsetse. This
phenomenon provides evidence of yet another important physiological adaptation that
further anchors the obligate symbiosis between tsetse and
Wigglesworthia.
| Beneficial bacterial symbionts, which are ubiquitous in nature, are often
characterized by the extent to which they interact with the host. In the case of
mutualistic symbioses, both partners benefit so that each one can inhabit diverse
ecological niches where neither could survive on its own. Unfortunately, little is
known about the functional mechanisms that underlie mutualistic relationships.
Insects represent a group of advanced multi-cellular organisms that harbor
well-documented symbiotic associations. One such insect, the tsetse fly, harbors a
maternally transmitted bacterial mutualist called Wigglesworthia
that provides its host with essential metabolites missing from its vertebrate
blood-specific diet. In this study, we further examine the relationship between
tsetse and Wigglesworthia by investigating the interaction between
this bacterium and its host's immune system. We have found that when
Wigglesworthia is absent from tsetse during the maturation of
immature larval stages, subsequent adults are characterized by an underdeveloped
cellular immune system and thus highly susceptible to infection with a normally
non-pathogenic foreign microbe. These findings represent an additional adaptation
that further anchors the steadfast relationship shared between tsetse and its
obligate symbiont.
| Bacteria comprise the most abundant and diverse life form on earth. The ubiquity of
bacteria means they have colonized virtually every ecological niche, including
habitation within more evolutionarily sophisticated multi-cellular animals. Co-evolution
over millions of years has provided an opportunity for beneficial symbiotic associations
to develop between phylogenetically distant taxa. Such affiliations are often
mutualistic, meaning both partners benefit so that each can successfully inhabit diverse
environments that neither could survive in on its own [1],[2]. Deciphering the mutualistic
relationships between prokaryotic bacteria and multi-cellular eukaryotic animals is a
rapidly advancing field of research.
Performing detailed investigations into the relationships between symbiotic bacteria and
higher eukaryotes often involves costly and complex procedures. However, insects have
well-documented symbioses that are attractive to study because they have relatively
short generation times and are easy and less costly to rear. One insect that harbors
multiple symbionts is the tsetse fly, Glossina morsitans. These
microbes include the commensal Sodalis, the parasite
Wolbachia, and the obligate mutualist Wigglesworthia
glossinidia
[3]. Molecular
phylogenetic analysis indicates that tsetse's symbiosis with
Wigglesworthia is ancient, dating back 50–80 million years
[4]. The concordant
nature of tsetse's obligate association with Wigglesworthia has
driven the co-evolution of biological adaptations that are beneficial to both partners.
For example, the localization of Wigglesworthia cells within host
bacteriocytes provides a protective and metabolically favorable niche for this bacterium
[5]. In return
tsetse derives benefit from Wigglesworthia in at least two distinct
ways. First, tsetse feeds exclusively on vertebrate blood, which is deficient in
vitamins essential for survival. In accordance, a large proportion of
Wigglesworthia's streamlined (700 kB) genome encodes vitamin
biosynthesis pathways that presumably supplement tsetse's restricted diet [5],[6]. Second, more recent studies indicate
that Wigglesworthia may serve an immunologic role in tsetse. These
flies are the sole vector of pathogenic African trypanosomes, the causative agent of
sleeping sickness in humans [7]. In laboratory experiments infection with immunogenic
trypanosomes results in a decrease in tsetse fecundity [8]. Furthermore, parasite infection
prevalence is higher in flies that lack Wigglesworthia when compared to
age-matched wild-type (WT) individuals [9]. Wigglesworthia is thought to influence
tsetse's vectorial competence by modulating its host's humoral immune system
[10]. Thus, by
preventing energetically costly parasite infections, this obligate symbiont may
indirectly benefit the reproductive fitness of tsetse.
Symbiotic bacteria are rapidly gaining recognition for their important contributions to
host development and immunity. The most well-known example of this type of interaction
involves the mammalian microbiome, which modulates gut development during early
postnatal life and subsequently shapes our mucosal and systemic immune systems [11],[12]. Symbionts also
serve similar functions in invertebrate hosts. For example, light organ morphogenesis in
juvenile bobtail squid (Euprymna scolopes) initiates only after
symbiotic Vibrio fischeri cells have stably colonized this tissue [13]. The pea aphid,
Acyrthosiphon pisum, harbors a secondary symbiont,
Hamiltonella defensa, which can be infected with a lysogenic
bacteriophage (A. pisum secondary endosymbiont; APSE). A.
pisum that harbor both H. defensa and APSE are protected
from being consumed by a parasitic wasp larvae through the action of phage-encoded
toxins [14].
To the best of our knowledge no evidence exists that demonstrates insect symbionts can
confer an impending protective phenotype in their adult host by directing immune system
development during immature stages. In the present study we investigate the mechanism by
which Wigglesworthia contributes to the development and function of
cellular and humoral immune responses in adult tsetse. Our results show that the
maturation and normal function of adult cellular immune responses in tsetse are severely
compromised when Wigglesworthia is absent during larval development.
This study reveals an important new facet that further anchors the obligate relationship
between tsetse and Wigglesworthia and may serve as a useful model to
understand the highly integrated and dynamic relationship between hosts and their
beneficial bacterial fauna.
Insects are normally capable of mounting an immune response that combats infection
with various groups of bacteria. Interestingly, in comparison to
Drosophila, tsetse flies are uniquely susceptible to septic
infection with 103 colony-forming units (CFU) of normally non-pathogenic
Escherichia coli (E. coli) K12 [15]. In the present
study we further investigated tsetse's unique susceptibility to E.
coli infection by subjecting wild-type
(GmmWT) and adults from two age groups to hemocoelic
infections with varying quantities of E. coli K12. Three-day-old
GmmWT individuals (flies from this age group are
hereafter referred to as “young”) were highly susceptible to this
treatment, as 103 CFU resulted in the death of all flies by 8 d
post-infection (dpi; Figure
1A, top graph). In contrast, 77% and 55% of 8-d-old WT
individuals (flies 8 d old and older are hereafter referred to as
“mature”) survived for 14 dpi with 103 and 106 CFU
of E. coli K12, respectively (Figure 1A, middle graph).
We demonstrated previously that feeding pregnant female tsetse the antibiotic
ampicillin results in the generation offspring that lack
Wigglesworthia
(GmmWgm−) but still harbor
Sodalis
[9] and presumably
Wolbachia. In the present study we determined the survival
outcome of mature GmmWgm− inoculated
with 106 CFU of E. coli K12 (the dose required to kill
∼50% of mature WT flies; Figure 1A, middle graph). In comparison to age-matched WT tsetse,
GmmWgm− flies were highly
susceptible to this infection. In fact, at 14 dpi, only 1% of these
individuals remained (Figure 1A,
bottom graph). We reasoned that the dramatic susceptibility of mature
GmmWgm− flies to infection with
E. coli could result from one of two scenarios. In the first
scenario Wigglesworthia may directly contribute to the adult immune
response to a foreign micro-organism. In the second scenario, the absence of
Wigglesworthia during larvagensis may give rise to
GmmWgm− adults with compromised
immune functions. To determine if the first scenario is correct we treated mature WT
tsetse with the antibiotic tetracycline to eliminate all of their microbiota,
including bacteriome-associated Wigglesworthia (Figure 1B), and subsequently challenged these adult
flies (GmmWT/Wgm−) with
106 CFU of tetracycline resistant E. coli K12. We
found that, similar to their WT counterparts (GmmWT;
Figure 1A, middle panel),
about 70% of GmmWT/Wgm−
individuals survived this infection (Figure 1C). Conversely,
GmmWgm− adults were highly
susceptible (Figure 1A, bottom
graph). This result suggests that when Wigglesworthia is absent from
tsetse during larval development subsequent adults are severely
immuno-compromised.
We demonstrated that when larval tsetse lacks exposure to
Wigglesworthia in utero their immune system is highly compromised
during adulthood. This finding does not exclude the possibility that treatment of
female flies with ampicillin to produce
GmmWgm− offspring induces
perturbations in other constituents of tsetse's microbiota. Because
tsetse's other two known endosymbionts, Sodalis and
Wolbachia, are present at similar densities in
GmmWT and
GmmWgm− adults (Figure 1D), we looked for the presence of
uncharacterized endosymbionts or digestive tract-associated microbes that are
potentially passed on to developing intra-uterine larvae where they subsequently
impact immunity. We generated a clone library containing 16s rRNA gene sequences PCR
amplified from 3rd instar
GmmWgm− and
GmmWT larvae and then sequenced multiple clones from
each tsetse line. Our results indicate that the proportion of
Wigglesworthia, Sodalis, and
Wolbachia in 3rd instar
GmmWT larvae is 19%, 75%, and 6%,
respectively. As expected no Wigglesworthia 16s rRNA sequences were
present in GmmWgm− larvae, and the
proportion of Sodalis and Wolbachia sequences
present was 71% and 29%, respectively (Figure 1E). No uncharacterized endosymbionts or
gut-associated environmental microbes were present in any sample. Taken together,
these results suggest that the presence of Wigglesworthia
specifically is responsible for enabling immune system maturation in WT tsetse.
Our host survival curves indicate that mature GmmWT and
GmmWT/Wgm− can survive
infection with E. coli while young GmmWT
and mature GmmWgm− perish. To determine
a cause for the variation in survival we observed between the four groups following
infection with E. coli, we monitored bacterial growth dynamics over
the course of the experiment in each group. Bacterial number within mature
GmmWT and
GmmWT/Wgm− peaked at
3.8×104 and 1.9×104
E. coli, respectively, over the 2-wk period, indicating that these
groups appeared able to control their infections. In contrast, E.
coli increased exponentially in young GmmWT
and mature GmmWgm− and reached a maximum
density at 6 dpi of 4.2×106 and 8.8×106,
respectively (Figure 1F). These
results implicate bacterial sepsis as the cause of high mortality observed in young
GmmWT and mature
GmmWgm−. All of the above-mentioned
results taken together indicate that mature GmmWT are
considerably more resistant to infection with a foreign microbe than are their
younger counterparts. Furthermore, tsetse's obligate mutualist,
Wigglesworthia, must be present during the development of
immature stages so that mature adults are able to overcome infection with E.
coli.
To understand the basis of the compromised immunity we observed in mature
GmmWgm− flies, we evaluated the
expression profile of a set of immunity-related genes from age-matched mature adult
GmmWT and
GmmWgm− flies that were either
uninfected or 3 dpi with E. coli K12. We included the antimicrobial
peptides (AMPs) attacin, cecropin, and
defensin, which distinctly target gram-negative
(attacin and cecropin) and gram-positive
bacteria (defensin) [16]. Our analysis also included
thioester-containing proteins (tep2 and tep4),
prophenoloxidase (PPO), and inducible nitric oxide synthase
(iNOS). In insects TEPs likely function as pathogen-specific
opsonins that bind to bacteria or parasites and promote their
phagocytosis/encapsulation [17], while PPO initiates a proteolytic cascade that results in
melanin deposition [18]. iNOS catalyzes synthesis of the signaling molecule nitric
oxide (NO), which plays a role in humoral and cellular immunity in Anopheline
mosquitoes [19], reduvid bugs [20], and Drosophila
[21],[22] by inducing AMP
expression and recruiting hemocytes to the site of infection.
Our expression results indicate that symbiont status plays little or no role in the
expression of immunity-related genes in uninfected adults. In fact, with the
exception of the AMP defensin, no significant differences in
immunity-related gene expression between mature uninfected
GmmWT and
GmmWgm− adults (Figure 2A). However, we observed a considerably
different profile of immunity-related gene expression when these different fly
strains were infected with E. coli K12. Under these circumstances,
all of the genes evaluated (with the exception of defensin) were expressed at
significantly higher levels in GmmWT compared to
GmmWgm− individuals (Figure 2B). Particular striking was
the fact that the induction of pathways associated with cellular immunity, such as
pathogen recognition (tep2 and tep4) and
melanization (PPO), were significantly compromised in mature
GmmWgm− adults. The absence of a
robust cellular immune response is likely the cause of high mortality among these
individuals following E. coli infection. This analysis indicates
that Wigglesworthia must be present during the development of
immature tsetse in order for immune-related genes to subsequently be expressed in
mature E. coli–infected adults.
Our analysis of immunity-related gene expression in tsetse suggests that cellular
immune pathway functions in adult tsetse are particularly compromised when
Wigglesworthia is absent during immature development. The most
prominent cellular immune mechanisms include melanization and phagocytosis. These
processes, which ultimately result in the removal of foreign invaders in
Drosophila
[23] and A.
aegypti
[24], both arise
from distinct crystal cell and plasmatocyte hemocyte lineages, respectively [25].
We investigated the role hemocytes might play in determining the susceptible
phenotype we observed in tsetse following infection with E. coli. We
infected mature GmmWT individuals with GFP-expressing
E. coli and were able to observe that hemocytes had engulfed a
large number of the introduced cells by 12 hpi (Figure 3A). We next inhibited phagocytosis by
introducing blue fluorescent microspheres directly into tsetse's hemocoel and 12
h later infected the bead-treated individuals with GFP-expressing E.
coli. Microscopic inspection of hemocytes harvested 12 hpi with
E. coli revealed the presence of internalized microspheres and
the absence of engulfed E. coli. This observation indicated that we
were successful in blocking hemocyte phagocytosis (Figure 3B). We subsequently maintained our
microsphere-injected tsetse for 2 wk with the intention of determining the impact of
impaired phagocytosis on host survival outcome. Mature
GmmWT flies exhibiting impaired phagocytosis were
highly susceptible to infection with both 1×103 and
1×106
E. coli K12. In fact, by day 12 post-infection, all of these flies
had perished regardless of the initial dose used for infection (Figure 3C). This observation contrasts starkly with
infection outcome in mature GmmWT that exhibit normally
functioning hemocytes (Figure 1A,
middle graph). Our results suggest that defects in phagocytosis severely compromise
the ability of these tsetse flies to overcome bacterial infection.
A notable result of our immunity-related gene expression analysis was a 37-fold
decrease in PPO levels in
GmmWgm− flies. This enzyme is an
essential component of the melanization pathway, and its expression ultimately
results in host wound healing and the melanization, encapsulation, and subsequent
removal of foreign microorganisms [26]–[28]. In conjunction with the remarkable variation in
PPO expression observed between GmmWT
and GmmWgm−, we were also able to
visually observe the absence of a melanization response to E. coli
infection in flies lacking Wigglesworthia. In fact, 30 min
post-injection with E. coli, hemolymph was still actively exuding
from the inoculation wound of GmmWgm−
flies. Conversely, in WT individuals no hemolymph was detectable and melanin was
deposited at the wound site (Figure
4). These results further suggest that hemocyte-mediated cellular immunity
provides an imperative defense against the establishment of bacterial infections in
tsetse's hemocoel. Furthermore, the absence of this response in
GmmWgm− individuals was likely
responsible for the compromised host survival phenotype we observed following
infection with E. coli.
We observed that young GmmWT were markedly more
susceptible to infection with E. coli K12 than were their mature
counterparts. Furthermore, symbiont status also altered infection outcome, as mature
GmmWgm− perished following
E. coli infection while age-matched WT individuals survived.
These differential infection outcomes appeared to result from disparities in cellular
immune system function between the different tsetse lines we examined. Based on these
observations we hypothesized that the obligate mutualist
Wigglesworthia plays a crucial role in regulating the development
of cellular immunity in tsetse during immature stages. To test this hypothesis we
quantified the number of circulating and sessile hemocytes present in young and
mature adult GmmWT and mature adult
GmmWgm−. Our results indicate that
a 1.4-fold increase in circulating hemocyte number occurs between day 3 and day 8 in
WT tsetse, while no significant change in circulating hemocyte number was observed
between young and mature GmmWgm− (Figure 5A). Interestingly, mature
GmmWT adults harbored 3.4× more circulating
hemocytes than did mature GmmWgm− adults
(Figure 5A). We also looked at
sessile hemocyte abundance as a further indicator of
Wigglesworthia's impact on the development of cellular
immunity in tsetse. In Drosophila, this hemocyte subtype
concentrates in large quantities around the anterior end of the fly's dorsal
vessel [29].
Thus, we indirectly quantified sessile hemocyte abundance immediately adjacent to the
anterior-most chamber of tsetse's dorsal vessel by measuring the fluorescent
emission of microspheres that were found engulfed in this region. Young
GmmWT adults engulfed 1.2× more microspheres
than their mature counterparts and 15.7× more than mature
GmmWgm− adults. Furthermore,
mature WT adults engulfed 13.2× more microspheres than did age-matched
GmmWgm− adults (Figure 5B).
Our results demonstrate that Wigglesworthia must be present during
the development of immature tsetse in order for cellular immunity to develop and
function properly in adults. Thus, we speculate that the absence of hemocytes in
adult tsetse should reflect a lack of blood cell differentiation during the
development of immature stages. In Drosophila the process of blood
cell differentiation, or hematopoiesis, begins in the embryo and proceeds through all
larval stages [30].
During Drosophila embryogenesis early hematopoiesis can be
distinguished by the expression of the zinc finger transcription factor
“Serpent.” Subsequently, another transcription factor,
“Lozenge,” directs the differentiation of
serpent-expressing precursor cells into a specific lineage of
hemocytes called crystal cells. To address the relationship between the presence of
Wigglesworthia and early hematopoiesis in tsetse, we used qPCR to
evaluate the relative number of serpent and lozenge
transcripts present in 1st, 2nd, and 3rd instar
larvae dissected from pregnant GmmWT and
GmmWgm− females. Larval instars
L1, L2, and L3 from WT females expressed 1.7, 2.1, and 1.9 times more
serpent transcripts, and 4, 4.4, and 3.9 times more
lozenge transcripts, respectively, than did their counterparts
from females that lacked Wigglesworthia (Figure 5C). This observed attenuated expression of
both serpent and lozenge may account for the
depleted hemocyte population, and compromised cellular immune function, we observed
in adult GmmWgm− individuals.
In the present study we demonstrate that Wigglesworthia is intimately
involved in regulating the maturation and function of tsetse's cellular immune
system during immature larval development. We present a model that links the presence of
Wigglesworthia in larval progeny with host immune system maturation
during development and the subsequent ability of adult tsetse to overcome infection with
foreign microbes (Figure 6).
Obligate symbioses between intracellular bacteria and multi-cellular eukaryotes
represent millions of years of co-evolution during which time both partners have adapted
to increase each other's overall fitness. The association between tsetse and
Wigglesworthia is an example of this reciprocal relationship in that
neither organism can survive in the absence of the other.
We present several lines of evidence indicating that tsetse's resistance to
E. coli positively correlates with fly age and symbiont infection
status during juvenile stages. Specifically, our results signify that mature
GmmWT adults are resistant to E. coli
infection, while young GmmWT adults are susceptible. In
contrast, both young and old adult flies that lack Wigglesworthia
throughout all developmental stages are killed by E. coli infections.
Like their WT counterparts, GmmWgm− larvae
acquire both Sodalis and Wolbachia while in utero
[31]. Although the
intrauterine larval environment is otherwise aseptic, adult
GmmWgm− can be exposed to a wide range
of environmental microbes during adulthood. However, neither the presence of other
symbiotic bacteria (Sodalis and Wolbachia) in larvae,
nor environmental microbes acquired during adulthood, appear to be sufficient to induce
immunity in GmmWgm− adults. Additionally,
we show that mature GmmWT adults treated with antibiotics to
eliminate Wigglesworthia
(GmmWT/Wgm−) remain resistant
to E. coli infections. The resistant phenotype of
GmmWT/Wgm− further signifies
that obligate Wigglesworthia is not directly responsible for the
ability of mature adult GmmWT to overcome septic bacterial
infection. Instead, it appears that Wigglesworthia's presence
during the maturation of larval progeny stimulates the development of host immunity in
adults.
Our quantitative analysis of gene expression indicates that
GmmWT and
GmmWgm− adults exhibit no significant
difference in the expression of genes that serve as hallmarks of humoral and cellular
immunity in the absence of microbial challenge. However, following infection with
E. coli, all pathways were significantly compromised in
GmmWgm− versus
GmmWT adults. The most notable discrepancy observed
between the two tsetse lines involved the expression of
prophenoloxidase. Following E. coli infection, the
expression of this gene increased 37-fold in WT tsetse but remained virtually unchanged
in individuals that lacked Wigglesworthia. In WT insects, PPO, which is
an inactive zymogen, is proteolytically cleaved to produce phenoloxidase upon mechanical
injury or the presence of foreign pathogens. Phenoloxidase then facilitates the process
of melanin deposition [18]. A functional melanotic pathway would likely increase
tsetse's resistance to E. coli infection in several ways. First,
sequestration of E. coli in a melanotic capsule at the site of
inoculation would likely help prevent their dissemination into adjacent host tissues.
This phenomenon has been observed in both the hawk moth, Manduca sexta,
and Drosophila following infection with pathogenic bacteria
(Photorhabdus luminescens) and parasitic wasp eggs, respectively
[32],[33]. In both cases the
lack of melanization at the wound site severely compromised the host's ability to
subsequently overcome the foreign invader. Second, the melanization cascade results in
the production of reactive oxygen intermediates that are directly toxic to foreign
pathogens. In the flesh fly Sarcophaga peregrina and M.
sexta, melanization intermediates such as DOPA exhibit direct antimicrobial
activity [34],[35]. Finally, melanin at
the site of inoculation expedites wound healing that could prevent the spread of
secondary infections [26].
The E. coli–susceptible phenotype of mature
GmmWgm− adults is likely reflective
of their blood cell (hemocyte) deficit in comparison to E.
coli–resistant WT flies. In Drosophila
90%–95% of the total hemocyte population is composed of sessile and
circulating plasmatocytes, which are a distinct hemocyte lineage predominantly
responsible for engulfing and digesting foreign pathogens [36]. The inability of mature
GmmWgm− tsetse to survive infection
with E. coli may result from their significantly reduced population of
phagocytic hemocytes available to engulf bacteria injected into the hemocoel. In fact,
by injecting polystyrene beads as a means of blocking this physiological process, we
demonstrate that phagocytosis is a critical component of tsetse's ability to manage
septic infection with E. coli in WT flies. Similarly, mutant
Drosophila that contain a depleted plasmatocyte population, or have
been injected with beads to prohibit phagocytosis, exhibit a remarkable susceptibility
to a variety of gram positive and negative bacteria [37],[38]. Interestingly,
Drosophila that lack functional plasmatocytes also exhibit a reduced
capacity to activate humoral immune responses, further inhibiting their ability to fight
bacterial infection [29].
In Drosophila the crystal cell hemocyte lineage controls the humoral
melanization cascade via the release of PPO stored in large cytoplasmic inclusion bodies
[30],[39]. In addition to the
dramatic reduction in PPO expression in
GmmWgm− tsetse, two further lines of
evidence indicate that this fly strain harbors a significantly reduced population of
hemocytes that function in a homologous manner to Drosophila crystal
cells. First, lozenge expression in all three larval instars from
GmmWgm− females is significantly
lower than in their GmmWT counterparts. In
Drosophila larval crystal cells fail to form in the absence of
lozenge expression, yet the differentiation of other hemocyte types
proceeds normally [40].
Second, we observed that hemocyte-deficient
GmmWgm− flies were unable to produce a
viable clot at the site of bacterial inoculation. In mutant Drosophila
strains that lack crystal cells, PPO is absent from the hemolymph. Consequently, the
melanization cascade fails to initiate and hard clots do not form at wound sites [41].
We discovered that young GmmWT adults harbor significantly
less circulating hemocytes than do mature WT adults. This observation implies that
circulating hemocyte number in adult WT tsetse increases as a function of age, although
the specific mechanism underlying this process in tsetse is currently unknown. In adult
Drosophila intact lymph glands are absent and no evidence exists for
the de novo synthesis of hemocytes following metamorphosis [30],[40]. Interestingly, we did observe more
subepidermal sessile hemocytes in young compared to mature
GmmWT adults, although the number was not statistically
significant between the two groups. We speculate that the increased abundance of
circulating hemocytes we observed in mature compared to young
GmmWT may reflect a shift in the proportion of sessile to
circulating cells instead of de novo production of new hemocytes. A comparable process
occurs in Drosophila larvae and adults, where the proportion of sessile
to circulating hemocytes changes following immune stimulation [42],[43]. Furthermore, after larval
Drosophila receive an epidermal wound, circulating hemocytes are
rapidly recruited to the site of injury. These circulating cells are phagocytically
active and likely function as a front-line surveillance system against tissue damage and
microbial infection [44]. In the mosquito malaria vector Anopheles
gambiae, exposure to Plasmodium parasites stimulated an
increase in the number of granulocytes circulating in the hemolymph. These primed
mosquitoes were subsequently more resistant to infection with pathogenic bacteria than
their wild-type counterparts [45]. We propose that young GmmWT
adults and mature GmmWig− adults may be
susceptible to E. coli infection in part because they harbor
significantly less circulating hemocytes than do older WT individuals that are
resistant. Our previous studies indicated that
GmmWig− adults are highly susceptible to
trypanosome infection [9],[10] and that this phenotype may be modulated by tsetse's
humoral immune system [46]–[48]. Studies are ongoing to determine if cellular immunity also
modulates tsetse's trypanosome vectorial competence.
Neonatal humans (and presumably other mammals as well) acquire probiotic gut microbes
through the act of breast feeding [49]. These bacteria subsequently stably colonize the naïve
intestine where they promote immune system maturation and enhance defense against
infection with pathogenic microbes [50]. Most lower eukaryotes,
including insects, hatch from an egg deposited into the environment and thus rely mainly
on environmental microbes to stimulate their innate immune systems during the
development of immature stages. For example, Anopheline mosquitoes and
Drosophila are exposed to environmental microbes throughout all
stages of their development. Under these natural conditions subsequent adults exhibit
potent innate immunity [51],[52]. However, when these insects are reared under germ-free
conditions, they exhibit a severely compromised humoral immune response [51],[53]. Tsetse flies are unique among other
insects because they lead a relatively aseptic existence. Not only do they feed
exclusively on sterile vertebrate blood, but they also exhibit a unique viviparous
reproductive strategy. During this type of reproduction, all embryonic and larval stages
develop within the female's uterus. The protected in utero environment in which
viviparous offspring develop limits their exposure to environmental microbes. However,
throughout tsetse larvagenesis maternal milk gland secretions provide the developing
offspring with nourishment as well as Wigglesworthia,
Sodalis, and Wolbachia
[31]. Interestingly,
recently established Sodalis and Wolbachia appear
unable to influence their host's physiology to the same extent as mutualist
Wigglesworthia despite their presence throughout larval development.
Future studies on this system will focus on determining the chemical and/or metabolic
elements provided by Wigglesworthia that stimulate the maturation of
its host's immune system during larval development.
The results of this study provide evidence of a novel functional role for obligate
insect symbionts–host immune system activation during immature developmental
stages to ensure robust function during adulthood. We demonstrate that the obligate
Wigglesworthia provides tsetse offspring the stimuli necessary for
immune system development, a process that exhibits functional parallels with the
mammalian system following the transfer of beneficial microbes from mothers to their
neonates during that act of breast feeding. This phenomenon represents an adaptation
that further anchors the steadfast relationship shared between tsetse and its obligate
mutualist. The essential nature of tsetse's dependence on
Wigglesworthia provides a potentially exploitable niche to
experimentally modulate host immunity, with the intention of diminishing this
insect's capacity to vector deadly trypanosomes.
Throughout the text, 3-d-old and 8-d-old tsetse are referred to as
“young” and “mature,” respectively. Wild-type G. m.
morsitans (GmmWT) were maintained in
Yale's insectary at 24°C with 50%–55% relative humidity.
These flies received defibrinated bovine blood every 48 h through an artificial
membrane feeding system [54]. Two Wigglesworthia-free tsetse lines were
generated for use in our experiments. The first,
GmmWgm−, was generated by
supplementing the diet of pregnant females with 25 µg of ampicillin per ml of
blood. GmmWgm− (offspring of
ampicillin-fed females) adult are devoid of Wigglesworthia
throughout all developmental stages [9]. PCR using Wigglesworthia thiamine
C-specific primers (forward, 5′-TGAAAACATTTGCAAAATTTG-3′; reverse,
5′-GGTGTTACATAGCATAACAT-3′) confirmed that this
tsetse strain lacked Wigglesworthia. The second
Wigglesworthia-free tsetse line,
GmmWT/Wgm−, was generated by
feeding mature WT adults three blood meals supplemented with 40 µg/ml
tetracycline. GmmWT/Wgm−
individuals thus harbored Wigglesworthia, Sodalis,
and Wolbachia throughout the development of all immature stages and
early adulthood, but lacked these symbionts thereafter. The absence of
Wigglesworthia from mature
GmmWT/Wgm− adults was
confirmed microscopically by comparing the bacteriome
(Wigglesworthia-harboring organ) contents of mature WT and
GmmWT/Wgm− adults.
Septic infection of tsetse was achieved by anesthetizing flies with CO2
and subsequently injecting individuals with live bacterial cells using glass needles
and a Narashige IM300 micro-injector. The methods used to produce
luciferase-expressing E. coli K12 (recE.
colipIL), and the assay used to quantify luciferase expression
in vivo, were described previously [15]. GFP-expressing and tetracycline-resistant E.
coli K12 were produced via electroporation with pGFP-UV (Clontech,
Mountain View, CA) and pBR322 (Promega, Madison, WI) plasmid DNA, respectively. All
flies treated with tetracycline were subsequently infected with tetracycline
resistant E. coli K12. The number of bacterial cells injected and
control group designations for all infection experiments are indicated in the
corresponding figures and their legends. For all survival experiments, treatments
were performed in triplicate, using 25 flies per E. coli treatment
replicate. LB media controls were performed once using 25 flies.
For analysis of immunity-related gene expression, sample preparation and qPCR were
performed as described previously [15]. Quantitative measurements were performed on three
biological samples in duplicate and results were normalized relative to tsetse's
constitutively expressed β-tubulin gene (determined from each corresponding
sample). Fold-change data are represented as a fraction of average normalized gene
expression levels in bacteria-infected flies relative to expression levels in
corresponding uninfected controls.
For symbiont quantification, total RNA was prepared from 40-d-old adult
GmmWT and
GmmWgm− flies. Symbiont genome
numbers were quantified using single-copy Sodalis fliC and
Wolbachia groEL. Relative symbiont densities were normalized to
tsetse β-tubulin. All qPCR was performed with an icycler iQ real
time PCR detection system (Bio-Rad). Values are represented as the mean
(±SEM). qPCR primer sequences are shown in Table S1.
Universal bacterial 16S rRNA gene primers 27F (5′-AGAGTTTGATCCTGGCTCA
G-3′) and 1492R (5′-GGTTACCTTGTTACGACTT-3′) [55] were used to produce a clone
library of 16s rRNA gene sequences found in 3rd instar
GmmWT and
GmmWgm− larvae. Five individual
larvae from each tsetse line were dissecting under sterile conditions and washed in
DNase (Ambion, Austin, TX) to remove any surface contamination prior to DNA
extraction. Genomic DNA was isolated using Holmes-Bonner buffer (0.1 mol/L Tris-HCl,
pH 7.5; 0.35 mol/L NaCl; 10 mmol/L EDTA, pH 8.0; 2% SDS; 7
mol/L Urea), purified via phenol-chloroform extraction and precipitated in
100% EtOH. PCR was performed under standard reaction conditions [15], and the resulting
products were cloned in the pGEM-T vector (Promega, Madison, WI). Twenty clone
inserts from each larvae (100 in total from each tsetse line) were sequenced using
the T7 vector specific primer, and homology to previously described 16s rRNAs was
determined using the blastn database.
Depending on the subsequent experiment, tsetse hemolymph was collected using one of
two methods. For hemocyte quantification, undiluted hemolymph was collected by
removing one front fly leg at the joint nearest the thorax and then applying gentle
pressure to the distal tip of the abdomen. Hemolymph exuding from the wound was
collected using a glass micro-pipette and placed into a microfuge tube on ice.
Hemocytes were quantified microscopically using a Bright-Line hemocytometer, and
hemocyte numbers are represented as cells per µl of hemolymph.
When hemocytes were required for microscopic visualization, hemolymph was collected
by employing a modified version of the high injection/recovery method previously
developed for use in mosquitoes [56]. In brief, tsetse flies were sedated on ice and injected
with 25 µl of chilled anticoagulant buffer [70% MM medium,
30% anticoagulant citrate buffer (98 mM NaOH, 186 mM NaCl, 1.7 mM EDTA, and 41
mM citric acid, buffer pH 4.5), vol/vol] between the last two abdominal
schlerites using a glass needle and a Narashige IM300 micro-injector. Following a 30
min incubation on ice, a front leg was removed at the joint most proximal to the
thorax. At this point internal pressure forced hemolymph diluted with anticoagulant
buffer to be expelled from the wound site (more liquid could be recovered by applying
gentle pressure to the distal end of the abdomen). Liquid was collected using a
pipette and either placed into a chilled microfuge tube or directly into a 24-well
cell culture plate. In the latter case, cells were allowed to adhere to the plate
bottom, after which anticoagulant buffer was replaced with MM media.
Sessile hemocytes were observed by intra-thoracically injecting young and mature
GmmWT, and mature
GmmWgm−
(n = 5 of each strain), with 2 µl of blue
fluorescent (365/415 nm) 0.2 µm carboxylate-modified beads (Invitrogen corp.).
Prior to use, beads were washed once in PBS and resuspended in 100% of their
original volume. Flies were dissected 12 h post-injection to reveal their dorsal
vessel and surrounding tissue, which was gently washed 3 times with PBS to remove any
potentially contaminating circulating (non-adherent) hemocytes. Engulfed microspheres
were visualized using a Zeiss steriomicroscope (Discovery v8) equipped with a coaxial
fluorescence module. Semi-quantitative comparison of sessile hemocyte number between
young and mature GmmWT adults, and mature
GmmWT and
GmmWgm− adults, was performed by
quantifying fluorescent signal intensity (n = 4
individuals from each group) using ImageJ software (http://rsbweb.nih.gov/ij/).
Phagocytosis by circulating tsetse hemocytes was observed by intra-thoracically
infecting mature GmmWT flies (n
= 10) with 1×106 GFP-expressing E.
coli K12. Twelve hours post-infection, hemolymph was collected and
hemocytes monitored to determine if they had engulfed the GFP-expressing bacterial
cells. Hemolymph samples were fixed on glass microscope slides via a 2 min incubation
in 2% PFA. Prior to visualization using a Zeiss Axioscope microscope, slides
were overlayed with VectaShield hard set mounting medium containing DAPI (Vector
Laboratories, Burlingame, CA).
Phagocytosis by tsetse hemocytes was inhibited with blue fluorescent (365/415 nm) 0.2
µm carboxylate-modified beads (Invitrogen Corp.). Prior to use, beads were
washed once in PBS and resuspended in 100% of their original volume.
Inhibition assays were performed by inoculating 8-d-old
GmmWT with 2 µl of beads via their thoracic
compartment. Twelve hours later, these flies were similarly infected with
1×103 and 1×106 GFP-expressing E.
coli K12 (experiment was performed in triplicate; n
= 25 flies per replicate). Finally, 12 hours post-infection
with E. coli, hemolymph was collected and processed as described
above (these samples were overlayed with VectaShield hard set mounting medium that
lacked DAPI).
Melanization assays were performed by intra-thoracically inoculating mature
GmmWT and
GmmWgm− (n
= 10 of each strain) with 1×103
E. coli K12. Subsequently, three individuals from each group were
monitored microscopically every 10 min for the presence of melanin at the wound site.
The remaining seven flies from each group were maintained for 2 wk in order to
observe infection outcome.
Statistical significance of survival curves was determined by log-rank analysis using
JMP (v8.02) software (www.jmp.com). Statistical analysis of qPCR data was performed by
Student's t test using JMP (v8.02) software. Statistical
significance between various treatments, and treatments and controls, is indicated in
corresponding figure legends.
|
10.1371/journal.pgen.1003621 | Genetic Mapping of Specific Interactions between Aedes aegypti Mosquitoes and Dengue Viruses | Specific interactions between host genotypes and pathogen genotypes (G×G interactions) are commonly observed in invertebrate systems. Such specificity challenges our current understanding of invertebrate defenses against pathogens because it contrasts the limited discriminatory power of known invertebrate immune responses. Lack of a mechanistic explanation, however, has questioned the nature of host factors underlying G×G interactions. In this study, we aimed to determine whether G×G interactions observed between dengue viruses and their Aedes aegypti vectors in nature can be mapped to discrete loci in the mosquito genome and to document their genetic architecture. We developed an innovative genetic mapping strategy to survey G×G interactions using outbred mosquito families that were experimentally exposed to genetically distinct isolates of two dengue virus serotypes derived from human patients. Genetic loci associated with vector competence indices were detected in multiple regions of the mosquito genome. Importantly, correlation between genotype and phenotype was virus isolate-specific at several of these loci, indicating G×G interactions. The relatively high percentage of phenotypic variation explained by the markers associated with G×G interactions (ranging from 7.8% to 16.5%) is consistent with large-effect host genetic factors. Our data demonstrate that G×G interactions between dengue viruses and mosquito vectors can be assigned to physical regions of the mosquito genome, some of which have a large effect on the phenotype. This finding establishes the existence of tangible host genetic factors underlying specific interactions between invertebrates and their pathogens in a natural system. Fine mapping of the uncovered genetic loci will elucidate the molecular mechanisms of mosquito-virus specificity.
| The outcome of invertebrate host-pathogen interactions often depends on the specific pairing of host and pathogen genotypes. This genetic specificity challenges our current understanding of invertebrate resistance to pathogens because it contrasts the limited discriminatory power of known invertebrate defense mechanisms. However, genetic factors underlying this observed specificity have remained elusive, questioning their very existence. In this study, we developed an innovative strategy to localize factors in the genome of the mosquito Aedes aegypti that govern specific interactions with dengue viruses. We used large mosquito families derived from a natural population in Thailand that we experimentally challenged with virus isolates obtained from human patients living in the same area. We identified several regions of the mosquito genome that control specific interactions with dengue viruses and contribute significantly to the observed variation in vector competence. Our study establishes the existence of tangible host genetic factors underlying specific interactions between invertebrates and their pathogens in a natural system that is relevant to human health. This represents a critical step towards identification of mechanisms underlying the genetic specificity of insect-virus interactions.
| Most organisms engage in ecological interactions with organisms of different species that have profound effects on their fitness. These interactions, which can be antagonistic (e.g., parasitism, competition) or mutualistic (e.g., cooperation), are major drivers of adaptive evolution and diversification. Understanding the evolution of traits mediating ecological interactions can be complicated by their genetic specificity, whereby fitness of a genotype depends on the genotype of the interacting species [1], [2]. Such genotype-by-genotype (G×G) interactions, sometimes referred to as intergenomic epistasis, occur in both antagonistic [3] and mutualistic [4] relationships. Importantly, G×G interactions imply that the genetic basis of interaction traits is a composite entity that involves distinct genomes. Therefore, dissecting the genetic architecture (i.e., the number, position, effect and interactions between genetic loci underlying the phenotype) of these traits requires accounting jointly for genetic variation in different species [5].
Among the most intriguing examples of G×G interactions are those involved in invertebrate host susceptibility to pathogens [6]. Indeed, specific interactions between host and pathogen genotypes have been documented in a wide variety of invertebrate systems [7]–[12]. This observation challenges the long-held view that invertebrate defense against pathogens relies on broad-spectrum recognition and effector mechanisms [13], [14]. Lack of a mechanistic explanation, however, has questioned the nature of host factors underlying the observed G×G interactions [15]. For instance, the effect of host genotype can be confounded with that of symbiotic microbiota [16], raising the possibility that G×G interactions may be environmentally driven. A critical question is whether G×G interactions observed at the phenotypic level truly result from the effect of discrete genetic factors within host and pathogen genomes. More generally, understanding the ecological and evolutionary dynamics of host-pathogen interactions requires a detailed knowledge of their genetic architecture [17]. In this study, we addressed this question in a natural insect-virus association that is relevant for human health.
Aedes aegypti mosquitoes are the main vectors of dengue viruses, which cause the most prevalent mosquito-borne viral disease of humans [18]. Successful virus transmission requires that following mosquito blood feeding on a viremic host, infection is initially established in the insect's midgut cells and then disseminates throughout the rest of the body. The mosquito becomes infectious when the virus reaches the salivary glands and is released into the saliva. Vector competence defines the intrinsic ability of a mosquito to become infected following ingestion of infectious blood and to subsequently transmit the virus [19]. It varies substantially between and within Ae. aegypti populations throughout their wide geographical range [20], [21]. The existence of genetic factors underlying the observed variation in mosquito susceptibility to dengue was initially demonstrated by artificial selection of resistant and susceptible inbred lines of Ae. aegypti [22]. This finding confirmed that, as for many other host-pathogen systems [17], Ae. aegypti susceptibility to dengue has a genetic basis. Subsequent studies based on laboratory crosses of resistant and susceptible mosquito lines mapped several quantitative trait loci (QTL) controlling Ae. aegypti susceptibility to dengue virus infection and dissemination [23]–[25]. These QTL mapping studies, however, ignored the influence of viral genetic factors by exposing mosquitoes to a single, reference virus strain. A meta-analysis on the genetic architecture of host susceptibility in plants and animals revealed that QTL are recovered in only 25% of the cases when the mapping involves a different pathogen strain [17]. Dengue viruses exist in nature as four antigenically distinct serotypes (DENV-1 through DENV-4), which, in turn, consist of considerable genetic diversity [26]. Recently, we reported that several indices of Ae. aegypti vector competence for dengue viruses are governed by G×G interactions [9], [27]. Thus, the efficiency of dengue virus transmission by Ae. aegypti depends on the specific pairing of mosquito and virus genotypes.
Here, we surveyed genetic factors within the Ae. aegypti genome that are associated with G×G interactions influencing vector competence for dengue viruses. We developed an innovative genetic mapping strategy (Fig. 1) based on wild-type Ae. aegypti families that were experimentally exposed to four different dengue virus isolates (two DENV-1 isolates, designated as DV1-26A and DV1-30A, and two DENV-3 isolates, designated as DV3-10A and DV3-14A). The use of outbred families for genetic mapping was inspired from a validated study design previously developed to investigate the genetic basis of natural mosquito resistance to human malaria parasites [28], [29]. To simulate a natural situation, we used naturally circulating virus isolates contemporaneous with the mosquitoes that were obtained from the serum of human patients. Their complete genome sequence confirmed that they were genetically distinct (Fig. S1). Genetic mapping was based on a set of microsatellite markers distributed across the Ae. aegypti genome, which consists of three chromosomes (Fig. S2). With one marker every 9.0 centiMorgans (cM) on average, marker density was entirely adequate for chromosomes 1 and 3. For chromosome 2, however, the paucity of valid and/or informative microsatellites resulted in poor coverage (1 marker every 23.4 cM). Therefore, we focus here on chromosomes 1 and 3 and provide mapping results for chromosome 2 as supplementary data.
Our genetic mapping strategy allowed us to detect genetic linkage (i.e., non-independence between marker allele segregation and phenotype) at two different levels for each marker. The first level measured the dependence of the phenotype on the mosquito genotype regardless of the virus isolate (i.e., the main host genotype effect across virus serotypes and isolates). The second level measured the dependence of the phenotype on the genotype conditional on the virus isolate (i.e., the interaction between virus isolate and mosquito genotype, a measure of G×G interactions). The methodology of our genetic survey (Fig. 1) differs significantly from conventional genetic mapping strategies because it does not rely on controlled crosses between inbred lines that represent extremes of a trait. Although conventional strategies maximize QTL detection power, they are not best suited to identify multi-allelic QTL naturally segregating within unmanipulated populations [30], [31]. The large number of progeny produced by a single parental pair of mosquitoes can be used as outbred families that are suitable for QTL mapping [28], [29].
Vector competence was scored 14 days after an infectious blood meal according to three distinct phenotypes: (i) the proportion of mosquitoes that developed a midgut infection, (ii) the proportion of infected mosquitoes in which infection disseminated from the midgut to head tissues, and (iii) the infectious viral titer in virus-infected head tissues. Midgut infection and viral dissemination are prerequisites for virus transmission by mosquito bite [32]. Infectious titer of disseminated virus is used as a proxy for transmission potential [33]. All phenotypes were based on detection of infectious virus by standard plaque assay.
A total of 2,084 Ae. aegypti females from nine independent isofemale families (mean sample size per family: 232; range: 104–403) were individually phenotyped and genotyped (Table S1). Five of the families yielded at least one QTL statistically significant at the genome-wide level for the midgut infection phenotype (Fig. 2). Significant linkage at the genome-wide level was detected on chromosome 1 at marker 71CGT1 (29.7 cM) in family C01 (genome-wide p-value = 9.44×10−4) and family 5 (p = 2.9×10−2), at marker 335CGA1 (38.2 cM) in family C01 (p = 5.55×10−4), and at marker 88CA1 (44.9 cM) in family 7 (p = 4.94×10−3) and family 54 (p = 4.0×10−2). Linkage was also detected on chromosome 3 at marker 301ACG1 (0.0 cM) in family 51 (p = 7.47×10−5) and at marker B19 (13.6 cM) in the same family (p = 5.22×10−3). The proportion of phenotypic variation explained by each significant marker ranged from 3.5% to 12.0%. Importantly, we also detected significant virus isolate-specific linkage on chromosome 3 at marker 301CT1 (0.0 cM) in family 5 (p = 1.95×10−2, Fig. 2D). In this family, the proportion of infected females varied significantly among 301CT1 genotypes, but the genotype-phenotype relationship differed between virus isolates (Fig. S3). This isolate-specific genotype-phenotype association is interpreted as a G×G interaction between the mosquito and the viral genomes. An underlying assumption is that the isolate effect is primarily driven by genetic differences among isolates. When the isolate was replaced by the corresponding blood meal titer in the analysis, the interaction effect was no longer statistically significant (p = 0.083), which ruled out that uncontrolled variation in infectious dose among virus isolates (Table S2) might have confounded our interpretation of the isolate effect.
Significant linkage at the genome-wide level was detected in two of the nine families for the viral dissemination phenotype (Fig. 3). Linkage was significant on chromosome 1 at marker 335CGA1 (38.2 cM) in family J07 (p = 3.08×10−2) and family 42 (p = 3.1×10−2) and on chromosome 3 at marker 69TGA1 (32.1 cM) in family J07 (p = 4.4×10−2). The proportion of phenotypic variation explained by each significant marker ranged from 16.5% to 22.6%. Marker 335CGA1 on chromosome 1 was in linkage with the dissemination phenotype in two different families. In family J07 the marker effect was general across virus serotypes and isolates (Fig. 3A), whereas in family 42 it was isolate-specific (Fig. 3B). To verify that the isolate effect was not confounded with an effect of the infectious dose, we confirmed that the isolate by genotype interaction in family 42 was no longer statistically significant when the isolate was substituted by the blood meal titer (p = 0.287). For illustration, Fig. 4 shows the genotype-phenotype correlation for each virus isolate at marker 335CGA1 (the allele segregation pattern at this marker is shown in Fig. S4). Although marker genotype 439/439 confers protection against viral dissemination of isolates DV3-10A and DV3-14A, it does not have a detectable effect against isolates DV1-26A and DV1-30A. It is worth noting that because isolates DV3-10A and DV3-14A belong to DENV-3 whereas isolates DV1-26A and DV1-30A belong to DENV-1, in this particular case the effect could be serotype-specific rather than isolate-specific.
Significant linkage at the genome-wide level was detected in three of the nine families for the head titer phenotype (Fig. 5). Linkage was significant on chromosome 1 at marker 88CA1 (44.9 cM) in family 51 (p = 3.24×10−3). Linkage was also detected on chromosome 3 at marker 17ATA1 (22.4 cM) in family J07 (p = 1.70×10−5), at marker 69TGA1 (32.1 cM) in family J07 (p = 4.16×10−3), at marker 201AAT1 (57.1 cM) in family J06 (p = 5.18×10−4), and at marker 470CT2 (64.2 cM) in family J07 (p = 1.35×10−2). The proportion of phenotypic variation explained by each significant marker ranged from 8.9% to 75.6%. The genotype-phenotype association was isolate-specific at marker 201AAT1 in family J06 and at marker 470CT2 in family J07. Again, substituting the isolate by the corresponding blood meal titer ruled out a confounding effect of the infectious dose because the interaction was no longer statistically significant at marker 201AAT1 (p = 0.434) or at marker 470CT2 (p = 0.130). For illustration, Fig. 6 shows the genotype-phenotype correlation for each virus isolate at marker 201AAT1. Although marker genotype 338/338 confers protection against viral dissemination of isolates DV3-14A and DV1-26A, it results in increased head titer of isolate DV1-30A and no detectable effect against isolate DV3-10A. In this case the effect is truly isolate-specific (not serotype-specific) because isolates DV3-14A and DV1-26A (DENV-3 and DENV-1, respectively) share the same pattern whereas isolates DV1-26A and DV1-30A (both DENV-1) display opposite patterns. The isolate-specific genotype-phenotype correlation at marker 470CT2 is shown in Fig. S5.
Supporting information includes genetic mapping results for chromosome 2 (Fig. S6, S7, S8) and for families that did not produce any significant linkage (Fig. S9, S10, S11).
Our genetic survey demonstrates that G×G interactions between dengue viruses and mosquito vectors can be assigned to physical regions of the mosquito chromosomes. To the best of our knowledge, this is the first study to successfully locate G×G interactions in an invertebrate genome by marker-based genetic mapping. In agreement with the conclusions of a previous meta-analysis [17], we provide empirical evidence that the genetic architecture of host resistance depends on the pathogen strain. We establish the existence of tangible host genetic factors underlying G×G interactions in a natural invertebrate host-pathogen system. This is a critical first step towards their identification and characterization.
This study also provides important new information on the biology of dengue virus transmission in a natural situation. Phenotypic variation in the ability of field Ae. aegypti populations to serve as vectors of dengue viruses was previously observed [20], [21]. Genetic selection experiments [22] followed by QTL mapping studies using inbred selected lines [23]–[25] demonstrated a genetic basis for Ae. aegypti susceptibility to dengue virus infection and dissemination. Here, we provide direct evidence that a significant portion of natural phenotypic variation is genetically determined. We identify multiple genetic factors that control dengue susceptibility in a natural Ae. aegypti population, but show that the effect of these factors also depends on the virus genome.
Irrespective of G×G interactions, the relatively large proportion of phenotypic variation explained by the individual mosquito markers (up to 75.6%) reveals the existence of QTL with major effects. Interestingly, QTL underlying the midgut infection phenotype explained a smaller proportion of the phenotypic variation than QTL underlying the viral dissemination and dissemination titer phenotypes, suggesting a different genetic architecture. This hypothesis is supported by a similar observation in an earlier QTL mapping study [23]–[25]. Alternatively, this could be due to differences in marker informativeness or because exclusion of uninfected mosquitoes (on average, 57.5% of mosquitoes were uninfected in each family) for analysis of dissemination reduces the contribution of other QTL to overall phenotypic variation. Genetic linkage observed in different mosquito families could result from distinct loci or different allelic variants of the same locus. Based on the present data, we show that midgut infection by dengue viruses is controlled by at least two QTL in this wild Ae. aegypti population. In infected mosquitoes, viral dissemination from the midgut to secondary tissues is also controlled by two or more QTL. Infectious titer of disseminated virus, a proxy for transmission potential [33], is governed by three or more QTL.
Our mapping strategy relies on marker-by-marker tests and does not generate a confidence interval of the QTL location on the chromosomes. In other words, conventional techniques of interval mapping cannot be applied. Therefore, we cannot ascertain at this stage whether QTL identified on chromosomes 1 and 3 match those previously mapped for a DENV-2 strain in laboratory systems. On chromosome 1, a midgut infection QTL was previously identified at 19 cM [25] and a dissemination QTL at 31 cM [23]. On chromosome 3, a dissemination QTL was previously identified between 44 and 52 cM [23], [24]. No QTL was reported at the extremities of chromosome 3 in earlier studies. In the present study, significant linkage detected in the vicinity of the sex-determining locus (38.0 cM on chromosome 1) in four different families for the infection phenotype (Fig. 2A, 2C, 2D, 2E), in two families for the dissemination phenotype (Fig. 3A, 3B), and in one family for the head titer phenotype (Fig. 5C), could point to a major gene, or cluster of genes, controlling mosquito-virus interactions. Another important limitation of our marker-by-marker mapping strategy is that epistatic interactions between mosquito loci could not be measured. Intragenomic epistasis is a major component of the genetic architecture of quantitative traits [34], including host susceptibility to pathogens [17]. It is recognized as an essential determinant of the structure and evolution of complex genetic systems [35].
The main innovation of our study design was to explicitly account for viral genetic diversity in the genetic mapping of mosquito susceptibility loci. This allowed detection of both generalist and isolate-specific susceptibility loci. Several of the significant markers were in linkage with the phenotype independently of the virus isolate. Thus, the genetic basis of Ae. aegypti susceptibility to dengue viruses comprises a generalist component that is effective against diverse isolates, including isolates belonging to different serotypes. This result was previously unknown and gives hope to identify antiviral genes that confer a generalist protection against a diverse array of viruses. On the other hand, our genetic survey detected an isolate-specific component of the mosquito genetic basis for dengue susceptibility, which we interpret as G×G interactions between the vector and the virus. Markers associated with G×G interactions explained a significant proportion of phenotypic variation (from 7.8% to 16.5%). Identification of QTL associated with G×G interactions rules out the possibility that genetic specificity in this system is solely driven by environmentally inherited symbiotic microbiota that could have been confounded with the host genotype [16]. Note that this does exclude an indirect role of microbiota because the type of microbiota itself might be controlled by the host genotype.
It will be interesting to carry out fine-scale mapping experiments to identify the causal polymorphisms and their allelic profiles in the genomic regions where significant markers were found. An extension of the same protocol could be used to generate outbred isofemale lines beyond the F2/F3 generations to increase mapping resolution and locate candidate genes. Although several resistance mechanisms have been characterized in laboratory systems, mosquito genes underlying phenotypic variation in susceptibility to dengue viruses in nature have remained elusive. Leading candidates are genes known to be functionally involved in Ae. aegypti antiviral defense, including genes of the RNA interference (RNAi), JAK-STAT and Toll pathways [36]–[38]. A key gene of the RNAi pathway was recently associated with G×G interactions in this system [39]. The extremely low frequency (∼0.1%) of dengue virus infected Ae. aegypti in nature [40] and the relatively modest fitness cost of infection [41] make it unlikely that occasional challenge by dengue viruses is a strong enough selective pressure to drive the evolution of these genes. Rather, we speculate that their evolutionary dynamics are shaped by their concomitant role in the response to more prevalent pathogens in wild mosquito populations [42]. Conversely, natural selection of viruses that are able to evade or suppress resistance mechanisms is more likely to occur. Selection for enhanced transmission by mosquitoes has been proposed as a possible mechanism of adaptive evolution in dengue viruses [33].
Our results have at least two practical implications for the current development of novel strategies to interrupt virus transmission by genetically engineering resistant mosquitoes [43], [44]. First, the observation that Ae. aegypti vector competence for dengue viruses is controlled by multiple segregating QTL in a natural population suggests that such strategies may need to knock-down a larger number of genes than previously thought to confer complete resistance. Second, our discovery that the effect of several QTL is dengue virus serotype- and/or isolate-specific highlights the requirement for engineered resistance to be effective across all possible virus serotypes and strains encountered in nature.
In conclusion, our findings reinforce the idea that contributions from different genomes to the genetic architecture of ecological interactions cannot be fully disentangled because they depend on one another. By analogy with epistasis within the genome of a single organism, whereby the effect of a particular genotype on the phenotype depends on the genetic background, the direction and/or magnitude of the effect of host genes may depend on the pathogen genetic make-up. Like epistasis [45], [46], such G×G interactions between the genomes of two (or more) interacting organisms may constitute a significant component of the genetic architecture of complex traits resulting from ecological interactions. This may be true not only for antagonistically interacting organisms such as hosts and pathogens, but also for mutualistic interactions between, for example, animals and their gut microbiota or plants and their root microbiota [47], [48]. Accounting for the contribution of such genetic interactions between genomes will advance our understanding of the full genetic architecture of complex interaction traits in nature.
Wild mosquito eggs were collected using ovitraps in several households in the Nhong Pling, Kon Tee, Mae Na Ree, Nhong Ping Kai, and Thep Na Korn subdistricts, Muang district, Kamphaeng Phet Province, Thailand, during May 2010 and February 2011. Kamphaeng Phet Province is an agrarian, sparsely populated area located approximately 350 km northwest of Bangkok where dengue is endemic and the four dengue virus serotypes co-circulate [49]. All collections were made in rural villages located within a localized area of less than 850 km2. F0 eggs were brought back to the AFRIMS laboratory in Bangkok and allowed to hatch in filtered tap water. F0 pupae were separated and allowed to emerge in individual vials. Aedes aegypti adults were identified by visual inspection.
Single F0 pairs consisting of one virgin male and one virgin female were allowed to mate for 2–3 days following emergence. To avoid that F0 parents were siblings from the same wild mother, the male and the female of each pair were chosen from different collection sites. Inseminated females were offered daily blood meals and allowed to lay eggs. Egg batches from a single female were merged to obtain a pool of F1 eggs. F0 males and females were saved for later DNA extraction and typing. F2 and F3 families were produced by mass sib-mating and collective oviposition from the F1 offspring. Although the mass-mating step reduces statistical power to detect genetic linkage because parental genetic information is partially lost, it is traded for a considerable increase in sample size [28]. A single Ae. aegypti pair can produce several thousands progeny per generation after as few as 2–3 generations in the laboratory.
F1 adults were allowed to emerge in the laboratory, mate randomly, and feed on defibrinated sheep blood (National Laboratory Animal Center, Mahidol University, Bangkok, Thailand) through a membrane feeding system. The F2 and F3 eggs were collected and stored on dry pieces of paper towel and maintained under high humidity no longer than 6 months.
Although most Ae. aegypti females are inseminated by a single male in nature [50], using single pairs of newly emerged mosquitoes instead of naturally inseminated females allowed us to genotype both F0 parents prior to phenotyping. Families are not equal in the information they bring to QTL detection. Only families with the highest proportion of polymorphic markers were retained for genetic mapping. The aim of choosing families was to maximize the number of informative (i.e., segregating) meiosis at both marker and susceptibility loci. Out of a total of 184 initial mating pairs, nine families were selected that had >3,000 F2/F3 eggs and >80% polymorphic markers.
Four low-passage dengue virus isolates (two DENV-1 and two DENV-3) were used to orally challenge mosquitoes in vector competence assays (Table S2). They derived from serum samples collected between March and July 2010 during routine surveillance for diagnostic public health testing at AFRIMS from clinically ill dengue patients attending Kamphaeng Phet Provincial Hospital. Phylogenetic analysis assigned the viruses to known lineages of DENV-1 and DENV-3 that were circulating in Southeast Asia in the previous years (Fig. S1). Each isolate was amplified twice in Aedes albopictus cells (C6/36, ATCC CRL-1660), which is the minimum required to obtain a viral titer sufficiently high to infect mosquitoes orally using an artificial blood meal. To prepare virus stock, 0.2 ml of human serum was inoculated onto 2-day-old confluent C6/36 cells in a 25-cm2 flask and incubated for 7 days at 28°C. The virus-infected cell culture supernatant was harvested and inoculated into a fresh flask of 2-day-old C6/36 cells for the second passage, of which supernatant was aliquoted and stored at −70°C.
Viral genomic RNA was extracted from viral stock with the QIAamp viral RNA kit (Qiagen, Valencia, CA, USA). RT-PCR was performed using the SuperScript One-Step RT-PCR kit with platinum Taq polymerase (Invitrogen Life Technologies, Carlsbad, CA, USA) according to the manufacturer's recommendations, with a set of primers covering the entire genome (Table S3). RT-PCR products were purified by ultrafiltration. Sequencing reactions were performed using the Big Dye Terminator v1.1 cycle sequencing kit (Applied Biosystems, Foster City, CA, USA). Sequence chromatograms from both strands were obtained on an automated sequence analyzer ABI3730XL (Applied Biosystems). For sequence analysis, contig assembly and sequence alignments were performed using BioNumerics v6.5 (Applied-Maths, Sint-Martens-Latem, Belgium; www.applied-maths.com). Phylogenetic relationships were inferred using the maximum-likelihood method with the Tamura-Nei model implemented in MEGA v5 [51]. Reliability of nodes was assessed by bootstrap resampling with 1,000 replicates. The complete viral genome sequences were deposited to the GenBank database (accession numbers HG316481–HG316484).
Ae. aegypti females of the F2 or F3 generation were used in vector competence assays to score their relative susceptibility to the four low-passage dengue virus isolates. Experimental infections were run in three large experiments that involved different triplets of mosquito families (Table S1). F2/F3 eggs were hatched synchronously by placing them under low pressure for 30 min. Larvae were reared in 24×34×9 cm plastic trays filled with 2.0 liters of filtered tap water at a density of approximately 200 first instars per tray and fed a standard diet of approximately 1.0 g of fish food pellets (C.P. Hi Pro; Perfect Companion Group Co. Ltd., Bangkok, Thailand) per tray. Pupae were transferred to plastic screened 30×30×30 cm cages (Megaview Science Education Service Co. Ltd., Taichung, Taiwan) and adults were maintained on a diet of 10% sucrose. They were kept in an insectary at 28±1°C, under a relative humidity of 70–80% and a 12∶12 h light-dark cycle. The day before the oral challenge, females were transferred from the rearing cage to 1-pint feeding cups of ∼100 females.
Prior to experimental infections, 25-cm2 flasks of 2-day-old C6/36 cells were inoculated with a 1-ml aliquot from the viral stock and incubated for 45 min to 1 hour. At the end of the adsorption, 4.0 ml of maintenance medium were added and the cells were incubated at 35±1°C under 5% CO2 for 5 days. At day 5, 1.0 ml of heat-inactivated fetal bovine serum containing 15% of sodium bicarbonate 7.5% solution (HIFBS-NaHCO3) was added to the virus-infected cell culture supernatant, which was then harvested to prepare the infectious blood meal. The virus suspension was diluted 1∶3 or 1∶2 with RPMI 1640 medium containing 5% HIFBS and then mixed 1∶1 with defibrinated sheep blood (National Laboratory Animal Center). The infectious blood meal was placed in water-jacketed glass feeders maintained at a constant temperature of 37°C and covered with a piece of desalted porcine intestine. Four- to 7-day-old Ae. aegypti females deprived of sucrose and water for 24 h prior to blood feeding were offered an infectious blood meal for 30 min. Samples of the blood meals were saved for subsequent titration. Blood meal titers ranged from 2.0×104 to 1.5×106 plaque-forming units per ml (PFU/ml); the majority (83.3%) ranged between 1.0×105 and 1.0×106 PFU/ml (Table S2). Small differences in blood meal titers contribute to the isolate effect in the analysis, but we verified that it did not confound our interpretation (see below). After blood feeding, mosquitoes were briefly sedated with CO2 from dry ice, and fully engorged females were transferred to clean 1-pint paper cups. Unfed or partially fed females were discarded. Engorged females were maintained for 14 days at 28±1°C, under 70–80% relative humidity and a 12∶12 h light-dark cycle and provided cotton soaked with 10% sucrose ad libitum.
Vector competence was scored in the F2/F3 families at 14 days after the infectious blood meal according to three phenotypes: (i) midgut infection, (ii) viral dissemination from the midgut, and (iii) infectious titer in head tissues. Viral infection of midgut epithelial cells and subsequent dissemination to secondary tissues are two essential steps of dengue virus propagation in Ae. aegypti. Both events are prerequisites for virus transmission by mosquito bite and have been used to define a ‘midgut infection barrier’ and a ‘midgut escape barrier’ underlying Ae. aegypti susceptibility to dengue viruses [32]. These two vector competence indices were determined qualitatively (i.e., presence or absence of virus in mosquito bodies and heads, respectively). Although both phenotypes are binary traits (all-or-nothing), they are assumed to be consistent with a multifactorial basis and to result from continuous variation on an underlying (unobserved) scale. Infectious titer of virus disseminated to head tissues is strongly correlated with the probability to detect virus in saliva samples collected in vitro [33], and is therefore used as a proxy for transmission potential. Head titers were determined quantitatively by end-point titration.
Upon harvest, the head of each female was cut off on a chill table and placed individually in 500 µl of mosquito diluent (MD; RPMI 1640 medium with 10% HIFBS, 100 units/ml penicillin, 100 µg/ml streptomycin and 100 units/ml L-Glutamine). The remainder of the body (thorax and abdomen) was stored separately in 900 µl of MD with one 4.5 mm stainless steel bead in a 2-ml safe-lock tube. Samples were stored at −70°C until testing by plaque assay. They were quickly thawed in a water bath at 35±2°C and homogenized in a mixer mill (Qiagen) at 24 cycles/sec for 2 min. Four hundreds µl of each body homogenate were transferred into a new 1.5 ml safe-lock tube containing 400 µl of lysis buffer BQ1 (Macherey-Nagel, Düren, Germany) and stored at −20°C for DNA genotyping.
Infectious virus was detected and quantified by plaque assay performed in rhesus monkey kidney epithelial cells (LLC-MK2, ATCC CCL-7) as previously described [52]. Briefly, the homogenized body and head samples were filtered individually through a sterile, syringe-mounted 0.22-µm membrane. The samples were placed in an ice bath, 100 µl/well were inoculated onto a monolayer of 3-day-old LLC-MK2 cells in 24-well plates. The virus was adsorbed at room temperature (20–28°C) on a rocker platform for 90 min. The inoculum was removed and 0.5 ml/well of a first overlay of medium was added. The cells were incubated for 5 days at 35±1°C under 5±0.5% CO2. The cells were stained with a second overlay of medium containing 4% neutral red (Sigma Chemical Co., Perth, WA, USA). Mosquito infection and dissemination status was determined based on the presence of plaques in their body and head homogenates, respectively. Mosquito whose bodies were negative by plaque assay were considered uninfected, and their heads were not processed further. Head titer of infected bodies was determined by plaque assay of 1∶10 and 1∶100 dilutions of head homogenates.
QTL detection was performed in the outbred mosquito families using a set of 25 microsatellite markers broadly distributed across the genome (Fig. S2). Genetic position and PCR primers sequences for these markers were readily available from published literature [53], [54] with the exception of markers 210TTC1 and 14ATT1 that we developed (see below) in an attempt to increase chromosome 2 coverage. In our Ae. aegypti population, few existing chromosome 2 markers were valid and/or informative, and despite our efforts to find additional markers, coverage remained too low to provide a sufficient mapping density of markers. The paucity of unique sequences among supercontigs mapped to chromosome 2 made it extremely difficult to design primer pairs resulting in unique PCR products. Efforts are currently being made to develop alternative markers based on single nucleotide polymorphisms (SNPs). For each marker in the final map (Fig. S2), we verified that the pair of primers matched a unique supercontig of the unassembled Ae. aegypti genome [55], which in turn was anchored to the reference genetic map [56] by the co-presence of another marker with known genetic position that uniquely matched the same supercontig. The only exception is marker B19 that falls in an unmapped supercontig but was independently assigned to chromosome 3 by linkage analysis [53]. The 25 microsatellites represent 18 distinct genetic positions along the Ae. aegypti genome. Twenty-two of these microsatellites (15 genetic positions) are located on chromosomes 1 or 3. Based on an estimated genome size of 1,376 Mbp and a genetic size of 205 centiMorgans (cM), the relationship between physical and recombination distance is 6.71 Mbp/cM [55], [56]. Estimated genetic sizes of chromosomes 1 and 3 are 70.6 and 64.2 cM, respectively [56]. For these two chromosomes, adjacent markers in our genetic survey were separated by an average distance of 9.0 cM (60.3 Mbp). Thus, an unknown QTL was on average less than 4.5% recombination away from a marker.
The genetic survey was based on the analysis of outbred mosquito families at the F2 or F3 generation. Each mosquito family descended from a single pair of F0 parents collected in the field, providing an independent sample of up to four different alleles per locus from the original natural mosquito population. Based on the number of alleles present at the F0 generation, we verified at each marker that the correct number of genotypes was observed in the progeny. Three, six and ten different genotypes are expected in the progeny when F0 parents harbor two, three and four different alleles, respectively.
The originality of the strategy is to use families with incomplete pedigree information due to the mass-mating step [28]. Mosquitoes are classified according to their genotype so that identity by state (IBS) is used as a surrogate for identity by descent (IBD). Genetic linkage is not inferred from allele sharing proportions but from genotype-phenotype associations. Therefore, allele segregation in Mendelian proportions is not required by the study design. During mass mating and collective oviposition allele frequencies may be distorted because of random genetic drift or natural selection. Genetic drift is particularly likely to occur at the F1 generation because the number of reproducing adults is relatively small. Some genotypes could also be selected because they have a fitness advantage over other genotypes in insectary conditions. Departure from a neutral reproductive model may reduce the statistical power to detect marker-trait associations, but not the statistical significance of results. The same is true for null alleles or genotyping errors that would confound the observed genotypes. Our genetic model does not specify allelic codominance or recessivity. It simply compares genotypes (or groups of genotypes if a null allele segregates) regardless of their frequency.
Statistical power is also limited by the extent of heterozygosity in the family. There is no guarantee that every F0 parent is heterozygous both at a QTL and at a linked segregating marker, which is a prerequisite to generate a marker-trait association in the progeny. We maximized statistical power by genotyping F0 parents and choosing the most informative families (i.e., with >80% of markers being polymorphic) for phenotyping. In addition, the linkage phase between the marker and the QTL can vary in the progeny. This can reduce QTL detection power, if for example the same marker allele is associated with different QTL alleles in the F0 parents. Again, this would increase the probability to declare significant evidence against marker-trait association (i.e., in support of the null hypothesis) but not the statistical significance of results.
Microsatellite markers 210TTC1 and 14ATT1 on chromosome 2 were developed as previously described [54]. Briefly, supercontig sequences containing genetic markers mapped to chromosome 2 were retrieved from VectorBase (http://aaegypti.vectorbase.org/) and submitted to the Tandem Repeats Finder program [57] using default parameters with the exception of a maximum period size of 3. For tandem repeats with a consistent motif and a repeat copy number <30, a ∼500 bp sequence encompassing the microsatellite was subjected to BLASTn analysis against the Ae. aegypti genome in VectorBase to verify their occurrence in single copy. PCR primers were designed in flanking sequences of selected microsatellites using Primer3 v0.4.0 [58], with an amplicon size target of 100–500 bp in length. The primer sequences were 5′-TCATTCCCAGTACCACACAAACG-3′ (forward) and 5′-ACTCGTTACTGGATGTGCTATCCC-3′ (reverse) for marker 14ATT1 and 5′-GAACGCGCTCGTAAGCGAGA-3′ (forward) and 5′-CACTGTGCGTTGGTTTCGGCT-3′ (reverse) for marker 210TTC1. Individual primer pairs were further subjected to BLASTn analysis to verify that they were predicted to amplify single copy sequences in the Ae. aegypti genome. PCR products were run by electrophoresis on 2% agarose gel to confirm that amplicons were unique.
Genomic DNA was extracted from mosquito homogenates using the NucleoSpin 96 Tissue Core Kit (Macherey-Nagel) and stored at −20°C until use. Genotyping of microsatellite repeats was performed by PCR amplification using fluorochrome-labeled forward primers (5′-FAM, 5′-HEX or 5′-ATTO550) (Eurofins MWG Operon, Ebersberg, Germany) to generate fluorescent PCR products. Primer pairs producing different amplicon sizes were assembled into multiplex groups of 4–6 markers. Amplification was performed in 25 µl volumes in Thermo-Fast 96-wells PCR plates (ABgene, Epsom, Surrey, UK) in a Veriti thermal cycler (Applied Biosystems). Each reaction contained 1× Taq buffer (50 mM KCl, 20 mM Tris pH 8.4) (Invitrogen Life Technologies), 1.5 mM MgCl2, 200 µM dNTPs (Invitrogen Life Technologies), 0.2 µM of each primer, 1 unit of Taq DNA polymerase (Invitrogen Life Technologies), and 2 µl of genomic DNA purified as described above. Thermocycling conditions were 5 min at 94°C, followed by 35 cycles of a 30-sec denaturation at 94°C, a 30-sec annealing at 50°C, and a 1-min extension at 72°C, followed by a 7-min final extension at 70°C. Multiplexed PCR products were examined by electrophoresis on 1% agarose gel and diluted 1∶10 in sterile water. Two µl of this dilution was added to 10 µl of Hi-Di Formamide (Applied Biosystems) containing 7.5% of 6-carboxy-X-rhodamine (ROX)-labeled Geneflo 625 size standards (EurX, Gdansk, Poland). Capillary electrophoresis of multiplexed PCR products was performed on a 3730xl DNA Analyser (Applied Biosystems). Sizes of microsatellite alleles were called and manually checked using the GeneMapper v4.0 software package (Applied Biosystems).
Our approach is a combination of linkage and association analyses. Linkage analysis generally uses pedigrees to infer the location of a susceptibility locus based on coinheritance of the disease phenotype with genetic markers whose chromosomal location is known. Association analysis does not rely on pedigree structure but assumes that strong associations between marker alleles and disease phenotype in a population will be due to linkage, rather than chance. In association studies, IBD due to coancestry is inferred from IBS in the form of observed allelic associations. In the present study, linkage was inferred from IBS as in association studies. Tests of genotype-phenotype associations, however, were performed in sibships (single-generation families) at the at the F2 or F3 generation. In contrast with association studies performed at the population level, high linkage disequilibrium in the families strongly reduces the marker density required for the genetic mapping.
Genetic linkage was inferred from the significance of the genotype effect in a generalized linear model of the phenotype that included the factors mosquito genotype, virus isolate and their interaction as explanatory variables. Response variables were the three vector competence indices that we measured: (i) midgut infection status, (ii) viral dissemination status of midgut-infected mosquitoes, and (iii) head titer in mosquitoes with a disseminated infection. For binary phenotypes (infection and dissemination), the model was fitted with a binomial error structure and a logit link function (i.e., a logistic regression). For the continuous phenotype (head titer), the variable was log-transformed and the model was fitted with a normal error distribution and an identity link function (i.e., a linear regression). The model was fitted separately for each informative microsatellite marker in each mosquito family. Depending on the number of alleles of the marker, the factor genotype had from three to ten different categories, whereas the factor isolate always had four categories (i.e., the four isolates used in the study). If, due to random sampling effects in the progeny, one category of the genotype was not encountered in one or more categories of the isolate, this genotype category was excluded from the analysis so that the genotype by isolate interaction could be tested in the model. Depending on the marker, this could result in a different number of mosquitoes included in the analysis for the same family.
Statistical significance of the genotype effect or the genotype by isolate interaction effect in the above model was determined differently for binary (infection and dissemination) and continuous (head titer) variables. For binary phenotypes, statistical significance was tested with an analysis of deviance [59]. The deviance measures the unexplained variation of the data for a given model. The difference between the deviances of two models measures whether the two models fit the data differently. We first tested whether a model with the factors isolate and genotype fitted the data significantly better than a model with only the isolate (i.e., testing whether the genotype is a significant predictor of the phenotype). Then we tested whether a model with isolate, genotype and genotype by isolate interaction fitted the data better than the model with only the main effects of isolate and genotype (i.e., testing whether the interaction is a significant predictor of the phenotype). To estimate the proportion of variation explained by a significant factor we compared the mean deviance (deviance divided by the number of degrees of freedom) of the model including the factor and the mean deviance of the model without the factor. For the continuous phenotype, statistical significance was tested with an analysis of variance. To estimate the proportion of variation explained by a significant factor we followed the approach described above for the binary phenotypes. We compared the residual variance (sum of squares divided by the number of degrees of freedom) of the model including the factor and the residual variance of the model without the factor.
Because we performed multiple tests for each mosquito family, we used a Bonferroni correction of the p-values to ensure a genome-wide type I error of at most α = 0.05 (i.e., no more than 5% false positives overall). The genome-wide significance level of the test at each marker was α/N, where N is the number of informative markers tested in each family. A genotype-phenotype association was declared significant at the genome-wide level if the nominal p-value was smaller than α/N. When a significant genotype by isolate interaction was found, we verified that uncontrolled differences in the infectious titer of the artificial blood meal (Table S2) did not confound our interpretation of the factor isolate as an approximation of viral genetic identity. We performed an analysis based on the same model as previously but replacing the isolate by the corresponding blood meal titer (log-transformed). If the isolate effect were only due to differences in blood meal titer, we expect that the effect would remain statistically significant. Conversely, if the effect became insignificant, it would mean that the isolate effect resulted primarily from an effect of the viral genetic polymorphism rather than a simple effect of the infectious dose.
All statistical analyses were performed in the statistical environment R [60].
|
10.1371/journal.pgen.1007464 | BMP signaling downstream of the Highwire E3 ligase sensitizes nociceptors | A comprehensive understanding of the molecular machinery important for nociception is essential to improving the treatment of pain. Here, we show that the BMP signaling pathway regulates nociception downstream of the E3 ubiquitin ligase highwire (hiw). hiw loss of function in nociceptors caused antagonistic and pleiotropic phenotypes with simultaneous insensitivity to noxious heat but sensitized responses to optogenetic activation of nociceptors. Thus, hiw functions to both positively and negatively regulate nociceptors. We find that a sensory reception-independent sensitization pathway was associated with BMP signaling. BMP signaling in nociceptors was up-regulated in hiw mutants, and nociceptor-specific expression of hiw rescued all nociception phenotypes including the increased BMP signaling. Blocking the transcriptional output of the BMP pathway with dominant negative Mad suppressed nociceptive hypersensitivity that was induced by interfering with hiw. The up-regulated BMP signaling phenotype in hiw genetic mutants could not be suppressed by mutation in wallenda suggesting that hiw regulates BMP in nociceptors via a wallenda independent pathway. In a newly established Ca2+ imaging preparation, we observed that up-regulated BMP signaling caused a significantly enhanced Ca2+ signal in the axon terminals of nociceptors that were stimulated by noxious heat. This response likely accounts for the nociceptive hypersensitivity induced by elevated BMP signaling in nociceptors. Finally, we showed that 24-hour activation of BMP signaling in nociceptors was sufficient to sensitize nociceptive responses to optogenetically-triggered nociceptor activation without altering nociceptor morphology. Overall, this study demonstrates the previously unrevealed roles of the Hiw-BMP pathway in the regulation of nociception and provides the first direct evidence that up-regulated BMP signaling physiologically sensitizes responses of nociceptors and nociception behaviors.
| Although pain is a universally experienced sensation that has a significant impact on human lives and society, the molecular mechanisms of pain remain poorly understood. Elucidating these mechanisms is particularly important to gaining insight into the clinical development of currently incurable chronic pain diseases. Taking an advantage of the powerful genetic model organism Drosophila melanogaster (fruit flies), we unveil the Highwire-BMP signaling pathway as a novel molecular pathway that regulates the sensitivity of nociceptive sensory neurons. Highwire and the molecular components of the BMP signaling pathway are known to be widely conserved among animal phyla, from nematode worms to humans. Since abnormal sensitivity of nociceptive sensory neurons can play a critical role in the development of chronic pain conditions, a deeper understanding of the regulation of nociceptor sensitivity has the potential to advance effective therapeutic strategies to treat difficult pain conditions.
| In spite of its clear medical importance, the molecular mechanisms of pain signaling remain poorly understood. Pain pathways in large part depend on sensory input from specialized sensory neurons called nociceptors [1]. Since the activation of nociceptors leads to pain sensation and the sensitization of nociceptors is thought to be a major contributor of pain pathogenesis, understanding the molecular mechanisms controlling nociceptor function is essential for improving the treatment of pain [2].
Drosophila melanogaster is a powerful model system for neurogenetic studies of nociception. Larval Drosophila show stereotyped behavioral responses to potentially tissue-damaging stimuli, such as noxious heat or harsh mechanical stimulation [3]. The most unambiguous larval nociception behavior involves a corkscrew-like rolling around the long body axis (termed nocifensive escape locomotion (NEL) or simply “rolling”). Since rolling is specifically triggered by noxious stimuli and is clearly separable from normal larval locomotion, the analysis of NEL provides a robust behavioral paradigm to study nociception. Class IV multidendritic (md) neurons are polymodal nociceptors that are necessary for thermal and mechanical nociception in larvae [4]. Optogenetic activation of the Class IV neurons is sufficient for triggering NEL [4,5]. Accumulating evidence in studies of fly nociception suggests that the molecular pathways of nociception are conserved between Drosophila and mammals [3,6–15].
To identify genes important for nociceptor function, we recently performed thermal nociception screens in which we targeted the RNAi knockdown of nociceptor-enriched genes in a nociceptor-specific manner [16]. In this screen, we found that two RNAi lines targeting highwire (hiw) caused driver dependent hypersensitivity in thermal nociception assays (revealed as a rapid response to a threshold heat stimulus) indicating a potential role for hiw as a negative regulator of nociceptor activity [16]. hiw is an evolutionally conserved gene encoding an E3 ubiquitin ligase, whose function has been implicated in various aspects of neuronal development, synaptic function, and neuronal degeneration [17]. However, in contrast, very little is known about the roles of hiw in sensory processing and in controlling behavior. Here, we present additional and more specific evidence that hiw plays an important role in the regulation of behavioral nociception and nociceptor sensitivity through the bone morphogenetic protein (BMP) pathway.
To further investigate the potential function of hiw in nociception that was suggested by our previous study, we tested mutants for a strong loss-of-function allele of hiw (hiwND8) in thermal nociception assays [18]. Unexpectedly, we found that genetic mutants of hiw showed insensitivity to a noxious temperature probe of 42 or 46°C, which was, surprisingly, the opposite of the previously described hiw RNAi phenotype (Fig 1A) [16]. Similar thermal insensitivity was also seen with other hiw alleles (S1 Fig). Although hiw is widely expressed in the nervous system [18], nociceptor-specific restoration of hiw expression rescued this insensitivity (Fig 1A), indicating that hiw function in nociceptors is sufficient for restoration of normal thermal nociception and the relevant site of action was in nociceptors.
Intrigued by the clear phenotypic distinction between genetic mutants and RNAi animals, we further dissected the nociception phenotype of hiw mutants by employing an optogenetic strategy. Optical activation of larval nociceptors via the blue light-gated cation channel Channelrhodopsin-2 (ChR2) is sufficient to induce larval NEL [4,5,19]. Since nociceptor activation by ChR2 circumvents receptor potential generation but still depends on the machinery essential for downstream signaling (Fig 1B), this technique has been utilized to distinguish genes that are important for primary sensory function from those that function in downstream aspects of signaling, such as action potential generation/propagation and/or synaptic transmission [10,20]. Using low intensity blue light (3.8 klux), which elicits NEL in about 20–30% of control animals expressing ChR2::YFP in nociceptors (Fig 1C), we found that the hiwND8 mutants had a significantly increased probability to show NEL, indicating that the mutant for this allele is hypersensitive in response to optogenetic activation of nociceptors (Fig 1C) even though it was insensitive in thermal nociception assays. Tissue specific rescue experiments again showed that nociceptor specific expression of hiw was sufficient to rescue this optogenetic hypersensitivity (Fig 1C). Taken together, these findings suggested that hiw has multiple, but dissociable, effects in the regulation of nociceptors. On the one hand, hiw regulated a sensory reception-dependent function causing insensitivity, but it also regulated a function downstream of sensory reception that caused hypersensitivity. Thus, the hypersensitivity seen in our earlier RNAi experiments is likely reflective of effects on the latter process.
To further examine hiw’s role, we tested the effects of expressing hiw△RING in nociceptors. The hiw△RING transcript encodes a mutated form of hiw lacking the RING domain that is responsible for E3 ligase activity [21,22]. This mutated protein has been proposed to function as a dominant-negative poison subunit in multimeric Hiw E3 ligase complexes. Similar to our original observations with hiw RNAi, expression of hiw△RING in nociceptors resulted in significant hypersensitivity in thermal nociception (Figs 1D and S2). This manipulation also caused hypersensitive optogenetic nociception responses (Fig 1E). As hiw encodes a large protein with many functional domains, and phenotypes of hiw mutants are known to show varied sensitivity to gene dosage [21], the observed similarity between hiw△RING overexpression and hiw RNAi is suggestive of dosage-dependent effects of hiw in nociceptors. For instance, the dominant negative approach may lead to an incomplete loss of function for hiw that is similar to the effects of RNAi.
It has been very recently shown that the canonical BMP pathway in nociceptors is required for nociceptive sensitization after tissue damage in Drosophila [23]. Since the BMP signaling pathway has also been proposed to be a downstream pathway regulated by Hiw in motoneurons [24], we tested whether the BMP signaling pathway is regulated downstream of Hiw in nociceptors. We first examined the level of phosphorylated Mad (pMad) in nociceptor nuclei by quantitative immunohistochemistry, which is an established method for evaluating the activation level of intracellular BMP signaling [25–31]. In nociceptor nuclei, hiw genetic mutants showed significantly elevated pMad levels (33%) in comparison to wild-type, even when processed together in the same staining solution (see also Materials and Methods) (Fig 2A, 2B and 2F). A similarly modest change in pMad accumulation in motor neuron nuclei is associated with effects on presynaptic function and morphology at the neuromuscular junction (NMJ) [32,33]. An increased accumulation of pMad in the nucleus and the cytoplasm was observed in nociceptors expressing hiw△RING (Fig 2C and 2F). Expression of wild-type hiw in nociceptors of hiw mutant animals rescued the elevated pMad level (Fig 2D and 2F). We also confirmed that our immunohistochemistry successfully detected the increase of nuclear pMad caused by expressing the constitutively active form of thick veins (tkvQD), which activates the intracellular BMP signaling cascade independently of BMP ligands [34] (Fig 2E and 2F). These data together suggest that BMP signaling is negatively regulated downstream of hiw in larval nociceptors. In the larval motoneurons, it is known that pMad signals can be locally detected at synaptic boutons as well as nuclei [26,35,36]. However, in our samples no detectable pMad signals were observed at synaptic terminals in larval nociceptors (Fig 2G).
Next, we tested whether up-regulated BMP signaling in nociceptors is responsible for the hypersensitive nociceptive responses caused by hiw loss-of-function. mad1 encodes a dominant-negative form of Mad with disrupted DNA-binding ability [37]. When mad1 was expressed together with hiw△RING in nociceptors, the hypersensitive phenotype that was normally induced by the expression of hiw△RING alone was not detected (Fig 2H). Since neither expressing Mad1 together with hiw△RING nor expressing Mad1 alone in nociceptors induced insensitivity to noxious heat (S3 Fig), these results indicate that hypersensitive nociception caused by weak hiw loss of function requires an intact BMP signaling pathway that normally operates through Mad. This result is consistent with the elevated pMad observed with hiw loss of function as playing a causal role in the hypersensitive phenotypes.
The MAP kinase (MAPKKK) wallenda (wnd) is a well-characterized target substrate of Hiw ligase [17]. Hiw negatively regulates the protein level of Wnd, and the Hiw-Wnd interaction is crucial for normal synaptic growth, but not for normal synaptic function in NMJ [31,38–40]. In addition, hiw interacts with wnd in Class IV neurons in the regulation of dendritic and axonal morphology [41]. In larval motoneurons, it has been suggested that wnd is not involved in the regulation of BMP signaling [31]. To test whether wnd is involved in the control of BMP signaling downstream of hiw in nociceptors, we examined a genetic interaction between hiw and wnd in double mutants. A wnd mutation in hiw mutant background did not suppress the elevated nuclear pMad level in nociceptors that we observed in the hiw mutant (Fig 3A–3D and 3F), nor did wnd single mutants show altered nuclear pMad accumulation relative to controls (Fig 3E and 3F). Interestingly, significant up-regulation of nuclear pMad signal was observed in nociceptors overexpressing wnd, but not with a kinase-dead version of wnd (S4 Fig). Taken together, these results suggest that elevated nuclear pMad in hiw mutant nociceptors does not depend on the activity of Wnd, although overexpression of wnd with GAL4/UAS can cause elevated BMP signaling in nociceptors.
To gain insight into which regions of Hiw protein are involved in attenuating BMP signaling in nociceptors, we performed an expression study of a series of Hiw dominant negatives with various deletions, which has been established by Tian et al. [39] (Fig 4A). Expressing HiwNT (N-terminal half of Hiw) caused a greater than 200% increase in nuclear pMad signals compared to controls (Fig 4B, 4C and 4H). HiwCT (C-terminal half of Hiw) and Hiw△RCC1 resulted in 99% and 68% increases in nuclear pMad signals, respectively (Fig 4D, 4E and 4H). HiwCT and Hiw△RCC1 also caused marked accumulation of pMad signals in the cytoplasm of nociceptors (Fig 4D and 4E), which was also observed with Hiw△Ring expression (Fig 2C). This cytoplasmic accumulation of pMad signals is unlikely due to technical variability of immunostaining since the control samples treated in the same staining solutions with HiwCT or Hiw△RCC1 never developed such accumulations and cells nearby the nociceptors showed the normal pMad signal. In contrast, Hiw△HindIII and HiwCT1000 (C-terminal only region of Hiw) did not cause any changes in nuclear pMad signals in nociceptors (Fig 4C, 4F and 4H). Thus, the attenuation of BMP signaling in nociceptors through Hiw appears to depend on different regions of Hiw from those that have been proposed to be involved in the regulation of NMJ morphology (Hiw△RCC1, and Hiw△HindIII function as dominant-negative in NMJ morphology while HiwNT and HiwCT1000 do not [39]). Because both HiwNT and HiwCT, which are largely non-overlapping N-terminal and C-terminal halves of Hiw, caused increased nuclear pMad signals, multiple regions of the Hiw protein must be intact for normal suppression of BMP signaling in nociceptors.
Although a previous study by Follansbee et al. suggests that the canonical BMP signaling pathway in larval nociceptors is a necessary component for nociceptive sensitization after tissue-damage, whether up-regulation of BMP signaling in nociceptors is sufficient to sensitize nociception has not been proven and potential mechanisms leading to sensitization are unknown. Because our data support the notion that the up-regulation of BMP signaling in nociceptors plays a key role in inducing sensitized nociception, we tested whether up-regulation of intracellular BMP signaling in nociceptors is sufficient to induce nociceptive hypersensitivity. In thermal nociception assays, animals expressing the constitutively active BMP receptor tkvQD in nociceptors did exhibit significant hypersensitivity (Figs 5A and S2), and tkvQD also caused hypersensitive responses in optogenetic nociception assays. The latter suggests that elevated BMP signaling in nociceptors was able to sensitize nociception through a mechanism that was downstream of sensory reception (Fig 5B). Although the dendritic structure of nociceptors in tkvQD overexpressing animals was not significantly altered (Fig 5C–5E), overexpression of tkvQD caused overextension and overexpansion of nociceptor axon termini (Fig 5F–5H). Combined, these data demonstrate that elevated BMP signaling in nociceptors is sufficient to sensitize thermal and optogenetic nociception behaviors in addition to causing increases in axon terminal branching.
Since nociceptor-specific up-regulation of BMP signaling sensitizes thermal and optogenetic nociception behaviors, we next explored whether the up-regulation of intracellular BMP signaling actually sensitizes physiological responses of nociceptors. To observe neuronal responses of larval nociceptors to a range of thermal stimuli, we developed a preparation for optical recording from axon terminals of the nociceptive neurons. We then observed these terminals while we locally applied a thermal ramp stimulus to the larval body wall (Fig 6A). To monitor Ca2+, the genetically encoded sensor GCaMP6m was expressed under the control of ppk-GAL4 [42]. In control animals we observed a steep increase of the GCaMP6m signal in nociceptors when the ramping temperature reached the 39–47°C temperature range (Fig 6B, 6B’ and 6D). We found that nociceptors expressing tkvQD showed a significantly greater increase of GCaMP6m signals through 36–50°C in comparison to those in controls (Fig 6C, 6C’ and 6D), while basal fluorescence levels of GCaMP6m (F0) were comparable between the control and tkvQD-expressing nociceptors (Fig 6E). These results suggest that the significantly greater increase of GCaMP6m signals observed in nociceptors expressing tkvQD is due to the greater level of Ca2+ influx triggered by the heat ramp stimulus, and not to unintended transcriptional upregulation of GCaMP6m. Thus, elevated BMP signaling in nociceptors results in exaggerated Ca2+ signals at the terminals of nociceptors in response to heat in the noxious range. This conclusion is consistent with the behavioral nociceptive sensitization induced by the same intracellular up-regulation of BMP signaling in nociceptors.
Chronic up-regulation of BMP signaling in nociceptors caused sensitization of behavioral nociception responses of larvae and an increased Ca2+ response of nociceptors to noxious heat, but also expansion of nociceptor terminals. To further separate the physiological and morphological effects of BMP up-regulation in nociceptors, we up-regulated BMP signaling during a shorter 24-hour time-window in larval stage. Using the temperature sensitive repressor of GAL4 activity (GAL80ts) [43], we activated expression of tkvQD in larval nociceptors by shifting ppk-GAL4 UAS-Chr2::YFP tub-GAL80ts animals to 30°C for 24 hours. We then tested these larvae for sensitized optogenetic nociception. The 24-hour induction of tkvQD induced hypersensitivity in the optogenetic nocifensive responses and also significantly increased nuclear pMad levels relative to controls (Fig 7A and 7B). However, no detectable axonal overgrowth was induced by 24-hour tkvQD expression (Fig 7C and 7D). Unfortunately, we were not able to investigate the effects of this manipulation on nociception responses with a 39°C thermal stimulus because the prolonged incubation at 30°C interfered with 39°C NEL behavior in both controls and experimental animals (S5 Fig). This latter finding indicates that the sensitivity of thermal nociception in Drosophila is modulated by the ambient temperature. Collectively, these data demonstrate that 24-hour activation of BMP signaling in nociceptors is sufficient to sensitize larval nociceptive response in the absence of detectable changes to axonal morphology. Taken together with our Ca2+ imaging results, these data suggest a physiological role for BMP signaling in the regulation of nociceptor sensitivity.
Identifying novel conserved molecular pathways that regulate nociception in model animals is a promising strategy for understanding the molecular basis of pain signaling and pain pathogenesis [44,45]. Using Drosophila, we found that both the E3 ligase Hiw and the downstream BMP signaling pathway play crucial roles in regulating nociceptor sensitivity.
The data we present in this study suggest that hiw has at least two distinct functions in the regulation of nociceptor sensitivity. We found that strong loss-of-function mutants of hiw showed insensitivity to noxious heat but hypersensitivity to optogenetic stimulation of nociceptors (Fig 1A and 1C). Since expressing wild-type hiw in nociceptors of hiw mutants rescued both phenotypes, loss of hiw in nociceptors is responsible for these two ostensibly opposing phenotypes (Fig 1A and 1C). We also found that nociceptor-specific expression of hiwRNAi or hiw△RING caused only hypersensitivity (Fig 1D and 1E) [16], indicating that the process that governs hypersensitivity is separable from the cause of insensitivity. As insensitivity was epistatic to hypersensitivity in thermal nociception assays, we used optogenetics to show that hypersensitivity is actually present in hiw genetic mutants as well as in previously described RNAi animals. The use of optogenetic stimulation of neurons allowed us to bypass the endogenous sensory reception step(s) and to reveal this role. Our data suggest that hiw is a) required for the negative regulation of a neural pathway that is downstream of sensory reception and b) required to confer normal sensitivity to noxious heat via sensory reception pathways. As strong hiw loss of function causes reduced dendritic arbors [41] while hiw RNAi does not [16], it is possible that the reduced dendrite phenotype accounts for the insensitivity of the strong hiw alleles. Consistent with this hypothesis, many manipulations that cause insensitive thermal nociception are associated with a reduction in the dendritic arbor [16]. The phenotypic difference between strong loss-of-function mutants and RNAi or Hiw dominant-negative animals suggests that insensitive and hypersensitive phenotypes observed in hiw mutants have different sensitivity to the dosage of hiw. This has also been seen in the larval motor neuron system where it has been demonstrated that two different phenotypes of hiw in larval NMJ (overgrowth of synaptic boutons and diminished synaptic function) are separable by their different sensitivity to the dosage of hiw [21].
Our data also suggest that hiw may regulate distinct molecular pathways in motor neurons and in nociceptors. In the larval NMJ, mutations of hiw or expression of hiw△RING cause a diminished evoked excitatory junction potential (EJP) amplitude due to decreased quantal content in synaptic vesicles [18,21,46]. However, this diminished evoked EJP amplitude phenotype is apparently opposite to the hypersensitive nociception phenotype observed in this study. Thus, the downstream targets and/or pathways of Hiw in nociceptors may be distinct from those in motor neurons.
We identified the BMP signaling pathway as an important signaling pathway in nociceptors that is regulated downstream of hiw. In fly motor neurons, it has been proposed that BMP signaling is a direct target of Hiw ligase [24]. However, a later study reported that pMad up-regulation was not detected in motor neuron nuclei in hiw mutants [31] and controversy has arisen over this interaction. We found that nuclear pMad signals were up-regulated in hiw mutant nociceptors, and that this molecular phenotype was rescued by wild-type hiw expression (Fig 2). In addition, we also detected striking accumulation of pMad in both the nuclei and cytoplasm of nociceptors expressing Hiw dominant negative proteins (Figs 2 and 4). Finally, using UAS-mad1, we showed that a Mad-dependent pathway is responsible for the hypersensitive thermal nociception caused by hiw△RING expression (Fig 2H). Our data therefore support the idea that the nociceptor BMP signaling pathway is regulated downstream from hiw.
Although we demonstrated that BMP signaling is downstream of hiw in nociceptors, we have yet to determine the precise mechanism for Hiw regulation of BMP signaling. Our genetic analysis suggests that BMP signaling in nociceptors is regulated independently from the wnd pathway (Fig 3). Wnd is the best characterized target substrate of Hiw in the regulation of NMJ morphology [31,38–41,47]. Our expression analysis using various hiw deletion series showed that the set of hiw deletion constructs that induced up-regulation of BMP signaling in nociceptors was not identical to the set that induced abnormal synaptic morphology in motoneurons [39]. This finding is somewhat consistent with the existence of a Wnd-independent mechanism in the regulation of BMP signaling in nociceptors, since the Hiw-Wnd pathway plays a pivotal role in regulating synaptic morphology in larval NMJ.
Intriguingly, our expression study of the hiw deletion series showed that the expression of HiwNT caused a prominent accumulation of nuclear pMad, while the expression of HiwCT or Hiw△RCC1 caused accumulation of pMad signals in both the nuclei and cytoplasm in nociceptors (Fig 4C–4E). These data raise the possibility that Hiw is involved in at least two different mechanisms which regulate pMad: one pathway affecting nuclear pMad and another for cytoplasmic pMad. Given that hiw is a large protein with many functional domains for interacting with multiple molecules, the notion that hiw is involved in multiple processes regulating various aspects of neuronal functions in both motor neurons and nociceptive sensory neurons is perhaps unsurprising. Further studies are necessary to reveal the mechanisms of Hiw-dependent regulation of BMP signaling in nociceptors.
We have presented a new physiological preparation for investigating the calcium levels in nociceptor terminals with a physiologically relevant noxious heat stimulus. This allowed us to demonstrate that up-regulation of BMP signaling in nociceptors sensitizes the physiological responses of nociceptors in response to noxious heat in addition to its effects on behavior (Figs 5 and 6). We also demonstrated that 24-hour activation of intracellular BMP signaling in nociceptors is sufficient for the nociceptive sensitization (Fig 7). Although it has been previously reported that BMP signaling in nociceptors is required for nociceptive sensitization after tissue-injury in Drosophila [23], the mechanisms of the regulation of nociception by BMP signaling was totally unknown. Our study provides the first evidence to implicate BMP signaling in regulating physiological processes in nociceptors that control its sensitivity to noxious stimuli.
The BMP signaling pathway plays crucial roles in various developmental processes, such as embryonic patterning, skeletal development, and the development of neuronal circuits [48,49]. The roles of BMP signaling in the regulation of neuronal activity has also been extensively investigated in larval motor neurons, where BMP signaling plays crucial roles in the homeostatic regulation of synaptic morphology and transmission [50,51]. In larval NMJ, the expression of active-form Tkv increases evoked EJP amplitude which is a similar effect on neuronal output to that we observed in nociceptors in this study [52]. A similar effect of active-form Tkv on evoked synaptic currents has been also reported in aCC interneurons in larval CNS [53]. These previous studies and this study together indicate that BMP signaling may function as a positive regulator of neuronal outputs. However, the previous studies and our current study also highlight differences in the functions of BMP signaling in different neurons. First, interfering with BMP signaling with dominant negative Mad did not cause nociception insensitive phenotypes (S3 Fig) (consistent with another study that found that nociceptor-specific knockdown of BMP signaling components did not affect basal thermal nociception [23]). In contrast, loss of BMP signaling components in motor neurons decreased evoked EJP amplitude [24,36,54]. Second, expression of activated-Tkv in nociceptors resulted in an expansion of axonal projections (Fig 5F–5I), the same manipulation does not increase the size of NMJ, while it increases nuclear pMad level also in motor neurons [24]. Although a full understanding of the mechanisms through which BMP signaling regulates nociceptor sensitivity requires further investigation, these results indicate that BMP signaling may act, at least in part, differently in the nociceptors and motor neurons to regulate neuronal outputs and morphology.
Hiw and BMP signaling pathway components are all evolutionally well-conserved. The role of hiw in the negative regulation of nociceptive signaling may be as well. A mammalian hiw orthologue Phr1/MYCBP2 has been previously implicated in a negative regulation of nociception processing. Specifically, it has been reported that Phr1/MYCBP2 is expressed in DRG neurons, and that intrathecal injection of an antisense oligonucleotide against Phr1/MYCBP2 causes hypersensitivity in formalin-induced nociceptive responses [55]. Furthermore, nociceptive and thermoceptive neuron-specific Phr1/MYCBP2 knock-out mice show prolonged formalin-triggered sensitization in thermal nociception, whereas no obvious phenotypes are observed for basal nociception in the knock-out animals [56]. Decreased internalization of the TRPV1 channel (which is mediated through a p38 MAPK pathway) has been implicated in this prolonged nociceptive sensitization in MYCBP2 knock-out mice [56]. In contrast, whether BMP signaling plays a role in regulating nociception in mammals is unknown. Similarly, the degree to which the role of Hiw and BMP signaling is conserved in the physiological regulation of mammalian nociceptors represents a fascinating topic for future investigation.
Intriguingly, Hiw and BMP signaling have been implicated in nerve regeneration/degeneration processes after axonal injury in both Drosophila and mammals [17,57]. In flies, axonal injury leads to decrease of Hiw, which leads to the upregulation of Wnd that promotes axonal degeneration in motor neurons [47]. Phr1/MYCBP2 is also involved in promoting axonal degeneration after sciatic or optic nerve axotomy [58]. Smad1 is known to be activated and play an important role for axonal regeneration after peripheral axotomy of DRG neurons [59–62]. Because nerve injuries are thought to be one of key contributors for neuropathic pain conditions and peripheral axotomies are widely used to generate neuropathic pain models in mammals, it will be of particular interest in the future to determine whether the Hiw-BMP signaling pathway and up-regulation of intracellular BMP signaling in nociceptors play a role in the development of a neuropathic pain state in mammals.
Canton-S and w1118 were used as control strains as indicated. The other strains used in this study were as follows: ppk1.9-GAL4 [63], UAS-mCD8::GFP [64], UAS-ChR2::YFP line C [4], hiwND8 [18], hiwΔN, hiwΔC, UAS-hiw, UAS-hiwΔRing [21], UAS-hiwNT, UAS-hiwCT, UAS-hiw△RCC, UAS-hiw△HindIII, UAS-hiwCT1000 [39], wnd1, wnd2, UAS-wnd [31], ppk1.9-GAL4; UAS>CD2 stop>mCD8::GFP hs-flp, UAS-tkvQD [34], tub-GAL80ts [65], ppk-CD4-tdGFP [66] and UAS-GCaMP6m [42]. UAS-mad1 [37]
The thermal nociception assay was performed as described previously [3,6,10,16,67]. NEL latency was measured as initial contact of the thermal probe on the lateral side of the larval body wall to the completion of NEL (a 360° roll). Stimulation was ceased at 11 seconds. A thermal probe heated to 46°C was used to examine the insensitive phenotype since it usually evokes NEL in less than 3 seconds [3,6,10,16,68]. A 39°C probe, which usually results in NEL in 9–10 seconds, was used to examine thermal hypersensitivity, as using a lower temperature probe is important to detecting the hypersensitive phenotype [16].
The optogenetic nociception assay was performed as described previously [5] with slight modifications. 3.8 klux was used to test for optogenetic hypersensitivity, but 76 klux blue light was used in the analysis of 24-hour tkvQD induction (Fig 7). Because male larvae showed a lower responsiveness to optogenetic nociceptor activation than females (S6 Fig), male larvae were used to allow for more easily detectable hypersensitivity.
Antibodies used in this study were as follows: rabbit anti-GFP (Invitrogen, 1:1000), mouse anti-GFP (Invitrogen, 1:250), mouse anti-rat CD2 (AbD Serotec, 1:200), rabbit anti-pMad (gift from Ed Laufer, 1:1000), goat anti-rabbit Alexa488 (Invitrogen, 1:1000), goat anti-rabbit Alexa568 (Invitrogen, 1:1000), goat anti-mouse Alexa488 (Invitrogen, 1:1000) and goat anti-mouse Alexa568 (Invitrogen, 1:1000). Larvae were filleted, fixed in 4% paraformaldehyde for 30 minutes and then stained according to a standard protocol [69].
Wandering third instar larvae expressing mCD8::GFP in nociceptors were filleted and immunostained as described above. To minimize variation due to processing controls, experimental specimens were processed side-by-side within the same staining solutions. In order to avoid skewing results from potential biases of pMad staining among different segments, one dorsal Class IV mutidendritic neurons (ddaC) each from segments A4, 5 and 6 (three neurons in total) was imaged in each sample (Zeiss LSM 710 with a 100x/1.4 Plan-Apochromat oil immersion or Olympus FV1200 with a 100x/1.4 UPLSAPO oil immersion). Z-stack images were converted to maximum intensity projections. To quantify nuclear pMad signals, nociceptor nuclei were identified based on the absence of GFP signal, and a region of interest (ROI) outlining the nucleus was delineated. The average signal intensity of nuclear pMad staining in the ROI was then calculated. Background signal intensity was determined as the mean from ROIs (identical size and shape of the nucleus from the image) drawn in the four corners of each image. The calculated background signal intensity was then subtracted from the nuclear pMad signal level. Data are plotted as nuclear pMad levels normalized to that of the co-processed control specimens. Image analyses were performed in Adobe Photoshop.
Wandering third instar larvae expressing mCD8::GFP in nociceptors under the control of ppk1.9-GAL4 were anesthetized by submersion in a drop of glycerol in a chamber that contained a cotton ball soaked by a few drops of ether. ddaC neurons in segments A4-6 were imaged on Zeiss LSM 5 Live with a 40x/1.3 Plan-Neofluar oil immersion objective lens. A series of tiled images were captured and assembled to reconstruct the entire dendritic field of the three A4-6 ddaC neurons. Z-stack images were then converted to maximum intensity projections. Dendritic field coverage was quantified as described previously [16].
A ppk1.9-GAL4; UAS>CD2 stop>mCD8::GFP hs-flp strain was used to induce single cell flip-out clones in order to sparsely label nociceptors. Six virgin females and three males were used to seed vials containing a cornmeal molasses medium for a period of 2 days. The seeded vials were then heat-shocked in a 35°C water bath for 30 minutes. After an additional 3 to 5 days, wandering third instar larvae were harvested from the vials and dissected. In order to precisely identify the neurons responsible for the axons labeled in the CNS, the incision made in filleting the larvae was along the dorsal side, and the CNS remained attached to the fillet prep during immunostaining. mCD8::GFP and rat CD2 were detected using rabbit anti-GFP and mouse anti-rat CD2 primary antibodies, and visualized by anti-rabbit Alexa488 and anti-mouse Alexa568 secondary antibodies, respectively. Axon terminal branches of single cell flip-out clones were imaged in the abdominal ganglion using a Zeiss LSM 5 Live with a 40x/1.3 Plan-Neofluar oil immersion objective. The cell body of origin for each flip-out clone was then determined by inspecting the body wall of the corresponding fillet. Flip-out clones belonging to A1-7 segments were imaged and analyzed.
To analyze the projection patterns for axon terminals, the presence or absence of terminal branches in each neuromere and longitudinal tract was manually identified for each single nociceptor clone. In order to align clones projecting to different segments, positions relative to the entry neuromere were used. The neurons that aligned were then used to calculate the percentage projecting to each neuromere and longitudinal tract. Heat-maps were color-coded according to these percentages using Microsoft Excel and Adobe Illustrator.
The quantification of axon terminal area was performed in Matlab. Z-stack images of axon termini were converted to maximum intensity projections and manually cropped to exclude signals from other clones in the same sample. The green channel (GFP) and red channel (CD2) of the cropped images were separately binarized using Otsu’s method [70]. The number of GFP-positive pixels were counted to calculate the area innervating the termini. To compensate for differences in the size and shape of the ventral nerve cord, the number of GFP-positive pixels was normalized to the average size of a single neuromere, which was calculated as the number of CD2-positive pixels divided by the number of neuromeres in the cropped image. To analyze axon terminals in nociceptors after 24-hour tkvQD expression (see below), GFP and CD2 signals were linearly enhanced to match to the control images in order to compensate low expression level of GFP and CD2. The clones whose signal intensities were too low to be binarized by Otsu’s method were excluded from the analysis.
Larvae raised in normal fly vials for 5 or 6 days at 25°C, or larvae raised on apple juice plates containing ATR for 4 days at 25°C, were transferred to 30°C for 24 hours. In every experiment, experimental genotypes and control animals were treated side-by-side to minimize the effect of potential variations in temperature.
The ppk1.9-GAL4 UAS-GCaMP6m strain was crossed to either a control strain (w1118) or UAS-tkvQD strain. Activity of larval nociceptors were monitored at their axon terminals in the larval ventral nerve cord (VNC), which was exposed for imaging by a partial dissection as follows: wandering third instar larvae expressing GCaMP6m in Class IV md neurons were immobilized in ice cold hemolymph-like saline 3.1 (HL3.1) (70 mM NaCl, 5mM KCl, 1.5 mM CaCl2, 4 mM MgCl2, 10 mM NaHCO3, 5 mM Trehalose, 115 mM Sucrose, and 5 mM HEPES, pH 7.2)[71]. The outer cuticle of each larvae was cut at segment A2/A3 to expose the central nervous system from which intact ventral nerves innervate the posterior larval body. The partially dissected animals were transferred to an imaging chamber containing HL3.1 equilibrated to the room temperature (23–25 °C). A strip of parafilm was placed over the larval VNC and was used to gently press the nerve cord down onto a coverslip for imaging. A Zeiss LSM5 Live confocal microscope and a 20x/0.8 Plan-Apochromat objective with a piezo focus drive were used to perform three-dimensional time-lapse imaging. Z-stacks consisting of 10–11 optical slices (Z depth of 63 to 70 μm) of 256 x 128 pixel images were acquired at approximately 4 Hz. During imaging, and using a custom-made thermal probe, a heat ramp stimulus was applied locally to one side of the A5 to A7 segments. The temperature of the thermal probe was regulated using a variac transformer. 10V was used to generate a 0.1 °C/sec heat ramp stimulation and no voltage was applied during cooling. A thermocouple probe (T-type) wire was placed inside of the thermal probe to monitor the probe temperature, and the data were acquired at 4 Hz through a digitizer USB-TC01 (National Instruments) and NI Signal Express software (National Instruments). The acquired images and temperature data were analyzed using Matlab software (Mathworks). Maximum intensity projections were generated from the time-series Z-stacks. Region of interest (ROI) was selected as a circular area with a diameter of 6 pixels, whose center was defined as the centroid of the A6 neuromere. Averaged fluorescent intensities (F) were calculated for the ROI for each time point. The average of Fs from the first 30 frames was used as a baseline (F0), and the percent change in fluorescent intensity from baseline (ΔF/F0) was calculated for each time point. Since acquisitions of images and probe temperatures were not synchronized, probe temperature for each time point was estimated by a linear interpolation from the raw probe temperature reading. For a comparison of controls and tkvQD OE, ΔF/F0, data were binned and averaged in 1°C intervals.
To statistically compare proportional data, Fisher’s exact test was used. Multiple comparisons of proportional data were corrected by the Bonferroni method. For non-proportional data, Mann-Whitney’s U-test was used for pair-wise comparisons, and Steel’s test (non-parametric equivalent of Dunnet’s test) was used for multiple comparisons. Statistical analyses were performed in R software and Kyplot.
|
10.1371/journal.pgen.1004798 | Quantitative Genetics of CTCF Binding Reveal Local Sequence Effects and Different Modes of X-Chromosome Association | Associating genetic variation with quantitative measures of gene regulation offers a way to bridge the gap between genotype and complex phenotypes. In order to identify quantitative trait loci (QTLs) that influence the binding of a transcription factor in humans, we measured binding of the multifunctional transcription and chromatin factor CTCF in 51 HapMap cell lines. We identified thousands of QTLs in which genotype differences were associated with differences in CTCF binding strength, hundreds of them confirmed by directly observable allele-specific binding bias. The majority of QTLs were either within 1 kb of the CTCF binding motif, or in linkage disequilibrium with a variant within 1 kb of the motif. On the X chromosome we observed three classes of binding sites: a minority class bound only to the active copy of the X chromosome, the majority class bound to both the active and inactive X, and a small set of female-specific CTCF sites associated with two non-coding RNA genes. In sum, our data reveal extensive genetic effects on CTCF binding, both direct and indirect, and identify a diversity of patterns of CTCF binding on the X chromosome.
| We have systematically measured the effect of normal genetic variation present in a human population on the binding of a specific chromatin protein (CTCF) to DNA by measuring its binding in 51 human cell lines. We observed a large number of changes in protein binding that we can confidently attribute to genetic effects. The corresponding genetic changes are often clustered around the binding motif for CTCF, but only a minority are actually within the motif. Unexpectedly, we also find that at most binding sites on the X chromosome, CTCF binding occurs equally on both the X chromosomes in females at the same level as on the single X chromosome in males. This finding suggests that in general, CTCF binding is not subject to global dosage compensation, the process which equalizes gene expression levels from the two female X chromosomes and the single male X.
| A major challenge in human genetics is to understand the mechanisms that link variation in genomic sequence to phenotypes of interest, including disease. Since 2005, a growing number of genome-wide association studies (GWAS) have associated both disease and normal phenotypes with over 9,800 single nucleotide polymorphisms (SNPs) [1]. Association studies can identify either causative variants or SNPs in linkage disequilibrium (LD) with the causative variant. Considerable effort has been invested in identifying potential causative variants, because this is essential to understanding the mechanistic route from the change in genomic sequence to final phenotype. The majority of the loci that have been found are not in strong linkage disequilibrium with a protein coding variant, suggesting that a change in a non-protein coding DNA sequence is often responsible for the phenotypic effect [2].
One route to finding intermediates between genotype and whole organism phenotype is to study the effect of genetic variants on gene regulation. New technologies such as microarrays and RNA sequencing (RNA-seq) have enabled quantification of transcript levels for every gene in a genome. Similarly, genome wide measurements of transcription factor occupancy and chromatin structure via chromatin immunoprecipitation followed by sequencing (ChIP-seq) [3] and DNase I hypersensitivity assays [4]–[7] have made it possible to quantify the state of upstream activities important for regulating transcription. Using DNase I hypersensitivity and binding assays for the CTCF transcription factor on two family trios with known genome sequences, we showed that allele-specific binding patterns consistent with strong genetic effects could be readily measured at heterozygous sites [8]. Other studies have shown allele specific binding of RNA polymerase and NF-κB binding measured across a small number of individuals [9], or of a wider range of transcription factors in a single cell line [10]. Similarly, differences between mouse strains in binding of PU-1 and CEBP/α at enhancer regions correlate with sequence differences and adjacent gene expression [11]. Intriguingly, some sites with prominent SNPs in the binding motifs of CTCF did not show a genetic effect in a study of its binding across an extended family [12]. Reciprocally, differences in transcriptor factor binding were seen between closely related species even where there was no sequence difference in the binding region [13].
In order to examine these phenomena further, and infer potential causative connections to disease GWAS results, we need to identify specific cases where a genetic variant affects a binding site. To do this we can use a genetic association study, as in GWAS, that searches for statistical association of genetic variants to quantitative measurements taken across samples. The variants with statistically significant association are known as quantitative trait loci (QTLs). When applied to transcript expression levels as the measurements on 60 or more samples, this approach has identified thousands of expression quantitative trait loci (eQTLs) [14]–[16]. A QTL study of human open chromatin [17] found 8,902 DNase I hypersensitivity sites that were correlated with genetic variants. However, there are currently no systematic association studies of how genetic variation in human populations affects the binding pattern of a specific transcription factor. Here we carry out such a study.
To identify transcription factor binding QTLs, we measured the binding of CTCF across a panel of cell lines. CTCF is a highly conserved multifunctional protein that serves as both a transcription factor as well an insulator binding protein, preventing interactions between enhancers and promoters and demarcating chromatin domains. Working with cohesin, CTCF can also mediate chromosomal looping interactions, and is involved in imprinting as well as X-inactivation (see [18], [19] for reviews). There have been extensive locus specific studies [20]–[26] and specific genome wide screens [27]–[30] demonstrating the different roles of CTCF in different circumstances. Studies by ourselves and others have shown the extent of genetic effects on CTCF binding in families [8], [12], although specific loci underlying these effects have not been identified.
We used ChIP-seq to measure CTCF binding in 51 lymphoblastoid cell lines (LCLs) from the HapMap CEU population, each of which had already been sequenced as part of the 1000 Genomes Project [2] and had been subjected to RNA-seq analysis [31]. Our data and analysis identified thousands of CTCF binding QTLs across the human genome. These data, together with the available full genome sequence of the cell lines, allowed us to explore parameters of genetic effects on protein-DNA binding. For example, we defined the relationship of the QTL location to the TF binding motif, estimated the relative impact of substitutions and insertions/deletions (indels), and measured whether allele-specific differences are indicative of population-wide variation.
Furthermore, our study revealed a previously uncharacterized mode of CTCF binding on the X chromosome. In human females (XX), one X chromosome is randomly inactivated and does not express most protein coding RNAs (reviewed in [32]). Thus for most X chromosome genes, both male and female cells have just one active locus, resulting in dosage compensation between the two sexes. The X-inactivation process requires expression of the non-coding RNA Xist from the inactive X. When we looked at CTCF binding on the X chromosome across our samples, we observed three distinct classes of CTCF binding sites. One major class was sensitive to X inactivation such that the active X showed stronger binding. Another class showed similar binding by CTCF on both X chromosomes, and the third, minor class of sites exhibited female specific binding.
We performed ChIP-seq on extracted chromatin from genotyped LCLs as previously described [33] except that we sequenced the DNA fragments from both ends (Figure 1) (Materials and Methods). We quantified binding to binding regions similarly to previous work [33] but pooled all the samples and identified a composite set of binding regions with detectable CTCF binding at low threshold. We then counted the sequence fragments that overlap each binding region in each individual, and normalised the signal to correct for systematic biases as in Degner et al [17]. We discarded binding regions that showed very little inter individual variance or had only one or two individuals with significant binding scores. Overall, our normalized data showed good consistency across all 51 individuals, as well as variation in signal sufficient to motivate QTL analysis (Figure 1B).
To measure the variance due to growth differences between the cells, we grew two individual cell lines as four independent cultures started on four consecutive days. There was higher correlation between these biological replicates from the same individual than between samples from different individuals, although all data sets were modestly correlated as expected for CTCF ChIP-seq (Figure S1). We next examined the data to see whether there were any systematic biases between samples. A principal component analysis identified some systematic variance, with a particularly strong first component (24.1%, Figure S2) that on investigation was correlated to known experimental batches. We therefore removed the first principal component, significantly improving the recovery of QTLs (Figure S3, Methods). We used the resulting normalised adjusted binding intensity (NABI) for subsequent analyses.
To discover QTLs, we looked for correlations between the NABI measures and SNPs and small biallelic insertion or deletion (indel) variants within 50 kb of the relevant binding region, using a linear model (Table 1, Example in Figure 2A; see Methods). As expected, the majority of variants do not have a significant association with variation in CTCF binding, with the linear model P-value distribution following the expected distribution (>95% of tests, fraction of the overlap between the black line and red line, Figure 2B). When samples are permuted, the distribution of the test statistic falls on the expected line (see Figure S4). Using a non-parametric statistic we saw similar P values (Figure S5). Using a Bonferroni adjusted threshold of P<3.8E-9 (See details on association testing in Methods) we find 509 binding regions with significant QTLs. Using a more liberal False Discovery Rate (FDR) [34] approach to take advantage of the smaller number of effectively independent tests occurring in these limited cis-regions, we discovered 1,837 binding regions (3% of total binding regions) with at least one significant variant at the 1% FDR level; relaxing the threshold to 10% FDR we discover 6,747 binding regions (12% of the total) (Table 1).
We chose to focus further analysis on the 1% FDR threshold as this provided ample QTLs from which to derive insights. We only considered one association per binding region, because the small number of samples meant that there was insufficient power for a conditional analysis for secondary associations in almost all cases. Within this set of associations, the genetic variant accounted for a substantial fraction of the variation in CTCF binding (median R square 0.38).
We summarised the collective set of variants which might be involved in each binding region association as being the cluster of SNPs within one order of magnitude of the P-value of the lead variant. 24,534 variants were identified in at least one cluster at the 1% FDR level, 13.4 variants on average per binding region (Table 1). As expected, these variants were mainly clustered around the target binding region, and when a CTCF binding motif could be identified (1341 of the 1837 cases) and a cluster QTL variant was present in the motif, the frequency was correlated with the information content of the motif (Figure 2C), as seen previously [12]. However, only a minority of significant binding regions had a QTL candidate within the motif (433/1341), and in only a small majority of cases there was a QTL within 1 kb (747/1341), of the binding region (Table 2).
We explored further the cases where there was no proximal variant in the cluster. There was not a substantial difference in genotype quality around the associated binding regions in these cases compared to binding regions with proximal effects, suggesting that there is not a large missing data problem. When considering all 1000 Genomes Project variants including those with allele frequency below 5%, in 95.5% of these cases, there was a proximal variant within 1 kb of the binding region in linkage disequilibrium (LD) with the distal lead variant, where LD was defined as the absolute value of D′>0.5. In approximately half of these cases the P-value of the proximal association either fell just outside the one order of magnitude threshold to fall in the cluster, or was just under the FDR threshold (Figure S6). In the 99 such cases where such a proximal variant was within the CTCF binding motif, the position of the variant was correlated with the information content of the position in the motif (Figure S7). Therefore a substantial fraction of the apparently distal cases appear to be explained by proximal cases. However still only a minority can be explained by variants in the binding motif.
We also conducted the analysis excluding short indels to replicate the more commonplace association analysis using only SNPs. In an indel-free analysis we would have missed QTLs in 67 binding regions entirely (∼5% of significant binding regions), and for 56 additional binding regions the closest observed explanatory SNP would have been over 1 kb away from the motif inside the peak. For these SNPs, there is usually a short indel with similar direct P-value inside the binding region. We further explored whether another cause for distal QTL effects could be due to the distal variant affecting a second neighbouring binding region, which in turn influenced the primary binding region, but there was only one case where we could find any evidence for this model (Figure S8). We additionally investigated the cases where there exist binding interactions between the QTL binding region and the neighboring region. We observed corresponding changes in histone modifications depending on the direction of the interactions between two binding regions (Figure S9, S10).
The effect size distribution with respect to allele frequency shows increased effect sizes for lower frequency SNPs, with a clear absence of large effects of common alleles (Figure S11). There is no statistical difference in effect size distribution between SNP and indel variants (Figure S11).
The dual-end sequencing of the ChIP-seq fragments provides the resolution to discover specific binding modes that influence the spatial distribution of the recovered fragments. To analyse this, we characterised ChIP-seq binding regions by metrics that summarised the extent of the peak and the position of the summit on a per individual basis, and used these additional metrics as phenotypes in a quantitative trait analysis using the methods described above. We found 25 shifts in peak shape driven by a genetic locus at the 1% FDR. Ten cases were also associated with a change in peak height. An example is shown in Figure 3, with the two homozygous genotypes showing the creation of a new associated peak, and merging of a double peak, and from visual inspection the other cases also look as if they can be explained as two CTCF peaks in close proximity, one or both of which is under cis-genetic control.
There are 61 CTCF QTL variants that overlap with disease and trait associated variants from other studies (GWAS Catalog [1]). In particular there is a disproportionate overlap with immune system related diseases (20 variants; Chi-sq P-value 1.7E-9). This is consistent with the lymphocyte origin of LCLs, and suggests a causal pathway for CTCF binding in the molecular aetiology of the disease phenotype in at least some cases. However many of these variants fall within the MHC locus, and a full causal analysis would need to take account of the complex LD structure there.
In summary, these results are consistent with previous studies [9], [10], [12], [13] that observed substantial variation in transcription factor binding within and between species, only a minority of which could be accounted for by genetic differences in the binding site. We also found that only 25.7% of our QTLs could be explained by a genetic variant in the motif. The majority of the remainder can be explained by changes within 1 kb of the motif, consistent with observations that transcription factor binding differences between mouse strains are more likely if there are genetic differences within 200 bp of the binding site [11]. However there remain some genetic associations for which we are not able to identify any proximal candidate, suggesting that longer range influences can make some contribution to CTCF binding.
This data set represents an excellent resource to directly examine allele-specific biases in TF binding at heterozygous sites in a larger set of individuals than previous studies [8]. Allele-specific binding refers to statistically significant biases in binding to the two alleles in a diploid cell, at sites where a heterozygous polymorphism allows the two alleles to be distinguished. Allele-specific binding thus is an independent way of assessing how genetic variants at binding sites might affect binding variation. Although the two alleles at heterozygous SNPs are normally referred to as the reference or alternate allele (referring to which base is found in the reference genome sequence and which is the alternate base), here we chose to categorize the two alleles as ancestral (shared with chimp) or derived (human specific). This has two advantages. First, any residual effect of biases in aligning sequence reads to the reference allele will be minimized. Second, measuring allele-specific binding in terms of the ancestral and derived allele provides information about how evolutionary changes might affect CTCF binding.
After processing the reads, we identified allele-specific statistically sites using a binomial null model of equal occupancy of both alleles at heterozygous sites, using a 5% FDR corrected threshold (see Methods). This process identified 589 SNPs that have replicated in at least two individuals showing significant allele-specific bias. We examined the allele counts of all heterozygous individuals at these 589 SNPs. For most sites (91.5%) the allele-specific biases were consistent between individuals, confirming the predominantly genetic basis of allele-specific binding (Figure 4A). At such sites, the same ancestral or derived allele was preferred for binding across 2 or more individuals.
However, there were 50 (8.5%) sites which showed significant but opposite allele-specific biases between two or more individuals. Six of these 50 sites could potentially be explained by virtue of being close to loci known to be subject to allelic exclusion (the Immunoglobulin heavy chain), a process that affects one allele randomly (see Discussion). One site lies in the KCNQ1 imprinted locus, where the regulatory status depends on parent of origin rather than genotype. The 46 other sites at which the allele-specific binding bias switches between individuals (Table S1) could represent new random allelic exclusion loci or imprinted sites, or could arise because the site at which we see allele specificity is incompletely linked with the causal variant [35]. We tested whether there was a SNP which specifically explained the allele specific switching site; for 28 cases this was the case. We are not able to directly test whether any of these sites could be due to imprinting because parent-of-origin information is not available for the heterozygous alleles of these individuals.
Interestingly, a significant majority (68%, P<1E-16) of the SNPs showed increased binding to the ancestral allele (Figure 4A). Alignment bias towards the reference allele has been reported before [8] and because the ancestral allele is more likely to be the reference allele, the increased binding to the ancestral allele could be the result of the alignment bias. To rule out this possibility, we analyzed the cases where the ancestral allele is the alternate allele and found that the binding bias remained towards the ancestral allele (Figure S12). Additionally, we repeated the allele-specific analysis after using a variant-aware aligner (see Methods). The results were largely identical to what we observed as described above, indicating that the preference for the ancestral allele is not a trivial outcome of any alignment bias (Figure S13).
The allele-specific signal at binding regions (intra-individual measurements) mostly correlated linearly with the QTL effect size (inter-individual measurements) (Figure 4B). There were however exceptions to this, and these were mainly cases in which there was an allele-specific signal but not inter-individual QTL. We did not observe QTLs with strong effect size in binding regions that did not show strong allele-specificity (Figure S14).
While exploring the correlation of between CTCF sites, we observed an unexpected behaviour of CTCF signal on the X chromosome. Strikingly, for 87% of CTCF sites on X (excluding the pseudoautosomal regions) there was a strong gender effect (P-value <0.01, Mann Whitney on gender); in nearly every case females have a significantly higher signal on average than males. The higher peak amplitude observed in females indicates, in effect, that the vast majority (87%) of CTCF sites on the X chromosome are occupied on both chromosomes. This is in contrast to the transcription of protein-coding mRNA (3% not compensated, i.e. X-inactivation escape genes), ncRNA (9%) [35] or other transcription factor occupancy as measured by DNase I (4%) (data from Degner et al [17]). We created a simple metric of the relative levels of activity, being the difference between the average male and average female signal, in each case adjusted for library depth as for the QTL analysis (Figure 5A and B). Protein coding mRNA and the majority of DNase I sites are consistent with only one active chromosome, leading to dosage compensated mRNAs(reviewed in [32]). As expected, there is a larger set of female specific ncRNAs, in particular the three XIST transcripts (Figure 5A). Using the Mann Whitney test of gender bias per site, we classified sites first as having significant bias, and then split the significant bias to cases consistent with balanced haploid behaviour, which we call “both-active” sites, and a small number of female-specific sites where there is a strong CTCF signal for females but almost no signal in males (Figure 5B). The remaining CTCF sites, which show similar levels between males and females we describe as single-active. The both-active sites and the single-active sites are evenly distributed along the chromosome (Figure 5A), and the XIST site and two clusters of female specific sites are obviously distinct from the rest.
We first confirmed that the single-active and both-active sites represent different modalities of CTCF binding, using intra-individual allele-specific analysis and independent DNase I data. LCLs are Epstein Barr Virus (EBV) transformed lines from a mixed B-cell population, and can be clonal or polyclonal, so that some female-derived LCLs show consistent or clonal X inactivation, whereas others have a mix of both X chromosomes being inactivated. Because we cannot assume that our LCL lines were clonal and therefore have consistent X inactivation, we first assessed the 17 female cell lines for clonal X inactivation status using heterozygous SNPs in genes known to be silenced on the inactive X [36], and selected the 13 lines with consistently skewed expression of these genes indicating consistent X inactivation (Methods). In these 13 lines, heterozygous SNPs in CTCF binding regions showed strikingly different behaviour between the single-active and both-active sites. The single-active sites showed strong allele-specific CTCF binding behaviour (similar to mRNA) whereas the both-active sites showed balanced signal over the two alleles from the very same samples (Figure 5C). In addition, we projected the DNase I data from the independent Yoruban cell lines onto the CTCF classification. For the 451 DNase I sites overlapping the CTCF sites on the X chromosomes, there was a strong concordance of this independent assay, performed on independent cell lines, with the classification of CTCF sites (Figure 5D). Both these analyses strongly support the finding that there are two major distinct types of CTCF binding sites on the X chromosome, with the both-active sites being bound on both the active and inactive X chromosome and the single-active sites being bound on only one chromosome (most likely the active X chromosome).
We then explored differences between these two classes of CTCF sites, using the ENCODE data from the GM12878 LCL [37], derived from a female individual. The majority of histone modifications associated with active chromatin (H3K4me4, H3K27ac) showed strong enrichments in the single-active class of CTCF sites but not in the both-active class, even when we excluded promoters (Figure 6A). The repressive histone mark, H3K27me3, implicated in X chromosome inactivation, is similar between both classes of sites. Interestingly both classes showed nucleosome phasing (Figure 6B) albeit stronger at the both-active sites. There is not a striking change in Cohesin co-binding, as shown by overlap with Rad21 and SMC3 (Figure S20). The mammalian conservation of the two classes of CTCF sites is high and approximately similar (62% for single-active sites overall with GERP conserved elements, and 53% for both active sites), showing that both classes have been under selection across mammalian evolution. Overall there is strong evidence for a dramatic distinction of these two classes of sites in terms of local chromatin behaviour. When we considered histone marks from a smaller set of cell lines, but with a broader set of marks we do not observe the same set of gender-biased signals except for H3K27me3, consistent with it's role in X inactivation. (Figure S21).
We then turned to the 23 female-specific sites. These sites were concentrated in two loci overlapping non-coding RNAs (X56 and X130), largely identical to sites previously identified as being involved in a repeat-specific X chromosome behaviour [38]. Although there are far fewer sites to analyse than the other classes, the female specific sites are all enriched for binding to YY1, which is known to tether XIST to the inactive X nucleation centre [39]. Horakova et al [38] explored the RNA expression of these ncRNAs in female cells; we performed fluorescence in situ hybridization (FISH) for RNA in both male and female cells. Consistent with the published results [38], we detected RNA from the active X at these loci in female cells (Figure 7A). In male cells we also detected RNA expression (despite the female specific nature of the CTCF sites, Figure 7B), suggesting that these CTCF sites are likely to be involved in a female-specific inactivation process at these loci. Using the data from Kilpinen et al, we can show that these sites are active in female lymphoblastoid cell lines, but not male (Figure S23). It is notable how few of these sites there are on the X chromosome, compared to the far more numerous single-active and both-active categories.
We then turned to the 23 female-specific sites. These sites were concentrated in two loci overlapping non-coding RNAs (X56 and X130), largely identical to sites previously identified as being involved in a repeat-specific X chromosome behaviour [38]. Although there are far fewer sites to analyse than the other classes, the female specific sites are all enriched for binding to YY1, which is known to tether XIST to the inactive X nucleation centre [39]. Horakova et al [38] explored the RNA expression of these ncRNAs in female cells; we performed fluorescence in situ hybridization (FISH) for RNA in both male and female cells. Consistent with the published results [38], we detected RNA from the active X at these loci in female cells (Figure 7A). In male cells we also detected RNA expression (despite the female specific nature of the CTCF sites, Figure 7B), suggesting that these CTCF sites are likely to be involved in a female-specific inactivation process at these loci. It is notable how few of these sites there are on the X chromosome, compared to the far more numerous single-active and both-active categories.
This study is the first systematic association based analysis of how normal genetic variation in humans affects the binding of a sequence-specific transcription factor, where the binding is measured as a quantitative trait. The properties of the binding quantitative trait loci (QTLs) that we identified are consistent with and extend previous smaller-scale studies of how genetic variation affects CTCF binding [8], [12], as well as similar analyses of chromatin QTLs underlying DNase I hypersensitive sites [17]. We find a large number of QTLs, with the majority being within or close to the binding region, and approximately a quarter inside the bound CTCF motif. By using 1000 Genomes Project cell lines, we can be reasonably confident that we have a full catalog of common variation of which some subset are the causal variants. Using this information we could show that for a large fraction of the associations where the initial analysis suggested a distal variant more than 1 kb away, there was a plausible causal candidate also within 1 kb of the binding motif. Overall this suggests that, at least for CTCF, the substantial majority (∼75%) of common genetic variants in the region with a reasonably strong effect on its binding lie within 1 kb of the binding motif, although only a minority are actually within the motif. This clarifies previous observations that genetic variants contributing to transcription factor binding were typically not in the motif itself [9], [13] but there was enrichment nearby [11].
We see hundreds of sites showing allele-specific binding. The idea that allele-specific events have similar effects inside one cell as genotypic effects do between individuals is commonplace [40]. Here we show that these two effects are well modeled by a linear relationship (at least for this assay), though with an interesting subset of allele-specific sites that show no QTL. In contrast there are few QTL loci that overlap binding regions without an allele-specific signal.
As expected, some of the allele specific sites switch specificity between the alleles in different samples, consistent with a nearby, incompletely linked causal allele, random allelic inactivation or parent-of-origin imprinting. Many of these sites can be explained by an incompletely linked nearby locus, highlighting that the causal variant is often not co-incident with the binding region.
Finally with more confident mapping of reads from paired read ChIP-seq data we are able to show that a consistent signal towards reference alleles is in fact predominantly due to a biological effect favouring ancestral alleles (at least for the CTCF transcription factor). This suggests that base pair changes segregating in the population tend to reduce binding of existing sites (rather than create new sites), at least for CTCF, and this is consistent with CTCF motif creation occurring by non-base pair changes, e.g. repeat deposition, as suggested in Schmidt et al [41].
We were initially surprised by the strikingly different behaviour of CTCF on the X chromosome compared to gene expression. Unlike transcribed genes, a large proportion of CTCF sites behave in a similar manner on both chromosomes. This is due to the same sites being bound on both the active and inactive X chromosome in females, as shown by the distribution of CTCF signal, the corresponding change in DNaseI signal in entirely separate cell lines and the lack of allele-specific signals in heterozygote sites in this class. This suggests that there is a subset of CTCF sites on the X chromosome that is bound on both copies despite the striking large scale compaction of the inactive X. This X chromosome-wide behaviour of CTCF is a very different phenomenon to the locus-specific interaction at the Xist/Tsix locus implicated in determining which X chromosome is inactivated [42], [43].
This observation has a number of implications. It is consistent with the multi-functional nature of CTCF, which has been commented on many times before in locus-specific [20]–[26] or specific genome-wide screens [27]–[30]. In this study we only examined behaviour in lymphoblastoid lines, and there might be cell type specific differences as well. Single-active sites show histone modifications and TF co-binding consistent with involvement with regulating expression on active chromatin. In contrast, the both-active sites show far less complex histone modification, consistent with structural functions that might apply to both chromosomes. Finally although we discovered this phenomenon on the X chromosome due to how these sites interact with X chromosome inactivation, it is consistent with the different binding behaviours of CTCF seen on the autosomes, with a diversity of different histone modification patterns at different CTCF sites [37].
The female-specific CTCF sites on the X chromosome are a very distinct subset; these are placed mainly over two non-coding RNAs expressed from the active X in females and males. The simplest explanation is that CTCF binding at these sites is involved in transcription repression on the inactive female X chromosome. This catalog of CTCF QTL sites is part of a growing set of molecular assays that are being examined in outbred individuals (for example, see [12], [17], [40], [44], [45]). It provides a specific hypothesis for the 63 disease related loci which overlap these QTLs, and for future overlaps with other molecular, cellular and disease related phenotypes. The gradual unraveling of the different variant effects on different molecular behaviour will provide a growing understanding of molecular and physiological processes in health and disease.
Cells were cross-linked with 1% formaldehyde for 7 min at room temperature. Formaldehyde was deactivated by adding glycine. Chromatin from harvested cells was sonicated with a Bioruptor to an average size of 500 bp DNA. Immunoprecipitation was performed using sonicated chromatin by adding anti-CTCF antibody (Millipore 07-729). ChIP DNA was used to generate a ChIP-seq library according to the standard Illumina protocol. The library was then sequenced using the Illumina HighSeq platform in 50 bp paired end reads. On average ∼85.54M reads were produced per sample. Sequence lanes were assessed for multiple quality metrics including total yield, read quality, mapping quality, GC content distribution and duplication rate. All sequencing reads were aligned to the human reference sequence (GRCh37) using BWA v0.5.9-r16 [46] using default parameter settings. Duplicate reads were marked by the “MarkDuplicates” function of the software Picard (v1.47 http://picard.sourceforge.net/) and removed. We applied a stringent filter by removing all the reads with MAQ quality score below 30, improperly paired (with 0x2 flag set in the BAM format), or with mate pairs more than 1 kb apart were removed. For allele specific analysis, we further performed local realignment using a variant-aware aligner glia (https://github.com/ekg/glia), which aligns reads against paths in a variant graph built by combining the reference sequence and known variants.
Read counts at each allele were counted for the 5.6M SNPs within 50 kb of a binding region. Heterozygous SNPs with significant allele-specific CTCF binding were identified. In detail, we calculated a binomial P value at all heterozygous SNPs with the null hypothesis that the two allele counts are equal. We then performed multiple testing adjustment at all heterozygous SNPs that have at least 2 reads at each allele and at least 2 reads difference between the two alleles using the Benjamini & Hochberg [52] method. Significant allele-specific binding was determined with an FDR 5%.
We analysed the gender specific CTCF binding on the X chromosome in the 27 female and 24 male LCLs. To ensure that our normalisation would not introduce any bias we used the raw CTCF binding intensities. For each of the 1,968 binding regions on the X chromosome, after blacklisted regions were removed, we assessed gender specificity by a Mann-Whitney U (MWU) test between the male and female samples. Binding regions were then classified as single-active and both-active based on the significance of the MWU test on the binding intensities. To classify the female specific binding regions we also incorporate the fold change between the average male and average female binding intensity. Similar analysis as for CTCF was performed on mRNA and ncRNA data from the Geuvadis project [35] and on DNase I [17].
To differentiate clonal and ployclonal sample, we analyzed allelic RNA expression on previously identified X inactivated genes [36] in 17 female samples. Samples where only one allele is expressed are determined to be clonal and polyclonal samples have RNA expressed from both alleles. Figure S22 shows examples of a clonal and a polyclonal sample.
For the single-active and both-active sites, we analysed the overlap of each category of sites with the ENCODE transcription factor and histone modification datasets for the female CEU lymphoblastoid cell line GM12878 [37]. To avoid bias introduced by unequal distribution of promoter sequences between the classes, we removed all binding regions that overlap with promoters identified in GM12878. For each binding region we define a partial overlap as an overlap. Signal aggregation of each of the classes of sites for histone modification and TF ChIP-seq data, and micrococcal nuclease cleavage was calculated using the ACT toolkit (http://act.gersteinlab.org/, [57]) with the parameters ‘-nbins = 50 -mbins = 0’. Only binding regions that are in the top 50% of bound sites were used. ENCODE bedGraph files for both TF and histone modifications were obtained from ftp.ebi.ac.uk: pub/databases/ensembl/encode/integration_data_jan2011/byDataType/signal/jan2011/bedgraph/and converted into signal files that are used as input for ACT.py.
Female and male human dermal fibroblasts cells (Invitrogen) were grown directly on Nunc Lab-Tek chamber slides, rinsed briefly using 1×PBS (PAA) and immediately fixed in 3% formaldehyde (Sigma-Aldrich) for 10 minutes, permeabilized using 0.5% Triton X-100 (BHD), 10 mM Ribonucleosidase Vanadyl complex (Biolabs) in 1×PBS (PAA) for 10 minutes, and then dehydrated through a 70%, 90% and 100% ethanol series, all at room temperature. The probes for X56 and X130 were selected according to their genomic locations reported in Horakova et al [38]. The probe for X56 consisted of the BAC clone RP11-416J22 and the fosmid G248P8472H8, the probe for X130 consisted of the BAC clone RP11-158M12, while the probe for XIST consisted of the fosmid G248P8779H11. All the clones were selected from the UCSC Genome Browser (GRCh37/hg19 assembly). Plasmid DNA was purified using the PhasePrep BAC DNA kit (Sigma-Aldrich) following manufacturer's protocol, amplified using the whole genome amplification kit (WGA2, Sigma-Aldrich) following manufacturer's recommendations. Clones were labeled using the whole genome re-amplification kit (WGA3, Sigma-Aldrich) as described before [58]. Briefly, X56 probe was labeled with Cyanine 3-dUTP (Enzo), the X130 probe was labeled with ChromaTide Texas Red-12-dUTP (Invitrogen) and the XIST probe was labeled with Green-dUTP (Abbott). For RNA-FISH, approximately 100 ng of labeled DNA from each probe and 2–4 µg of human Cot-1 DNA (Invitrogen) were ethanol precipitated, then resuspended in hybridisation buffer containing 50% formamide, 2×SSC, 10% dextran sulphate, 0.5 M phosphate buffer, pH 7.4. The probe mix was denatured at 65°C for 10 minutes before being applied onto cells on the chamber slides. Hybridisation was carried out in a 37°C incubator overnight. The post-hybridisation washes consisted of two rounds of 50% formamide/2×SSC washes followed by two additional washes in 2×SSC. All washes were done at 40°C, for 5 minutes. After detection, slides were mounted with SlowFade Gold mounting solution containing 4′,6-diamidino-2-phenylindole (Invitrogen). Images were visualised on a Zeiss AxioImager D1 fluorescent microscope. Digital image capture and processing were carried out using the SmartCapture software (Digital Scientific UK).
For the subsequent DNA-FISH, the same slides that have passed through the RNA-FISH assay described above were subject to the following treatment before denaturation in 70% formamide/2×SSC for 1.5 minutes, including one wash in 2×SSC for 5 minutes, digestion with RNase A (100 µg/ml RNase A in 2×SSC) for 30 minutes at 37°C, further digestion with 0.01% pepsin in 10 mM HCl for 5 minutes at room temperature, dehydration through an ethanol series as above and ageing on a 65°C hot plate for an hour. The X chromosome paint probe was labeled with biotin-16-dUTP (Roche). The making and denaturation of the X chromosome paint probe mix, hybridisation incubation, post-hybridisation washes and digital imaging were the same as above described, except that the biotin-labeled probes were visualised using Cy3 conjugated avidin (Sigma Aldrich).
The ChIP-seq data reported in this paper have been deposited in the European Nucleotide Archive, available with accession number ERP002168. The sample information and experimental design was deposited in ArrayExpress with accession number E-ERAD-141, linked to ERP002168.
|
10.1371/journal.ppat.1005095 | E3 Ubiquitin Ligase NEDD4 Promotes Influenza Virus Infection by Decreasing Levels of the Antiviral Protein IFITM3 | Interferon (IFN)-induced transmembrane protein 3 (IFITM3) is a cell-intrinsic factor that limits influenza virus infections. We previously showed that IFITM3 degradation is increased by its ubiquitination, though the ubiquitin ligase responsible for this modification remained elusive. Here, we demonstrate that the E3 ubiquitin ligase NEDD4 ubiquitinates IFITM3 in cells and in vitro. This IFITM3 ubiquitination is dependent upon the presence of a PPxY motif within IFITM3 and the WW domain-containing region of NEDD4. In NEDD4 knockout mouse embryonic fibroblasts, we observed defective IFITM3 ubiquitination and accumulation of high levels of basal IFITM3 as compared to wild type cells. Heightened IFITM3 levels significantly protected NEDD4 knockout cells from infection by influenza A and B viruses. Similarly, knockdown of NEDD4 in human lung cells resulted in an increase in steady state IFITM3 and a decrease in influenza virus infection, demonstrating a conservation of this NEDD4-dependent IFITM3 regulatory mechanism in mouse and human cells. Consistent with the known association of NEDD4 with lysosomes, we demonstrate for the first time that steady state turnover of IFITM3 occurs through the lysosomal degradation pathway. Overall, this work identifies the enzyme NEDD4 as a new therapeutic target for the prevention of influenza virus infections, and introduces a new paradigm for up-regulating cellular levels of IFITM3 independently of IFN or infection.
| IFITM3 is critical for limiting the severity of influenza virus infections in humans and mice. Optimal antiviral activity of IFITM3 is achieved when it is present at high levels within cells. Our results indicate that the E3 ubiquitin ligase NEDD4 decreases baseline IFITM3 levels by ubiquitinating IFITM3 and promoting its turnover. Depleting NEDD4 from cells results in IFITM3 accumulation and greater resistance to infection by influenza viruses. Therefore, we have identified NEDD4 as a regulator of IFITM3 levels and as a novel drug target for preventing influenza virus and other IFITM3-sensitive virus infections.
| Interferon (IFN)-induced transmembrane protein 3 (IFITM3) is a 15 kDa protein that restricts cellular infection by influenza virus [1,2,3]. IFITM3 is active against all strains of influenza virus that have been tested to date, regardless of serotype or species of origin [1,3,4,5,6], and it similarly inhibits many other medically important viruses such as HIV, SARS coronavirus, and Ebola virus [1,5,7,8,9]. Confirming its importance in vivo, IFITM3 knockout mice succumb to sublethal doses of influenza virus [10,11]. Likewise, IFITM3 is the only known protein for which a genetic polymorphism present in a significant percentage of the human population is associated with severe influenza virus infections [10,12,13,14]. In the cell, IFITM3 localizes to endosomes and lysosomes [15,16,17], and traps endocytosed virus particles within these degradative compartments by impeding the formation of the virus fusion pore [16,18,19]. Yet, even with this potent mechanism by which IFITM3 limits infections, influenza virus remains a significant health concern [20,21]. This may be explained by the fact that IFITM3 is present at low levels within most cells at steady state and is induced by IFNs only after infection has already been established [3,11,22]. The inability to up-regulate IFITM3 levels independently of infection or IFNs is a challenge preventing the field from harnessing the activity of IFITM3 for infection prevention.
We previously showed that ubiquitination increases the rate of IFITM3 turnover within the cell [15]. A non-ubiquitinated lysine-to-alanine mutant of IFITM3 possessed enhanced antiviral activity and a longer half-life as compared to WT IFITM3 [15]. These findings indicated that inhibition of IFITM3 ubiquitination could augment the activity and/or levels of endogenous IFITM3, thus offering a strategy for exploiting IFITM3 therapeutically or prophylactically against viral infections. The identification of the E3 ubiquitin ligase(s) capable of modifying IFITM3 among the more than 600 annotated E3 ligases in the human genome will be an important step toward validating this antiviral strategy.
Through our work studying tyrosine phosphorylation of IFITM3, we discovered that phosphorylation at tyrosine 20 (Y20) inhibited IFITM3 ubiquitination [23]. This led us to posit that phosphorylation of Y20 may block an E3 ubiquitin ligase recognition signal. Indeed, Y20 is part of a highly conserved PPxY motif (where P = proline, x = any amino acid, and Y = tyrosine, Fig 1A)[24]. PPxY motifs are commonly recognized by WW (characterized by two tryptophan residues spaced approximately 20 amino acids apart) domains of NEDD4-family E3 ubiquitin ligases, of which there are nine family members [25]. We chose to focus first on NEDD4, the prototypical member of this family, for several reasons: 1) NEDD4 and IFITM3 both have ubiquitous expression patterns while several other NEDD4-family members are tissue-specific (BioGPS.org [26]), 2) Like IFITM3, many of the known NEDD4 substrates are membrane proteins and are associated with endosomal and lysosomal pathways [25,27], 3) IFITM3 and NEDD4 are both S-palmitoylated, suggesting that they may localize to similar membrane subdomains [2,28], and 4) NEDD4 is reported to be inhibited by ISG15 [29,30,31], an IFN-inducible protein, thus providing an intriguing model whereby IFN might induce IFITM3 expression while also inhibiting its ubiquitination. Herein, we provide results demonstrating the ability of NEDD4 to ubiquitinate IFITM3 and identify a unique role for NEDD4 in decreasing steady state IFITM3 abundance, leading to increased cellular susceptibility to influenza virus infection.
To explore the possibility that NEDD4 ubiquitinates IFITM3, we first examined whether IFITM3 and NEDD4 are in proximity to one another within cells. We stimulated mouse embryonic fibroblasts (MEFs) with IFN-α to induce abundant expression of IFITM3. By performing immunofluorescence microscopy imaging of endogenous IFITM3 and NEDD4, we detected co-localization of these two proteins (Fig 1B). Further co-localization of these proteins with endogenous LAMP1, a lysosomal marker, indicates that NEDD4 and IFITM3 may interact at lysosomes (Fig 1B).
We next examined the effect of overexpressing HA-tagged human NEDD4 (HA-NEDD4) on IFITM3 ubiquitination. We observed a significant increase in myc-tagged human IFITM3 (myc-hIFITM3) ubiquitination when HA-NEDD4 was expressed as compared to the transfection control (Fig 1C). On the contrary, no increase in IFITM3 ubiquitination was seen upon overexpression of HA-tagged human CBL-B, another E3 ubiquitin ligase that has been reported to interact with NEDD4 [32], and that is associated with regulation of immune responses [33] (Fig 1C). Additionally, we examined the effect of overexpressing FLAG-tagged human NEDD4 (FLAG-NEDD4) on IFITM3 ubiquitination in comparison to FLAG-tagged human SMURF1 and SMURF2 (FLAG-SMURF1 and FLAG-SMURF2), both of which are members of the NEDD4-family of ubiquitin ligases. FLAG-NEDD4 caused an increase in IFITM3 ubiquitination while FLAG-SMURF1 and FLAG-SMURF2 were unable to robustly modify IFITM3 (Fig 1D). These results demonstrate that NEDD4 possesses a degree of specificity for IFITM3 that is lacking for CBL-B and the NEDD4-family members, SMURF1 and SMURF2.
As previously mentioned, NEDD4-family ubiquitin ligases possess two to four characteristic WW domains that interact with proline-rich motifs, including PPxY motifs, on substrate proteins [34]. NEDD4 has four WW domains and IFITM3 contains a highly conserved PPxY motif within its N-terminus (Fig 1A). To test whether these domains are required for IFITM3 ubiquitination, we generated an IFITM3 mutant in which each residue of the PPxY motif (17-PPNY-20 in IFITM3) was mutated to alanine (designated 17-20A) and utilized a FLAG-NEDD4 mutant in which its four WW domains were deleted (designated ΔWW). Upon co-overexpression of FLAG-NEDD4 with myc-hIFITM3, ubiquitination of IFITM3 was increased as expected, while the ΔWW mutant was unable to increase IFITM3 ubiquitination (Fig 1E). In fact, FLAG-NEDD4-ΔWW partially decreased steady state IFITM3 ubiquitination, perhaps indicating a dominant negative effect (Fig 1E). Moreover, the 17-20A mutant of IFITM3 showed less ubiquitination than WT IFITM3 and was unaffected by overexpression of NEDD4 (Fig 1E).
The IFITM3 PPxY motif shares its tyrosine with an overlapping YxxΦ motif known to be involved in the trafficking of IFITM3 from the plasma membrane to endosomes [23,35,36]. Thus, in order to be certain that the results we observed for the 17-20A mutant of IFITM3 was not because of interference with the YxxΦ motif, we tested additional PPxY mutants in which the two prolines were mutated to alanine (myc-hIFITM3-P17,18A) or in which the tyrosine was mutated to alanine (myc-hIFITM3-Y20A). Upon co-overexpression of FLAG-NEDD4, the ubiquitination of both of these mutants was only minimally increased as compared to the robust increase in ubiquitination of WT IFITM3 (Fig 2A). Interestingly, a truncated form of IFITM3 missing its first 21 amino acids, including the PPxY motif, is prevalent in certain human populations. This variant is associated with severe influenza virus infections [10,13,14] and more rapid progression of HIV-related disease [37]. A myc-hIFITM3 construct lacking these first 21 amino acids (Δ1–21) was, as expected, largely unaffected in terms of ubiquitination by overexpression of FLAG-NEDD4 (Fig 2B), identifying a potentially important difference between the truncated and full-length IFITM3 proteins.
Next, since NEDD4 has been shown to physically interact with the PPxY motifs of its substrate proteins [25], we examined whether or not NEDD4 and IFITM3 co-immunoprecipitate with one another. We found that myc-hIFITM3 and FLAG-NEDD4 indeed co-immunoprecipitated with one another (Fig 2C), suggesting a physical interaction. Importantly, this interaction was greatly diminished between FLAG-NEDD4 and the P17,18A mutant of IFITM3 (Fig 2C). In sum, these results indicate that the IFITM3 PPxY motif is required for a strong interaction with and ubiquitination by NEDD4.
To determine whether a non-enzymatic activity of NEDD4 might be mediating its effect on IFITM3 ubiquitination, we tested a catalytically inactive NEDD4 point mutant. We found that this mutant was unable to increase IFITM3 ubiquitination, establishing that catalytic activity of NEDD4 is indeed required for its ability to increase IFITM3 ubiquitination (Fig 3). Since murine (m)IFITM3 also possesses a PPxY motif (Fig 1A), we tested the ability of NEDD4 to affect mIFITM3 modification. Like myc-hIFITM3, we observed an increase in myc-mIFITM3 ubiquitination when HA-NEDD4 was co-overexpressed and observed no effect of the catalytic mutant (Fig 3), suggesting a possible evolutionary conservation of NEDD4 modification of IFITM3 in mice and humans. These data further implicate NEDD4 as an E3 ubiquitin ligase capable of enzymatically modifying mouse and human IFITM3.
While NEDD4 overexpression experiments suggest that NEDD4 directly ubiquitinates IFITM3 (Figs 1C, 1D and 1E, 2A and 2B and 3), this effect could be indirect. We therefore tested the ability of purified NEDD4 to ubiquitinate immunoprecipitated IFITM3 in vitro in order to confirm that NEDD4 can directly modify IFITM3. HA-hIFITM3 was incubated with purified NEDD4, enzymatic cofactors, and ubiquitin. We then re-immunoprecipitated IFITM3 and subjected it to anti-ubiquitin western blotting. Our results show that NEDD4 is capable of robustly ubiquitinating IFITM3 in vitro (Fig 4). Additionally, we employed ubiquitin mutants that could only be added via lysine 48 (K48) or lysine 63 (K63) linkages in order to examine whether NEDD4 preferentially utilizes one of these polyubiquitination linkages for modifying IFITM3. While both K48 and K63 linkages could be added to IFITM3 by NEDD4, we observed a preference for the K48 linkage in long polyubiquitin chains, which is traditionally associated with protein degradation (Fig 4). These results are consistent with our past results using linkage-specific anti-ubiquitin antibodies, which demonstrated that while both K48 and K63 ubiquitin linkages could be detected on IFITM3, K48 linkages are more prevalent [15]. These data are also consistent with our previous results indicating that ubiquitination of IFITM3 promotes its turnover [15].
In order to examine the effects of NEDD4 on endogenous IFITM3, we examined NEDD4 WT and knockout (KO) mouse embryonic fibroblasts (MEFs)[38]. We also utilized KO MEFs reconstituted with NEDD4 via retroviral transduction. Remarkably, Western blotting of lysates from NEDD4 KO cells showed an increase in steady state IFITM3 levels as compared to WT cells, while NEDD4 reconstitution decreased IFITM3 to WT levels (Fig 5A). To examine the requirement for NEDD4 in ubiquitinating IFITM3, we immunoprecipitated IFITM3 from large quantities of lysate from both WT and KO cells, expecting that the immunoprecipitation reagents would be saturated, thus providing us with comparable amounts of IFITM3 for examination of ubiquitination. Indeed, IFITM3 from NEDD4 KO cells was ubiquitinated much less than IFITM3 from WT cells (Fig 5B). These results demonstrate that NEDD4 is required for proper steady state ubiquitination of IFITM3, and that the absence of NEDD4 results in cellular accumulation of unmodified IFITM3.
Given the increase in baseline IFITM3 levels, we predicted that NEDD4 KO cells would be more resistant to influenza virus infection. We observed that NEDD4 KO MEFs were in fact significantly less susceptible to infections with influenza A virus (IAV) subtypes H1N1 and H3N2 (PR8 and X-31 strains, respectively) compared to WT control cells (Fig 5C). The decreased susceptibility of KO cells was returned to WT levels of infection upon NEDD4 reconstitution (Fig 5C). We also verified that the enhanced resistance of NEDD4 KO cells to influenza virus infection included resistance to recently circulating strains. NEDD4 KO cells were significantly less susceptible than WT cells to infection by both influenza B virus (IBV) and IAV H3N2 strains isolated in 2011 (Fig 5D). We also examined retrovirus pseudotyped with the vesicular stomatitis virus (VSV) G protein, which is also reported to be inhibited by IFITM3 [3,35,36,39,40]. As expected, the percent of NEDD4 KO cells infected with VSV G-pseudotyped virus was significantly less than WT cells (Fig 5D). Sendai virus (SeV), a parainfluenza virus that primarily fuses at the cell surface [41] and is thus only minimally affected by IFITM3 [4], was also tested. Unlike IAV, IBV, and VSV G-pseudotyped retrovirus, SeV was not appreciably affected by NEDD4 KO (Fig 5D). Thus, the pattern of virus restriction we observed is consistent with protection of NEDD4 KO cells by IFITM3.
To confirm that the increased resistance of NEDD4 KO cells to influenza virus infection was due to increased levels of basal IFITM3, we knocked down IFITM3 in NEDD4 WT and KO cells for 24 hours prior to infection. Knockdown was verified through Western blotting of cell lysates prepared at the time of infection (Fig 6A). Importantly, knockdown of IFITM3 in both NEDD4 WT and KO MEFs resulted in an increase in influenza virus susceptibility, and largely eliminated the resistance of NEDD4 KO cells to infection (Fig 6B). Overall, these experiments demonstrate that NEDD4 promotes cellular susceptibility to influenza virus infection by decreasing levels of IFITM3.
To extend our results to more relevant human lung cells, we utilized the A549 human alveolar epithelial cell line to study the role of NEDD4 in the regulation of steady state IFITM3 levels. Knockdown of NEDD4 with siRNA in A549 cells led to a significant increase in endogenous IFITM3 compared to non-targeting control siRNA (Fig 7A). As expected, NEDD4 knockdown led to a significantly greater resistance to IAV infection (Fig 7B and 7C). Importantly, we found that the relationship between NEDD4 knockdown and increased IFITM3 levels was preserved in two additional human lung cell lines (Fig 7D). Taken together with experiments presented in Figs 3, 5 and 6, these data confirm an evolutionary conservation between mice and humans in the regulation of cellular IFITM3 levels by NEDD4. This work also identifies NEDD4 as a novel target in human cells for improving resistance to influenza virus infection independently of IFNs or adaptive immunity.
The degradative pathway involved in the turnover of steady state IFITM3 has not been previously investigated. Since NEDD4 is known to associate with the endosomal and lysosomal system and to target several of its substrates for lysosomal degradation [25], our results identifying NEDD4 as the primary ubiquitin ligase for IFITM3 would suggest that IFITM3 is degraded in lysosomes. To test this hypothesis, we utilized chloroquine and bafilomycin, which inhibit endosomal and lysosomal acidification and thus the activation of pH-dependent lysosomal proteases. We observed that treatment of A549 lung cells with these two inhibitors caused an accumulation of IFITM3 (Fig 7E). Similarly, treatment with leupeptin, an inhibitor of specific lysosomal proteases resulted in a similar increase in IFITM3 levels (Fig 7E). This is in contrast to the treatment of cells with the proteasomal inhibitor MG132, which consistently caused a modest decrease in IFITM3 levels, perhaps due to up-regulation of lysosomal degradation pathways when proteasome activity is inhibited. Overall, these experiments demonstrate that, consistent with the co-localization of IFITM3 and NEDD4 at lysosomes (Fig 1B), and the ubiquitination of IFITM3 by NEDD4 (Figs 1C, 1D and 1E, 2,3,4 and 5B), IFITM3 is turned over by the lysosomal degradation pathway.
Our previous work established that ubiquitination promotes the turnover of IFITM3 [15]. Thus, identification of the IFITM3 ubiquitin ligase would provide a potential target for increasing IFITM3 abundance and resistance to virus infections. In our previous work studying regulation of IFITM3 endocytosis by phosphorylation, we made the serendipitous discovery that the amino acid Y20 within IFITM3 is involved in regulating IFITM3 ubiquitination [23], which led us to identify the involvement of the IFITM3 PPxY motif in its ubiquitination by NEDD4 (Figs 1E and 2). NEDD4 knockdown or knockout in human or mouse cells, respectively, resulted in substantially greater levels of steady-state IFITM3 (Figs 5A, 6A and 7A and 7D). This accumulation of unmodified IFITM3 is consistent with the observed decrease in IFITM3 ubiquitination in NEDD4 KO cells (Fig 5B).
An additional intriguing aspect of our finding that IFITM3 steady state levels are regulated by NEDD4 is the previously described role of the IFN effector ISG15 in inhibiting NEDD4 [29,30]. ISG15 is a ubiquitin-like protein that specifically binds to NEDD4, blocking its productive interaction with Ubiquitin-E2 ligase complexes [29,30]. The importance of this pathway was highlighted by two independent studies demonstrating that ISG15 blocks NEDD4-mediated monoubiquitination of the VP40 matrix protein of Ebola virus, thereby inhibiting the budding of Ebola virus-like particles [29,30]. Importantly, several studies have implicated ISG15 as a critical antiviral effector against IAV and IBV [42,43,44]. Two studies have demonstrated conjugation of ISG15 onto the IAV NS1 protein by the E3 ligase HERC5, and found that ISGylation of IAV NS1 antagonizes virus replication [44,45]. Interestingly, IBV NS1 specifically blocks human ISG15 conjugation by preventing ISG15 interaction with the ISG15 activating enzyme UbE1L, effectively counteracting its antiviral effect [43,46,47,48]. We posit that high levels of IFITM3 attained after IFN stimulation result from both IFITM3 gene induction, as well as increased IFITM3 protein stability as a result of ISG15 inhibition of NEDD4. We are currently investigating this exciting potential synergistic link between ISG15 and IFITM3.
Our work demonstrates that NEDD4 is required for proper basal ubiquitination of IFITM3 (Fig 5B). However, our results would also suggest that additional ubiquitin ligases are also able to modify IFITM3, particularly when IFITM3 is present at high levels. This is supported by detection of partial ubiquitination of our various IFITM3-PPxY mutants (Figs 1E and 2A and 2B) and by detection of modest IFITM3 ubiquitination in NEDD4 KO cells (Fig 5B). The identities of secondary ubiquitin ligases for IFITM3 are still unknown. Of particular interest are the ubiquitin ligases capable of modifying the truncated Δ1–21 splice variant of human IFITM3, which was not significantly ubiquitinated by NEDD4 (Fig 2B). Identifying the ubiquitin ligases that modify this disease-associated variant may implicate crucial differences in the stability and degradative pathways potentially underlying the defect possessed by this protein. Nonetheless, our data clearly implicate NEDD4 as the primary E3 ubiquitin ligase for IFITM3, and demonstrate that NEDD4 is essential for maintaining low steady state IFITM3 levels.
This current work is in contrast to a prior study that concluded the IFITM3 PPxY motif was not involved in regulating the levels or antiviral activity of overexpressed IFITM3 [36]. However, this previous work did not directly assess ubiquitination of IFITM3 upon mutation of the PPxY motif. Additionally, our experiences studying IFITM3 ubiquitination here and in our prior work suggest that when examining overexpressed IFITM3 constructs, ubiquitination has only subtle effects on total protein levels detected by Western blotting despite significant effects on the IFITM3 half-life13,20. Thus, overexpression likely masked any effects of mutating the PPxY motif on the parameters previously tested [36].
Our study has uncovered a novel mechanism by which NEDD4 indirectly promotes cellular entry of influenza virus by decreasing IFITM3 levels (Figs 5, 6 and 7). Although this work is the first of its kind to identify NEDD4 as a negative regulator of IFITM3 levels, NEDD4 is well described to be necessary for the replication of several important RNA viruses. For example, NEDD4 interacts with proline-rich motifs in the viral late budding domains of Ebola virus [49], rabies virus [50], and HIV [51]. Mono-ubiquitination of these domains promotes efficient budding and viral egress necessary for productive viral spread. However, it remains to be determined whether inhibition of NEDD4 will serve as an effective in vivo antiviral strategy, particularly since NEDD4 is a developmentally essential molecule as demonstrated by the embryonic lethality of NEDD4 KO mice [52]. Likewise, NEDD4 has been implicated in regulation of insulin-like growth factor signaling [53], T-cell-mediated immunity [54,55], and tumor suppression [56]. On the other hand, neuron- and skeletal muscle-specific NEDD4 KO mice are viable [57,58], and NEDD4 is naturally inhibited by ISG15 during virus infections [29,30], perhaps suggesting that short-term inhibition of NEDD4 can occur without adverse effects. Additional experimentation will be needed to answer these vital questions, and this will be aided by the development of selective NEDD4 inhibitors, which is an area of active investigation [59,60]. Overall, our study identifies inhibition of NEDD4 as a novel strategy for preventing infection by influenza virus and other IFITM3-sensitive viruses through the increased accumulation of the antiviral restriction factor IFITM3.
All cell lines used in these studies (HEK293T, A549, NCI-H358, NCI-H2009, and MEFs) were cultured in DMEM supplemented with 4.5 g/L D-glucose, L-glutamine, 110 mg/L sodium pyruvate, and 10% fetal bovine serum (Thermo Scientific) at 37°C and 5% CO2 in a humidified incubator. HEK293T and A549 cells were purchased from ATCC. NCI-H358 and NCI-H2009 cells were obtained from the ATCC and provided to us by Dr. Gustavo Leone (The Ohio State University). NEDD4 WT and KO MEFs used in this study were generated by Dr. Hiroshi Kawabe (Max Planck Institute)[38] and were kindly provided to us by Dr. Matthew Pratt (University of Southern California) who also generated the retrovirally reconstituted control cell lines. For Western blotting, cells were plated for 90% confluency in 6-well plates for 24 h prior to transfection with 2 μg/well of plasmids using Lipofectamine 2000 (Invitrogen). For microscopy, MEFs were plated for 50% confluency on glass coverslips in 12-well plates for 24 h prior to overnight treatment with IFN-α (BEI Resources). IFITM3 constructs were expressed from the pCMV-HA or pCMV-myc vectors (Clontech) as described previously [2,4,23]. IFITM3 mutants were made using the QuikChange Multi site-directed mutagenesis kit (Stratagene). Plasmids expressing HA-NEDD4 and HA-NEDD4-C867A were obtained from Addgene (plasmids 27002 and 26999, deposited by Dr. Joan Massagué, Memorial Sloan Kettering Cancer Center)[61], and plasmids expressing FLAG-NEDD4, FLAG-NEDD4-ΔWW, and HA-CBL-B were kindly provided by Dr. Jian Zhang (The Ohio State University). FLAG-SMURF1 and FLAG-SMURF2 were obtained from Addgene (plasmids 11752 and 11746, desposited by Jeff Wrana, University of Toronto)[62].
IFITM3 knockdown in MEFs was performed using Silencer Select Ifitm3 siRNA (Ambion, catalog no. 4390816) and negative control (Ambion, catalog no. 4390844). Human NEDD4 knockdown in A549, NCI H358, and NCI H2009 cells was performed using Dharmacon ON-TARGETplus SMARTpool Human NEDD4 (GE Healthcare, catalog no. L-007178-00) and Dharmacon ON-TARGETplus Control Pool Non-targeting control (GE Healthcare, catalog no. D-001810-10-20). siRNAs were transfected into cells using Lipofectamine RNAiMax transfection reagent (Invitrogen). Transfection of siRNA was performed for 24 h for mIFITM3 knockdown, and 48 h for NEDD4 knockdown. For Western blotting, cells were lysed with 1% Brij buffer (0.1 mM triethanolamine, 150 mM NaCl, 1% BrijO10 (Sigma), pH 7.4) containing EDTA-free protease inhibitor mixture (Roche) and 25 μM MG132 (Sigma). Immunoprecipitations were performed using EZview Red anti-c-myc or anti-HA affinity gel (Sigma), or with Protein G Plus Agarose Suspension (Calbiochem) in conjunction with anti-mIFITM3. Chloroquine, bafilomycin, and leupeptin were purchased from Sigma.
Co-immunoprecipitation assays were adapted from a previously described protocol [63]. HEK293T cells were co-transfected overnight with plasmids expressing myc-hIFITM3 and FLAG-NEDD4. Cells were washed twice with PBS, lysed on ice in Triton X-100 lysis buffer (50 mM Hepes, pH 7.5, 150 mN NaCl, 1% Triton X-100, 10% glycerol, 1.5 mM MgCl2, 1.0 mM EGTA, 10 μg/mL leupeptin, 10 μg/mL aprotinin, 10 μg/mL pepstatin, and 1 mM PMSF) for 5 min, and centrifuged at 1,000 x g for 5 min at 4°C. 50 μg of cell lysate was set aside for each sample in order to evaluate, via Western blotting, expression of myc-hIFITM3, FLAG-NEDD4, and GAPDH as a loading control. Equal concentrations of cell lysate were immunoprecipitated using 15 μL EZview Red anti-c-myc or anti-FLAG affinity gel (Sigma) per sample for 1 h at 4°C with gentle nutation. Immunoprecipitations were washed three times with lysis buffer and examined by Western blotting with both anti-myc and anti-FLAG for each immunoprecipitate.
Western blotting was performed with anti-myc (Developmental Studies Hybridoma Bank at the University of Iowa, deposited by Dr. J. Michael Bishop, catalog no. 9E 10), anti-HA (Clontech, catalog no. 631207), anti-hIFITM3 (Proteintech Group, catalog no. 11714-1-AP), anti-mIFITM3 (Abcam, catalog no. ab65183), anti-NEDD4 (Millipore, catalog no. 07–049), anti-FLAG (Sigma, catalog no. F7425), anti-actin (Abcam, catalog no. ab3280), or anti-GAPDH (Invitrogen, catalog no. 398600) antibodies. All primary antibodies were used at a 1:1000 dilution. Secondary antibodies, Goat Anti-Mouse IgG, HRP conjugate (Millipore catalog no. 12–349), Goat Anti-Rabbit IgG, HRP-linked (Cell Signaling, catalog no. 70745), and Goat Anti-Mouse, IgG1 Gamma 1 Heavy Chain Specific (SouthernBiotech, catalog no. 1070–05, specifically used for detecting immunoprecipitated protein ubiquitination) were all diluted at 1:20,000.
HEK293T cells were transfected overnight with plasmid expressing HA-hIFITM3. Protein collected from all wells of one 6-well plate was immunoprecipitated using anti-HA affinity gel and was washed extensively. Immunoprecipitated protein on affinity gel was resuspended in PBS. 10% of the retrieved protein was used in each reaction containing 500 μM ubiquitin or ubiquitin mutants (Boston Biochem, catalogue nos. U-100H, UM-K630, or UM-K480), 0.5 μM UbcH5b E2 ligase (Boston Biochem, catalogue no. E2-622), 100 nM UBE1 E1 ligase (Boston Biochem, catalogue no. E305), and 1x Ubiquitin Conjugation Reaction Buffer containing ATP (Boston Biochem, catalogue no SK-10) in the presence or absence of 100 ng recombinant human NEDD4 (Sigma, catalogue no. SRP0226). Reactions were allowed to proceed at 37°C for 1 h and were stopped by boiling for 5 min. The reactions were then diluted 1:100 in ice cold 1% Brij buffer, and IFITM3 was re-immunoprecipitated at 4°C using newly added anti-HA affinity gel prior to Western blot analysis.
Cells were fixed for 10 min with 3.7% paraformaldehyde, permeabilized with 0.1% Triton X-100 in PBS for 10 min, and blocked for 10 min with 2% FBS in PBS. Primary antibodies, anti-mIFITM3 (Fragilis, Abcam, catalogue no. ab15592) (1:500), anti-NEDD4 (1:500), and anti-LAMP1 (Santa Cruz Biotechnology, catalogue no. sc-19992), and Alexa Fluor-labeled anti-mouse and anti-rabbit secondary antibodies (Life Technologies, 1:1000) or anti-rat DyLight 550-labeled secondary antibody (Abcam, catalogue no. ab96888, 1:1000) were diluted in 0.1% Triton X-100 in PBS. Cells were treated with antibodies sequentially for 20 min at room temperature and washed five times with 0.1% Triton X-100 in PBS after each antibody treatment. Glass slides were mounted in ProLong Gold antifade reagent containing DAPI (Life Technologies). Images were captured using a Fluoview FV10i confocal microscope (Olympus).
Influenza viruses A/Puerto Rico/8/1934 (H1N1, PR8), a PR8 reassortant virus possessing the hemagglutinin and neuraminidase genes from A/Aichi/2/1968 (H3N2, X-31), A/Victoria/361/2011 (H3N2), and B/Texas/06/2011 were propagated in 10-day embryonated chicken eggs (purchased as day 0 eggs from Charles River Laboratories) for 48 h at 37°C as described previously [64]. PR8 and X-31 were provided to us by Drs. Bruno Moltedo and Thomas Moran (Mount Sinai School of Medicine) and the 2011 virus isolates were obtained from BEI Resources sponsored by the NIH/NIAID. SeV expressing green fluorescent protein (SeV-GFP)[65] was generated by Dr. Dominique Garcin (University de Geneve) and provided to us by Dr. Mark Peeples (Nationwide Children’s Hospital Research Institute). SeV-GFP was propagated in 10-day embryonated chicken eggs for 40 h at 37°C as described previously [66]. VSV G-pseudotyped retrovirus expressing green fluorescent protein was generated by transfection of the viral vector pLenti-CMV-GFP-puro (Addgene plasmid 17448, deposited by Dr. Eric Campeau)[67] and packaging plasmids (provided by Dr. Li Wu, The Ohio State University) along with plasmid expressing VSV G into HEK293T cells. Direct inhibition of GFP production by IFITM3 was not expected since this is driven by the CMV immediate early promoter and bypasses retrovirus-specific expression machinery [68]. Media was changed 18 h post-transfection, and media containing virus was then harvested 48 h post-transfection. Virus-containing media was centrifuged at 1200 x g for 5 min, filtered with 0.45 μm filters, frozen, stored at -80°C, and used for infection at a dose that provided approximately 70% infection of WT MEFs. MEFs were infected with IAV PR8, X-31, and H3N2 2011 strains, SeV and IBV at a multiplicity of infection of 5.0 or 10.0. MEFs were infected for 24 h, except in the case of VSV G-pseudotyped retrovirus infections, which were analyzed after 48 h of infection. A549 cells were infected with IAV strain PR8 at a multiplicity of infection of 2.5 for 6 h. Infected cells were washed with PBS and harvested in 0.25% trypsin EDTA. Cells were fixed in 3.7% paraformaldehyde for 10 min and permeabilized with 0.1% Triton X-100 for 10 min. IAV infected cells were stained with anti-influenza nucleoprotein (Abcam, catalog no. ab20343, 1:333) directly conjugated to Alexa Fluor 647 using a 100 μg antibody labeling kit (Life Technologies). IBV infected cells were stained with anti-IBV nucleoprotein (Thermo Scientific catalogue no. MA1-80712, 1:1000) followed by anti-mouse secondary antibodies conjugated directly to Alexa Fluor 488 (Life Technologies). Measurement of SeV and VSV G-pseudotyped retrovirus infection rates was done by detecting virus-encoded GFP. All antibodies were diluted in 0.1% Triton X-100 in PBS, and cells were stained for 20 min. Cells were washed three times with 0.1% Triton X-100 in PBS after each antibody treatment. PBS was used for final resuspension of cells for flow cytometric analysis using a FACSCanto II flow cytometer (BD Biosciences). Results were analyzed using FlowJo software.
|
10.1371/journal.ppat.1003610 | RIG-I and MDA-5 Detection of Viral RNA-dependent RNA Polymerase Activity Restricts Positive-Strand RNA Virus Replication | Type I interferons (IFN) are important for antiviral responses. Melanoma differentiation-associated gene 5 (MDA-5) and retinoic acid-induced gene I (RIG-I) proteins detect cytosolic double-stranded RNA (dsRNA) or 5′-triphosphate (5′-ppp) RNA and mediate IFN production. Cytosolic 5′-ppp RNA and dsRNA are generated during viral RNA replication and transcription by viral RNA replicases [RNA-dependent RNA polymerases (RdRp)]. Here, we show that the Semliki Forest virus (SFV) RNA replicase can induce IFN-β independently of viral RNA replication and transcription. The SFV replicase converts host cell RNA into 5′-ppp dsRNA and induces IFN-β through the RIG-I and MDA-5 pathways. Inactivation of the SFV replicase RdRp activity prevents IFN-β induction. These IFN-inducing modified host cell RNAs are abundantly produced during both wild-type SFV and its non-pathogenic mutant infection. Furthermore, in contrast to the wild-type SFV replicase a non-pathogenic mutant replicase triggers increased IFN-β production, which leads to a shutdown of virus replication. These results suggest that host cells can restrict RNA virus replication by detecting the products of unspecific viral replicase RdRp activity.
| Type I interferons (IFN) are critical for mounting effective antiviral responses by the host cells. For RNA viruses, it is believed that IFN is triggered exclusively by viral double-stranded RNA (dsRNA) or RNA containing a 5′-triphosphate (5′-ppp) that is produced during viral genome replication or transcription driven by viral replicases. Here, we provide strong evidence suggesting that the viral replicase also generates 5′-ppp dsRNA using cellular RNA templates, which trigger IFN. This finding indicates that viral replicase is capable of activating the host innate immune response, deviating from the paradigm that viral nucleic acid replication or transcription must be initiated in the host cell to trigger IFN production. Using Semliki Forest virus (SFV) as a model, we show that the magnitude of innate immune response activation by the viral replicase plays a decisive role in establishing viral infection. We demonstrate that in contrast to the wild-type SFV replicase, a non-pathogenic mutant replicase triggers increased IFN production, which leads to a shutdown of virus replication. Consequently, excessive IFN induction by the viral replicase can be dangerous for an RNA virus. Thus, we delineate a novel mechanism by which an RNA virus triggers the host cell immune response leading to RNA virus replication shutdown.
| The innate immune system is an ancient set of host defense mechanisms that utilize germline-encoded receptors for the recognition of pathogens [1]. This set of receptors, termed pathogen recognition receptors (PRRs), binds to the pathogen's own structural or pathogen-induced molecules and triggers an anti-pathogenic cellular state through various signal transduction pathways. The set of molecules brought into the cells or induced by pathogens are called pathogen-associated molecular patterns (PAMPs) [2]. The number of different germline-encoded PRRs is limited; therefore, PAMPs represent unique structural signatures that are characteristic of several groups of pathogens [1].
In the case of RNA viruses, double-stranded RNA (dsRNA) and 5′-triphosphate (5′-ppp) RNA are the most common pathogen-characteristic molecular structures recognized by PRRs. Viral RNA replicases generate 5′-ppp RNA and/or dsRNA in ample amounts during replication and transcription of viral RNA genomes. The presence of viral dsRNA in an animal cell is an indication of the pathogen invasion and is recognized by the innate immune system as a non-self entity, as vertebrate genomes do not encode RNA-dependent RNA polymerase (RdRp) activity. Recognition of viral dsRNA by specific PRRs leads to the induction of type I interferons (IFN; e.g. IFN-α and IFN-β) [3], which promote an antiviral state of the cell by inducing several hundred genes expression [4]. In vertebrates, type I IFNs and several other cytokines mediate innate immune system signals that determine the type of response elicited by the adaptive immune system [2].
Currently, three PRR families have been identified as innate immune sensors involved in the detection of virus-specific components in cells: Toll-like receptors (TLRs), retinoic acid-inducible gene I (RIG-I)-like receptors (RLRs), and nucleotide oligomerization domain (NOD)-like receptors (NLRs). Only TLRs and RLRs, however, are important for type I IFN induction. RLRs are the primary detectors of cytosolic 5′-ppp RNA and dsRNA generated by RNA viruses [3]. In addition to dsRNA [5], host PRRs detect dsRNA with 5′-ppp ends [6], single-stranded RNA (ssRNA) [7], and viral genomic DNA [8], [9]. Thus, type I IFN production is almost exclusively triggered by the recognition of viral nucleic acids. In fact, there seem to be only two exceptions. First, TLR4 receptors present on macrophages trigger type I IFN induction in response to lipopolysaccharide, which is not nucleic acid [10]. Second, TLR2 receptors present on “inflammatory” monocytes were recently reported to activate type I IFN in response to as yet unidentified components of DNA viruses [11]. For RNA viruses, however, it is believed that type I IFN is triggered exclusively by viral dsRNA [5] or 5′-ppp dsRNA [6], [12]. Accordingly, the presence of viral nucleic acids in a host cell is the absolute requirement for RNA virus detection and type I IFN production.
The main function for the viral RNA replicase is to drive the replication and transcription of viral RNA. Recently, however, several observations regarding an unusual extra property of positive-strand RNA virus replicases have been reported. In particular, the transient expression of the HCV replicase was shown to activate the IFN-β promoter in several human cell lines [13], [14], and IFN-β promoter activation was also observed for the human norovirus replicase [15]. Moreover, transgenic mice expressing the replicase of Theiler's murine encephalitis virus (TMEV) were resistant to infection by this virus and showed increased basal IFN levels [16]. Therefore, the expression of the viral replicase in the absence of a replication-competent viral genome can activate the IFN-β promoter. However, the mechanism and role of this innate immune response activation on the viral life cycle have not been determined.
In this report, we use Semliki Forest virus (SFV, Alphavirus) as a model to study the innate immune response of host cells to infection by a positive-strand RNA virus. Alphaviruses are single-stranded RNA (ssRNA) viruses that replicate in the cytoplasm. SFV RNA is translated into a replicase polyprotein, which consists of four multifunctional, non-structural viral proteins (nsP1, nsP2, nsP3, and nsP4). The replicase polyprotein is remodeled by the nsP2 protease through sequential cleavages to produce replicases with different specificities [17]. The core RdRp of the SFV replicase is represented by nsP4, which contains a conserved catalytic GDD triad [18]. However, when nsP4 is expressed separately from the other replicase proteins, it cannot function as an RdRp [19]. It has been reported that the alphavirus nsP2 protein is absolutely required for the suppression of the host cell antiviral response either by inducing macromolecular synthesis shutoff or by targeting the catalytic subunit of the cellular RNA polymerase [20], [21], [22]. In addition, it was demonstrated that nsP2 expression efficiently inhibited IFN-induced JAK-STAT signalling indicating a shutoff-independent mechanism [23]. However, the infection of heterogeneous bone-marrow-derived dendritic cells (BMDC) by wild-type Sindbis virus (SIN) resulted in high IFN-β induction, no prominent shutoff, and self-limiting infection [24]. In addition, IFN-α/β is very potently induced within the first 12 hr post-infection in the serum of mice infected by wild type SIN [25]. Thus, a simple model in which nsP2 antagonizes IFN-α/β and promotes virus infection cannot fully explain the virus phenotypes in non-established cell lines and animals [26]. On the other hand, SFV replication is associated with the production of viral dsRNA replication intermediates that are thought to be the molecules used by the host to detect SFV infection, even though these intermediates are located inside of the membrane-bound replicase complexes. Therefore, similar to other viruses, SFV possesses mechanism(s) that result in the prevention and/or suppression of the antiviral IFN response. A non-pathogenic SFV4 mutant that contains a mutation that disrupts the nuclear localization sequence (NLS) of nsP2 (SFV4-RDR) has been reported to be deficient in suppressing the antiviral IFN response [27], [28], [29], [30].
Here, we present evidence for a novel mechanism by which mouse embryonic fibroblasts (MEFs) detect SFV infection. This detection results in IFN-β induction and, in the case of MEF infection with a non-pathogenic SFV4-RDR, the shutdown of virus replication. The reconstitution of SFV replication by uncoupling the wild-type and mutant replicase expression from the viral RNA template led to the identification of the viral replicase as the enzyme responsible for IFN-β induction. Remarkably, wild-type and mutant SFV replicases were capable of IFN-β induction in the absence of viral RNA replication or transcription. IFN-β induction resulted from the production of IFN-inducing RNA ligands generated from cellular RNAs that could activate RIG-I and MDA-5. Increased levels of IFN-β were induced upon the expression of the viral replicase with a mutation in the NLS of nsP2, while inactivation of RdRp activity of nsP4 blocked the induction of IFN-β. These results indicate that during SFV infection, two concurrent processes are driven by the viral replicase. First, the viral replicase drives the replication and transcription of the viral genome. Second, the viral replicase generates PAMPs using host cell RNA as a template. Thus, the innate immune recognition of RNA virus infection is more complex than pattern recognition, which is based on the detection of invariant structures present in pathogens or viral replication intermediates. Furthermore, viruses must find ways to circumvent the increased efficiency of their recognition by utilizing powerful mechanisms targeting innate immune response. A failure to interfere with host cell detection mechanisms results in virus replication shutdown and elimination.
All previous studies on IFN induction by alphaviruses have been performed using either viral infection or transfection of viral replicons. It has been shown that wild-type SFV4 is capable of IFN-β induction upon infection of a host cell [27], [31], [32]. It is not clear whether this is valid for all alphaviruses as no type I IFN induction was observed in MEFs infected with wild type SIN [33]. In contrast it had been shown that transfection of cells with a SIN replicon, a modified alphavirus genome that replicates but does not produce virus due to the replacement of the viral structural genes with a heterologous sequences, also triggers IFN-β production [34]. Introduction of an RR649R→RD649R mutation into the NLS of SFV4 nsP2 resulted in a virus, termed SFV4-RDR, that induced almost seven times more IFN-β than the wild-type virus [27]. Hence, the replicase region is essential for triggering IFN-β induction and also for limiting that response. However, the exact mechanism(s) by which IFN-β is triggered during alphavirus replication and the reasons why the mutant virus induces more IFN production than the wild-type virus are poorly understood.
We have previously demonstrated that SFV replicase that was expressed from a non-replicating mRNA specifically and efficiently replicated the SFV RNA template when the template was provided as a separate RNA molecule [35], and this finding was later confirmed [36], [37]. Therefore, the approach used in the current study was based on the uncoupling of replicase expression from the viral RNA template. For this purpose, codon-optimized DNA sequences encoding only the replicases of SFV4 and SFV4-RDR were inserted under the control of a heterologous promoter to produce the pRep and pRep-RDR plasmids (Figure 1A). As a control, we generated a pRep-RDR/GAA plasmid containing two point mutations in the RdRp-specific catalytic motif [18] of the nsP4 protein (GD466D467→GA466A467) that inactivate the RdRp activity of the viral replicase. For accurate reconstitution of viral replication, we also produced the pSFVminRluc plasmid encoding a minimal viral RNA template, which contained the coding sequence of the Renilla luciferase (Rluc) reporter flanked by cis-sequences that are required for efficient viral replication (Figure 1A). To determine the basal level of Rluc activity expressed from the RNA template, we transfected mouse fibroblast COP-5 cells [38] with pSFVminRluc and compared the results to those obtained with co-transfection of COP-5 cells with pSFVminRluc and either pRep or pRep-RDR. As expected, the presence of functional replicase resulted in the accumulation of intracellular Rluc activity that exceeded the basal level obtained with the template alone. In contrast, no accumulation of Rluc activity above basal level was observed when COP-5 cells were co-transfected with pSFVminRluc and pRep-RDR/GAA (Figure 1B). Consequently, the intact RdRp activity of the viral replicase was responsible for the amplification of Rluc activity. This clearly indicated that replication of the SFVminRluc RNA template was driven by the viral replicases expressed from the pRep and pRep-RDR plasmids. Although the replication of SFVminRluc by the wild-type SFV replicase was more efficient (Figure 1B), we found that pSFVminRluc co-transfection with pRep-RDR induced an increase in IFN-β production that was an order of magnitude greater that that induced by co-transfection with pRep (Figure 1C).
There are at least three potential mechanisms by which pSFVminRluc co-transfection with pRep-RDR could lead to enhanced IFN-β induction. First, viral RNA replication products generated by Rep-RDR may be more abundant and/or more accessible to the host cell innate immune sensors than the viral RNA replication products generated by wild-type Rep. Second, the wild-type replicase of SFV may interfere with IFN-β production and/or signaling by multiple mechanisms [20], [21], [27], [28]. Alternatively, Rep-RDR may have additional, non-viral RNA template targets in the host cell transcriptome. To determine the exact mechanism, we transfected mouse COP-5 fibroblast cells with pRep, pRep-RDR, and pRep-GAA and measured the amount of IFN-β in the cell culture medium at different time points. Unexpectedly, the expression of the SFV replicase induced robust IFN-β secretion by the transfected cells. Rep-RDR induced approximately 3–4 times more IFN-β production than Rep, whereas Rep-GAA did not induce IFN-β production (Figure 1D). The expression of the SFV replicase subunits became readily detectable in a western blot assay only when pRep, pRep-RDR, and pRep-GAA were used at the highest dose, arguing against the possibility of over-expression (Figure S1). These results indicate that the SFV replicase is capable of inducing IFN-β in the absence of replication-competent viral RNA. Thus, the induction of IFN-β during SFV infection may not only be caused by the presence of viral RNA replication or transcription intermediates but may also result from additional intrinsic properties of the functional viral replicase.
We next wanted to compare the kinetics of IFN induction in COP-5 cells transfected with either pRep-RDR or poly(I:C), which is known to induce IFN-β [5]. To achieve the comparable IFN induction we needed to increase the SFV replicase expression level, which was accomplished by the exchange of the promoter in pRep-RDR and pRep-RDR/GAA (Figure S2). Poly(I:C) transfection resulted in steady, and at the highest dose, declining levels of IFN-β, whereas the SFV replicase expression triggered a delayed but very potent accumulation of IFN-β (Figure 1E). To determine whether the expression of the SFV replicase in primary MEFs also leads to IFN-β induction, we transfected MEFs with pRep-RDR, pRep-RDR/GAA DNA, or poly(I:C) dsRNA. In the MEFs, SFV replicase expression induced IFN-β secretion, whereas Rep-RDR/GAA again failed to induce IFN-β production (Figure S3). Chloroquine treatment, which was used to avoid the TLR-9-dependent induction of IFN-β in response to DNA [39], had no effect on IFN-β induction. Therefore, transfection of pRep-RDR into primary MEFs also triggers IFN-β production. Taken together, these results demonstrate that SFV replicase expression triggers the accumulation of previously unknown PAMPs, which induce an innate immune response.
Several reports demonstrated that the transient expression of a single-subunit HCV replicase (NS5B, nonstructural protein 5B) activated the IFN-β promoter [13], [14], [40], [41]. Hence, we wanted to compare the abilities of SFV and HCV replicases to trigger the IFN-β induction. Corresponding HCV replicase plasmids encoding active NS5B (pNS5B) and the NS5B with inactivated catalytic motif (pNS5B-GND, GD318D→GN318D) were generated. Transfection of pRep-RDR and pNS5B into COP-5 cells resulted in a similar and potent IFN-β production, whereas pRep-RDR/GAA and pNS5B-GND failed to induce interferon (Figure 1F). Thus, the ability to induce IFN-β expression is a property shared by replicases of positive-strand RNA viruses belonging to at least two different virus families (Togaviridae and Flaviviridae).
RNA viruses can be recognized by either MDA-5, RIG-I, or a combination of the two [3]. Therefore, we tested whether these RLRs would be involved in the detection of PAMPs produced by SFV replicase RdRp activity. We transfected COP-5 cells with siRNAs targeting the mouse Ddx58, Ifih1, and Dhx58 mRNAs, which encode the RIG-I, MDA-5, and LGP2 proteins, respectively. As a control, we used dsRNA poly(I:C) preparations shorter than ∼1.5 kbp in length, as dsRNAs of this size have been shown to induce IFN-β primarily through RIG-I [5]. As expected, silencing the expression of RIG-I either alone or in combination with MDA-5 strongly inhibited IFN-β induction by poly(I:C), which indicated that RIG-I is the primary poly(I:C) sensor in COP-5 cells (Figure 2A). Silencing both RIG-I and LGP2 expression enhanced IFN-β production as compared to silencing RIG-I alone, whereas knockdown of LGP2 expression alone or in combination with MDA-5 resulted in increased IFN-β production. These observations are consistent with the role of LGP2 as a feedback inhibitor of antiviral signaling [42] and indicate that there is both competition and interplay between receptors for a common substrate [43]. We then transfected pRep-RDR into COP-5 cells treated with siRNAs and measured IFN-β production. Again, silencing the expression of RIG-I alone strongly inhibited IFN-β induction, whereas knockdown of both RIG-I and MDA-5 blocked IFN-β production completely (Figure 2B). In this case, knockdown of RIG-I in combination with LGP2 did not enhance the IFN-β production as compared to the effect of RIG-I knockdown alone (compare Figures 2A and 2B). Notably, knockdown of LGP2 alone also did not increase IFN-β induction, whereas silencing of both MDA-5 and LGP2 did increase IFN-β induction. Importantly, silencing the expression of RLR sensors by different siRNAs gave essentially the same results, indicating that observed effects were not due to off-target effects (data not shown). The possibility that the siRNA transfections also affected the expression of the SFV replicase and thus contributed to the alteration of IFN-β production was also excluded, as the amounts of SFV nsP2 and nsP4 in different cell lysates were roughly identical (Figure 2B). Furthermore, the amounts of the replicase proteins in the transfected cells never exceeded the amounts in the cells that were infected with the corresponding virus. Thus, this analysis also excluded any possibility that the IFN-β induction was an artifact that resulted from the transient over-expression of the viral replicase. Therefore, the production of novel PAMPs is a natural function of SFV replicase RdRp activity, and RIG-I is the major sensor for these products. Moreover, for PAMPs generated by the SFV replicase, these data suggest that MDA-5 serves as an auxiliary sensor that also contributes to IFN-β induction, albeit to a much lesser extent than RIG-I.
It has been previously reported that the sensors for PAMPs generated during alphavirus infection include MDA-5 and RIG-I [31], [32]. Therefore, we determined which RLR sensors would be important for the detection of the mutant SFV4-RDR. For this purpose, SFV4-Rluc and SFV4-Rluc-RDR reporter viruses containing the Renilla luciferase (Rluc) coding sequence were used; in these viruses, the Rluc reporter is inserted between the coding sequences of nsP3 and nsP4, and this insertion does not interfere with correct replicase polyprotein cleavage. In addition, the expression of the reporter gene for these viruses is proportional to the RNA genome copy number [44]. Following transfection of the siRNAs, we infected MEFs with SFV4-Rluc-RDR at different multiplicity of infection (MOI) values. Analysis of the Rluc reporter activity indicated that transfection of the siRNAs had little or no effect on mutant SFV replication in MEFs (Figure 2D). In contrast, knockdown of both RIG-I and MDA-5 decreased IFN-β secretion at every MOI tested (Figure 2C). At the lowest MOI tested, it became clear that both RIG-I and MDA-5 were equally important for SFV recognition in MEFs (Figure 2C). These results indicate that the same RLR sensors are involved in both the detection of infection by the SFV and novel products of its replicase RdRp activity. However, MDA-5 knockdown had relatively little effect on the recognition of PAMPs generated by the intrinsic RdRp activity of the viral replicase, whereas the contribution of MDA-5 to the recognition of infection by the virus was similar to that of RIG-I.
Given that SFV replicase RdRp activity was absolutely necessary for the generation of PAMP structures to trigger the RIG-I pathway and IFN-β induction, we expected that these PAMPs may in fact be RNAs. Furthermore, the dominant role of RIG-I in the recognition of these PAMPs suggests that these potential RNAs must be shorter than the full-length dsRNA replication intermediates produced during viral infection. To address this question regarding the nature of these PAMPs, we isolated total RNA from COP-5 cells transfected with pRep-RDR, pRep-RDR/GAA and poly(I:C) and then separated the large RNAs fraction (>200 nt) from the fraction containing smaller RNAs (<200 nt). Only the large RNAs from COP-5 cells transfected with pRep-RDR were capable of inducing IFN-β in MEFs, whereas large RNA fraction from either pRep-RDR/GAA- or poly(I:C)-transfected cells was incapable of inducing a comparable IFN response (Figure 3A). Consequently, the RNA extracted from poly(I:C)-transfected cells, which was used for the secondary MEFs transfection, contained very little of the originally-transfected poly(I:C). Therefore, this RNA was unable to trigger the RIG-I pathway following re-transfection. Furthermore, this also indicates that the contribution of the IFN-β, translated directly from corresponding mRNAs purified from initially transfected cells, to the IFN-β production in re-transfected cells is negligible. Even when transfected at a 10-fold molar excess as compared to the large RNA fractions, the small RNA fractions were unable to trigger comparable IFN-β activity. This observation also indicates that the antiviral endoribonuclease L (RNase L), which has been reported to cleave host cell RNA to produce small RNA molecules (less than 200 nt) capable of triggering the RIG-I pathway [45], is not involved in the detection of products of SFV replicase RdRp activity.
Next, we transfected COP-5 cells with pRep-RDR and fractionated the total RNA into polyadenylated (polyA+) and nonpolyadenylated (polyA−) RNA using oligo(dT)-affinity chromatography. The polyadenylated RNAs comprised approximately 3% of the total RNA extracted from the COP-5 cells. We then transfected naïve COP-5 cells with polyA+, polyA−, or total RNA in stoichiometric amounts (polyA+ RNA : polyA− RNA : total RNA = 1 : 32 : 33) and measured the IFN-β response (Figure 3B). Approximately 90% of the IFN-β signal was induced by the nonpolyadenylated RNA, suggesting that this fraction contained the majority of the RNAs (PAMPs) generated by the SFV replicase RdRp activity.
It has been shown that RIG-I recognizes dsRNA or dsRNA containing a 5′-ppp [5], [6]. To determine the structural features of the IFN-β-inducing RNA generated by the SFV replicase, we treated RNA isolated from COP-5 cells with various ribonucleases (RNases). At the concentrations used, RNase A did not digest model dsRNA, whereas ssRNA was efficiently degraded (Figure 3C, upper and lower panels). RNase T1 is exclusively ssRNA-specific and does not degrade dsRNA, whereas RNase III should specifically digest only dsRNA unless used at a high concentration [46], which we confirmed to be the case (Figure 3C, compare upper and lower panels).
When RNA extracted from the COP-5 cells transfected with pRep-RDR was treated with RNase III, this RNA lost the ability to induce IFN-β production (Figure 3D). In contrast, RNase A had virtually no effect on the IFN-inducing activity of the RNA. Similarly, RNase T1 treatment of the RNA did not substantially alter its ability to induce IFN-β production (Figure 3D). These results suggest that the IFN-inducing RNA extracted from SFV replicase-transfected cells is in the form of dsRNA. As expected, DNase I treatment did not have any effect on the IFN-inducing of activity of the RNA. Finally, when we treated the RNA with alkaline phosphatase, the ability of the RNA to induce IFN-β was destroyed, which indicated that the terminal phosphate structure was absolutely required for IFN-β induction. Taken together, these results suggest that the SFV replicase generates ligands for RIG-I that consist of non-polyadenylated RNA species larger than 200 nt containing dsRNA regions and a terminal 5′-phosphate, which is most likely a 5′-triphosphate.
It has been shown that although the dsRNA containing SFV replicase complexes (spherules) are initially formed at the plasma-membrane, they are subsequently internalized [47] and localize to the cytoplasmic surface of both endosomes and lysosomes [48]. Moreover, expression of the viral replicase alone in type I IFN-deficient BHK (baby hamster kidney) cells resulted in the characteristic endo- and lysosomal localization of nsPs, although there was no spherule formation [37]. To determine whether the IFN-inducing RNAs produced by the SFV replicase associate with either endosomes or lysosomes, we performed subcellular fractionation of COP-5 cells transfected with pRep-RDR and pRep-RDR/GAA. The subcellular fractions containing various organelles were prepared from post-nuclear supernatants by flotation in sucrose step gradients. Subsequently, we extracted RNA from each of the fractions, and transfected these RNAs into naïve COP-5 cells. The RNAs extracted from each subcellular fraction of the pRep-RDR-transfected cells were capable of inducing IFN-β production; however, RNAs extracted from the endosomes and lysosomes were the most potent IFN-β inducers (Figure 4A).
Next, we determined whether the replicase-generated IFN-inducing dsRNA could be detected in transfected cells by immunofluorescence microscopy. For this purpose, we utilized the J2 monoclonal antibody, which recognizes dsRNA regions longer than ∼40 bp in length [49]. In this experiment, dsRNA-specific staining was detected in both pRep- and pRep-RDR-transfected COP-5 cells, whereas pRep-RDR/GAA-transfected cells contained no detectable dsRNA. Moreover, co-localization of dsRNA and the nsP1 of SFV was clearly detected (Figure 4B). Due to the enriched IFN-inducing RNA in endosomes and lysosomes, we further analyzed the potential co-localization between SFV nsP1 and dsRNA with the lysosomal marker protein LAMP2. The assay was performed using human rhabdomyosarcoma (RD) cells, as several mouse-specific LAMP2 antibodies produced a high background signal in COP-5 cells. Both dsRNA and LAMP2 showed co-staining with nsP1, and dsRNA and LAMP2 also showed co-staining with each other (Figure 4C). These results confirmed that the SFV replicase generates dsRNA structures and demonstrated that the dsRNA duplex region(s) is longer than ∼40 bp in length. Additionally, the endosomes and lysosomes are enriched in these IFN-inducing RNAs and consequently serve as the sites of SFV replicase docking and dsRNA generation. These findings indicate that our experimental system accurately reconstituted the conditions observed during actual SFV infection [37], [48].
Infection with a virulent strain of SFV, SFV4 [50], is cytotoxic for vertebrate cells and leads to the shutdown of cellular transcription and translation. This process has, at least in part, been attributed to the properties of nsP2 [22], [51]. Approximately half of the nsP2 produced is transported to the nuclei of SFV4-infected cells [29], and disruption of the nsP2 NLS by the RR649R→RD649R mutation attenuates the pathogenicity of the corresponding virus [30]. Previous studies with SFV4 and SFV4-RDR in cells deficient for the type I IFN response showed that both viruses grew to high titers with similar kinetics [27], [30]. However, in cells with an intact type I IFN response, the kinetics of viral accumulation differed; although both viruses grew to high titers by 12 hr post-infection (h.p.i.), the SFV4-RDR titer failed to increase further, whereas the SFV4 titer continued to increase until 24 h.p.i. [27]. These findings indicate that replication of the mutant SFV4-RDR was altered in cells with intact type I IFN signaling.
The analysis of the SFV4 and SFV4-RDR replication kinetics in MEFs did not reveal any differences in viral RNA accumulation up to 7 h.p.i. [27]. To analyze the replication of these viruses in greater detail, MEFs were infected with SFV4-Rluc and SFV4-Rluc-RDR at an MOI of 1, and the Rluc activity was measured. No significant difference in the replication kinetics of either reporter virus was observed (Figure 5A). Under closer examination, however, by 24 h.p.i., the replication of the mutant virus was almost entirely suppressed and only minor cytotoxic side effects were observed, whereas the decreased replication rate of the wild-type virus was associated with death. SFV4-Rluc-induced death of MEFs was also confirmed by the disappearance of abundant cellular protein that is recognized non-specifically by the nsP4 antibody (Figure 5C). Thus, the same phenotype (decrease of replication) seems to stem from a completely different origin. Similar to previously published results [27], we observed that MEFs produced considerably more IFN-β in response to infection with the mutant virus (Figure 5B). The observed difference (up to 40-fold), however, was nearly a magnitude larger than previously reported. These results suggest that the increased IFN-β production induced by SFV4-Rluc-RDR triggered an antiviral mechanism, which led to the restriction of viral replication in MEFs.
During several independent experiments, four aspects of the results were highly reproducible. First, the peak level of viral replicase expression, as indicated by the induction of Rluc reporter activity and the production of replicase subunits, was achieved faster in response to SFV4-Rluc-RDR (12 h.p.i.) than to SFV4-Rluc (15 h.p.i.); however, the levels of viral expression were roughly equal (Figures 5A,C,D). Thus, there are no major defects in replicase production and RNA replication of SFV4-Rluc-RDR. Second, for the mutant virus, IFN-β was detectable at earlier time points in comparison to the wild-type virus (Figure 5B). Third, peak levels of secreted IFN-β were obtained after peak levels of the viral replicase nsP3 and nsP4 (RdRp) subunits were established (Figures 5B,C,D). Fourth, after achieving the peak levels nsP3 and nsP4 of SFV4-Rluc remained at a relatively steady level, whereas the levels of nsP3 and nsP4 of SFV4-Rluc-RDR gradually decreased (compare Figure 5A and 5B). Taken together, these results indicate that the viral replicase that induces IFN-β production is formed faster during infection with the mutant virus. Furthermore, there is a clear correlation between replicase accumulation and IFN-β production. However, after the peak levels of IFN-β are achieved, the replicase of the mutant virus is degraded.
To test whether the increased IFN-β production that was induced by SFV4-Rluc-RDR triggered an antiviral mechanism leading to the restriction of viral replication, MEFs were infected with SFV4-Rluc-RDR at different temperatures. When MEFs were infected with SFV4-Rluc-RDR at 28°C, the IFN-β secretion was greatly delayed and clearly reduced as compared to that produced as a result of infection performed at 37°C (Figures 5H and 5F). This difference is most likely because at 28°C, the RDR mutation does not completely prevent the nuclear localization of nsP2 [51]. This delay in the kinetics of IFN-β secretion allowed for the efficient replication and spread of the mutant virus (Figure 5G). In contrast, when MEFs were infected with the mutant virus at 37°C, virus infection was restricted to the cells that were initially infected, and importantly, there was a complete suppression of viral replication at 24 and 36 h.p.i. for every MOI tested (Figure 5E). Thus, MEFs restrict the replication of SFV4-Rluc-RDR in an IFN-β-dependent manner.
To confirm the results obtained from the Rluc reporter activity analysis, we used northern hybridization approach to directly visualize the replication products. During SFV infection viral replicase first generates negative-strand RNAs (42S RNA −), which, together with the genomic RNA, forms dsRNA intermediate (Figure 5I). The replicase then utilizes the negative-strand RNA genomes to produce both positive-strand RNA genomes (42S RNA +) and subgenomic RNAs (26S RNA +). At 28°C, all RNA species produced by the SFV4-Rluc-RDR in MEFs accumulated in a time-dependent manner, indicating that there was efficient viral replication and spread of infection (Figure 5J). In contrast, at 37°C, the viral RNAs of positive polarity were detectable at 12 h.p.i. but were subsequently unable to be detected, whereas the amount of the 42S RNA negative strand remained at a steady level (Figure 5J). These results clearly demonstrate that the IFN-induced restriction of SFV4-Rluc-RDR replication is mediated by the destruction of the viral positive strands of RNA.
To compare the amounts of PAMP structures generated during the infection with SFV4-Rluc and SFV4-Rluc-RDR, we infected MEF cells with these viruses at a MOI of 1. The replication of each virus was analyzed at 12 h.p.i. by measuring Rluc activity. As observed in Figure 6A, both viruses replicated to the same extent, which is consistent with data from the previous experiment (Figure 5A). Also consistent with the previous data (Figure 5B), the production of IFN-β differed drastically (Figure 6B), suggesting that the wild-type virus produces less IFN-β-inducing PAMPs. To test this hypothesis, we extracted total RNA from infected MEF cells and used it to transfect COP-5 cells. To prevent the generation of additional dsRNA PAMPs after the transfection of COP-5 cells, we treated RNA samples that were extracted from the MEFs with UV (2000 µJ/cm2, 2 minutes). UV-treatment inactivated the infectious replication-competent SFV RNA, and as a consequence Rluc activity was fully abrogated in the RNA samples extracted from the infected cells (Figure 6C). Remarkably, the UV-treated total RNA extracted from the MEF cells infected with both viruses induced almost identical amounts of IFN-β in the COP-5 cells at all doses tested (Figure 6D). Thus, the wild-type and mutant SFV viruses generated equal amounts of PAMP structures during infection. Consequently, the wild-type virus, but not the mutant virus, efficiently blocks either the access of the RLR machinery to the PAMPs or disrupts host cell antiviral signaling (see the Discussion).
Next, we wanted to estimate the relative contributions of different RNAs generated by the viral replicase during the infection to IFN-β induction. To address this question, we needed to separate the viral RNA and its replicative dsRNA form from the dsRNAs of different origin. Here, we utilized the information resulting from the pioneering works on alphavirus RNA replication, which have shown that the only species of SFV-specific RNA of negative polarity detected in infected cells were the 42S (-) RNA strands [52], [53], [54], [55]. In addition, negative strands of replicative forms of alphaviruses either contain polyU sequences that are shorter than the corresponding polyA sequence at the 3′-terminus of the 42S genomic positive strand [53] or do not contain polyU sequences at all [56], [57]. In both cases the structure of SFV replicative dsRNA is ideal for purification by oligo(dT)-affinity chromatography similarly to both genomic (42S) and subgenomic (26S) single-stranded RNAs of SFV. Thus, polyA− RNA fraction, depleted from all types of SFV RNAs but containing the large majority of novel alphavirus replicase generated PAMPs (Figure 3B), could be obtained.
To address this question experimentally, we fractionated total RNA extracted from mock-, SFV4-Rluc-, and SFV4-Rluc-RDR-infected MEF cells (Figure 6) using oligo(dT)-affinity chromatography. The polyA+ RNA comprised approximately 5.6%, 10.8%, and 8.1% of the total RNA extracted from the mock-, SFV4-Rluc-, and SFV4-Rluc-RDR-infected MEF cells, respectively. Consequently, SFV increases the ratio of polyA+ RNA to polyA− RNA by synthesizing its own polyadenylated positive-strand RNAs. Next, equal amounts of the RNAs (polyA+ RNA : polyA− RNA = 1∶1, mass ratio) were resolved by performing electrophoresis on a native agarose gel and visualized by ethidium bromide staining. Only three major viral RNA species were present in the fraction containing the polyA+ RNAs that were isolated from the infected MEFs: 26S (+)strand, 42S (+)strand, and 42S (±) dsRNA (Figure 7A, lanes 3 and 5). The double-stranded nature of the 42S (±) RNA was confirmed by staining with acridine orange, as well as by its absence on a denaturing agarose gel (data not shown). These viral RNA species were also present in the corresponding polyA− RNA fractions (Figure 7A, lanes 4 and 6), but in highly reduced amounts. We found that polyA+ RNA fractions contained approximately 15-fold more viral dsRNA than the respective polyA− RNA fractions (Figure 7A; lanes 3 and 4, 5 and 6). The ratio of infectious units (42S(+) RNA genomes) between the two RNA fractions obtained from SFV-Rluc infected cells was also established using an infectious center assay. The infectivity was found to be 4.3×105 pfu/µg (plaque forming units per microgram) and 3×104 pfu/µg for polyA+ and polyA− RNAs, respectively, thus confirming the approximately 15-fold difference. In addition, the high infectivity of the isolated RNA reflected the high quality of the RNA and lack of RNA degradation during purification and fractionation.
The (+) strands of SFV RNAs have 5′ cap structure and, on their own, cannot serve as PAMPs. In contrast, the (−) strands of SFV genome RNA or the (−) strands of defective interfering RNA (DI-RNA) can be efficient PAMPs since they possess 5′-ppp structure. Therefore, we wanted to analyze which types of viral RNA (−) strands were present in both fractions of RNA. For this purpose, equal amounts of the RNAs (polyA+ RNA : polyA− RNA = 1∶1, mass ratio) were resolved by performing electrophoresis on a denaturing formaldehyde agarose gel and transferred to a nylon membrane. Next, we shredded the labeled full-length positive strand RNA of SFV4-Rluc into pieces with length of 300–600 nucleotides and used the obtained mixture as probe for northern hybridization analysis (Figure 7B). Only one major RNA species, corresponding to 42S(−) viral RNA, were detected in these experiments. The absence of additional major RNA species on the northern blot indicated that there were no incomplete or fragmented 42S (−) strands in cells infected with either SFV4-Rluc or SFV4-Rluc-RDR (Figure 7B). However, a single discrete negative-stand RNA present in trace amounts was detected in the polyA+ fractions of infected cells; its smaller size suggested that it could represent the negative strand of DI-RNA. Importantly, northern hybridization analysis revealed that the viral RNAs of negative polarity were present in abundance in polyA+ fractions and were barely detectable in polyA− fractions. For the RNA isolated from SFV4-Rluc infected cells, we found that polyA+ RNA fraction contained ∼15-fold more viral 42S(−) RNA than polyA− RNA fraction; the difference for RNAs obtained from SFV4-Rluc-RDR infected cells was even more prominent (Figure 7B). Moreover, for SFV4-Rluc infected cells, the nearly identical 15-fold excess of viral 42S (±) dsRNA and 42S (−) RNA species in the polyA+ RNA fraction strongly suggested that the latter species were derived from the former and no free 42S (−) strands existed in infected cells. Collectively, performed experiments demonstrated that the viral ssRNAs and dsRNAs fractionated with a similar efficiency and were present in 15-fold excess in the corresponding polyA+ RNA fractions.
Subsequently, we wanted to compare the potential of polyA+ and polyA− RNA fractions isolated from infected cells to induce the IFN-β. To block the possibility of additional PAMPs generation from infectious RNA, the polyA+ and polyA− RNA fractions were UV-treated as described above before being used in subsequent assays. The transfection experiments revealed that the virus-induced PAMPs were potent inducers of IFN-β expression, and at a dose of 1000 ng, both RNA fractions saturated the ability of the host cell to produce IFN-β. The same is true for 100 ng of polyA+ RNA from SFV4-Rluc-RDR infected cells. In contrast, an almost linear response of IFN-β production was observed for 100 ng and 10 ng of RNAs from SFV4-Rluc infected cells (Figure 7C). Therefore, the abilities of these RNAs to induce IFN-β expression were compared using these two RNA quantities. Based on the results of the quantification, the polyA+ RNA contains 15-fold more viral RNA than equal amounts of the corresponding polyA− RNA. Hence, if the viral dsRNA and/or negative strand RNA are the only PAMPs that are generated in the course of infection, then the transfection with 10 ng of polyA+ RNA should induce approximately 15-fold more IFN-β compared with the transfection with 10 ng of the corresponding polyA− RNA. However, as is evident from the graph in Figure 7C, this was not the case. Instead, the transfection of COP-5 cells with polyA+ and polyA− RNA fractions from SFV4-Rluc infected cells resulted only in approximately a 2-fold higher IFN-β induction in response to the polyA+ RNA species compared with the polyA− RNA. For the RNAs that were obtained from MEF cells infected with SFV4-Rluc-RDR, transfection with 10 ng of polyA+ RNA induced approximately 6.5-fold more IFN-β compared with the polyA− RNA fraction, which again is less than the values deduced from the content of viral dsRNA.
In addition to the major viral 42S (−) RNA species, a minor barely detectable and approximately six times faster migrating RNA species were observed in the polyA+ RNA fractions isolated from cells infected with SFV4-Rluc and SFV4-Rluc-RDR (Figure 7B). It has been reported that during SFV infection DI-RNA species might be generated [58], [59]. DI-RNA are viral deletion mutants that contain the RNA sequences important for replication but are unable to self-replicate and rely on virus replication machinery, similarly to our SFVminRluc RNA template [58]. The low abundance of this RNA and especially its enrichment in polyA+ fraction excludes possibility that this molecule may have significant role in IFN-β induction by polyA− RNA fraction. Nevertheless, the presence of DI-RNAs was subsequently analyzed using a highly sensitive strand-specific reverse transcription reaction on denatured RNA followed by PCR. This method allowed for the detection of negative strands of the DI-RNAs in all of the RNA samples from infected MEF cells; the length of the DI-RNA was between 1.5 and 2 kb (Figure S4). However, coherent with the results of northern blot analysis (Figure 7B) negative strands of DI-RNA were more abundantly present in polyA+ fraction. Furthermore, positive strands of DI-RNA were detected only in the polyA+ RNA fractions (Figure S4). Taken together these results confirm that DI-RNAs were efficiently removed from the polyA− RNA samples by oligo(dT) affinity chromatography and therefore could not contribute to IFN-β induction in response to this particular RNA fraction.
Next, we wanted to address whether the RNA PAMPs present in the polyA− RNA fraction from the infected cells were similar to the PAMPs generated by transfection of pRep-RDR. Most importantly, it was essential to determine if this fraction contained RNA molecules that were generated by the degradation of viral dsRNA by either RNase L or some other mechanism and would therefore lack a polyA sequence. We subfractionated the polyA− RNA fractions according to RNA size on silica-columns (Figure 7D) and transfected the fractions into COP-5 cells as described above. Only the fraction containing large (>200 nt) RNAs induced IFN-β production, whereas smaller RNAs (<200 nt), even when used in a 10-fold molar excess, failed to do so (Figure 7E). Consequently, the IFN-β that was induced by the polyA− RNAs isolated from the infected MEF cells was not dependent on RNase L and, importantly, could not be attributed to the presence of degradation products from the viral dsRNA, which could not be detected by the northern hybridization analysis due to their small size (Figure 7B). Taken together, these results demonstrate that polyA− RNAs isolated from SFV infected MEF cells contain large amounts of potent IFN-β inducers. These inducers are not viral RNAs and therefore must have been produced by the viral replicase using host cell RNAs as templates as it was observed in case of cells transfected with pRep.
To further characterize PAMPs present in polyA− RNA fraction of SFV infected cells the importance of 5′-ppp structure and double-stranded nature of these RNAs were analyzed. We found that neither RNA denaturation nor γ- and β-phosphates removal from the RNA 5′-end alone did not diminish IFN induction substantially. Combination of RNA 5′-end phosphates removal with subsequent denaturation, however, resulted in almost complete loss of the IFN signal (Figure 8A). This result indicated that polyA− RNA fractions of infected cells contained the mixture of PAMPs. Second, to analyze the properties of these PAMPs in a more detail, the SFV4-Rluc polyA− RNA fraction was resolved by performing electrophoresis on a native low melting agarose gel and different RNA species were excised and extracted. Subsequently, obtained RNA species were treated with various enzymes and transfected into COP-5 cells (Figure 8B). RNase T1 treatment of the isolated RNA species did not have any effect on their capacity to trigger IFN-β, indicating that all analyzed fractions contained dsRNA. Alkaline phosphatase treatment, however, substantially reduced the IFN induction only for RNA species co-migrating with cellular 28S and 18S rRNA (Figure 8B). Importantly, similar structural properties (RNase T1 insensitivity, AP sensitivity) were observed for the IFN inducing RNA generated by Rep expression in COP-5 cells (Figure 3D).
Finally, to prove that SFV replicase indeed transcribes host cell RNAs producing 5′-ppp dsRNAs in the context of viral infection we attempted to identify an example of such RNA. Since IFN-inducing RNAs co-migrated with 28S and 18S rRNAs (Figure 8B) it was obvious that multiple cellular RNAs are recognized and used by viral replicase. Therefore an identification of any of such RNAs represents a considerable challenge. It was recently demonstrated that RIG-I activation depends critically on 5′-ppp structure only for a short dsRNA (∼40 bp), whereas increasing the length of the dsRNA compensates for the 5′-ppp removal [60]. Therefore the previous results indicate that we were searching for a 5′-ppp RNAs (Figure 8A–B), complementary to a polyA− host RNAs longer than 200 nucleotides and forming ∼40 bp dsRNA structure (Figures 7E and 8B). Consequently, we had to adopt the recently described method used for micro RNA (miRNA) cloning, based on sequential linker ligations to 3′- and 5′-ends of the miRNA. In this procedure, the ligation of the 3′-linker is a very efficient process, whereas the efficiency of 5′-linker ligation may differ 100-fold depending on the target miRNA [61]. Several additional difficulties, however, were associated with cloning of viral replicase-generated 5′-ppp RNAs. First, while 5′-p miRNA may have some secondary and tertiary structure, they are essentially single-stranded RNA molecules, whereas the 5′-ppp RNAs generated by the viral replicase form dsRNA structures. Second, every ligation step in miRNA cloning is verified by the electrophoretical mobility shift and ligated miRNA can be directly purified, whereas due to the underrepresentation (for example as compared to miRNA) and/or heterogeneity of 5′-ppp RNAs generated by SFV replicase such manipulations were not possible. Consequently, both ligation reactions were performed in a single tube (“one-pot synthesis”). Therefore the second and the most critical RNA denaturation step was not particularly efficient due to the presence of buffer and proteins. In addition, due to higher concentrations of reactants, intermolecular ligation of 3′- and 5′-linkers was clearly more efficient than ligation to desired RNA molecule. Furthemore, host cell small RNAs (including miRNAs), could also interfere with the cloning of the replicase-generated RNAs due to the presence of a 5′-monophosphate (α-phosphate).
Taking into account all these above listed considerations a procedure illustrated in Figure 8C was developed. Briefly, we first depleted polyA− RNA samples from small RNA species by three rounds of purification on size-exclusion silica columns. Subsequently these RNAs from mock-, SFV4-Rluc-, and SFV4-Rluc-RDR-infected cells were incubated in the reaction buffer either in the absence (negative control) or presence of the RNA 5′ polyphosphatase. The purified RNAs were subjected to sequential ligation with 3′- and 5′-linkers (Figure 8C). Because the 5′-ppp structure does not allow the ligation of the 5′-linker, no unique clones should be present in the control samples which were not treated with RNA 5′ polyphosphatase. After performing reverse transcription and PCR we detected a unique discrete amplicon of approximately 100 bp only for the RNA 5′ polyphosphatase-treated polyA− RNA fraction isolated from SFV4-Rluc-RDR-infected cells (Figure 8C, lane 6). In contrast, no specific amplification was observed for polyA− sample originating from SFV4-Rluc infected cell. Most likely this was caused by the inefficiency of the developed detection procedure.
The results indicated that either a population or a single RNA of ∼30–40 nucleotides were successfully amplified. Subsequently, we excised the gel fragment containing this ∼100 bp amplicon and the corresponding fragments from remaining samples for DNA purification and cloning into a plasmid vector. As expected considerably more clones were obtained for RNA 5′ polyphosphatase treated sample from SFV4-Rluc-RDR infected cells. To avoid possible biases approximately half of clones, obtained for each probe, were sequenced. The analysis of obtained sequences revealed that the majority of the clones contained an amplified side-product of type 2, likely present due to its trace amounts after gel purification (Figure 8C, bottom). Unexpectedly, we found an additional type of clones that contained two copies of side-product of type 2, linked by a bacterial sequence, suggesting that a recombination leading to duplication took place. Almost 30% of the clones containing this type of inserts were generated as a result of the specific amplicon cloning, whereas in other samples these inserts constituted only a small fraction (2–8%). Consequently, conventional cloning is not an efficient approach for the analysis of amplicons representing the products of viral replicase and other methods such as direct analysis of PCR products by deep sequencing should be used. Nevertheless, one unique clone was obtained for the RNA 5′ polyphosphatase treated sample from SFV4-Rluc-RDR infected cells, providing an example from the putative population of ∼30–40 nucleotide RNAs (Figure 8D). This clone (named R2) contained a 33-nt copy of RNA, which was perfectly complementary to mouse antisense non-coding mitochondrial RNA 1 (ASncmtRNA-1, GenBank: GU332589.1), an equivalent of human ASncmtRNA-2 transcript (GenBank: EU863790.1) [62], [63], [64], [65]. The function of the human ASncmtRNAs is unknown, however it was demonstrated that these transcripts, co-migrating with the 18S rRNA and having a stem-loop structure, are down-regulated in tumor cell lines and tumor cells present in 17 different tumor types [63]. Furthermore, the possibility that R2 clone corresponded to a fragment of 16S rRNA encoded by mitochondrial precursor RNA (GenBank: V00665.1) was excluded due to the following reasons. First, the presence of ∼30 nt RNA fragment corresponding to 16S rRNA 3′-terminus would indicate that the integrity of RNA had been compromised. This was, however, highly unlikely since no ribonuclease activity was detected in either of the components of the reaction mixtures used for RNA enrichment and tagging (Figure 8C). Second, if R2 would have resulted from normal cellular RNA rather than from RNA containing 5′-ppp generated by SFV replicase, amplicons of corresponding size and similar clones should have been detected also for other samples, which was not the case. Third, R2 clone has an extra adenosine residue at its 3′-end (…GUUA) that is not present at the 3′ terminus (…GUU) of the major processed form of 16S rRNA [66], [67], indicating the absence of a perfect match between these sequences. Fourth, minor processed forms of the mouse 16S rRNA have GUUAU, GUUAUU, GUUAUUAGG, or GUUAUUAGGG sequences at their 3′-termini [66], however, these sequences were not present in R2 clone either. Finally, it was demonstrated that a small fraction of 16S rRNA contained polyA tails [66], [68] and was recoverable either by oligo(dT)-affinity chromatography [68] or by RT PCR [69], indicating that the presence of the polyA tail consisting of a single adenosine in the 16S rRNA is highly unlikely.
Moreover, the R2 clone corresponded to all PAMP criteria listed above – in addition to having expected length it was complementary to host polyA− RNA longer than 200 nt. Accordingly, as there is no cellular RdRp activity capable for synthesis of such RNA, these results strongly suggest that such RNA was synthesized by viral replicase. Moreover, the fact that we were unable to obtain this amplicon without the use of RNA 5′ polyphosphatase treatment indicated that this RNA had also 5′-ppp structure. Thus, an RNA molecule which represents an example of SFV replicase generated PAMP was experimentally identified. Based on our results (Figure 8B), it is also logical to assume that it does not represent the only PAMP of this kind generated by SFV replicase and its relative contribution to observed IFN induction is currently unknown. This as well as identification of full spectrum of sequences of SFV replicase generated PAMP RNAs represents topic for additional studies.
To subvert the vertebrate innate immune system, RNA viruses have evolved many different strategies. One of the most efficient of these strategies consists of blocking access of host sensors to viral dsRNA or 5′-ppp RNA. In the case of negative-strand RNA viruses and viruses with a dsRNA genome, the viral particles contain a replicase complex with genomic RNA. By retaining both their genomes and replicase in partially uncoated viral particles, negative-strand RNA viruses ensure that viral 5′-ppp RNA is not accessible to RLRs and that no exogenous (host cell) RNA can be transcribed. However, in the case of positive-strand RNA viruses, the viral particles deliver only viral genomic RNA into the cell, which serves as the mRNA for the synthesis of the viral replicase [70]. Therefore, it remains possible that host cell RNA could be transcribed by the viral replicase of positive-strand RNA viruses. At a later time in infection cycle, positive-strand RNA viruses remodel host cellular membranes to accommodate their replication complexes and thereby shield their own 5′-ppp dsRNA [71], [72].
In this study, we report a novel mechanism by which positive-strand RNA viral infection is sensed by the innate immune system. To demonstrate the importance of this mechanism for virus infection, we also describe a mechanism by which MEFs restrict the replication of a non-pathogenic SFV-Rluc-RDR. While others have shown that the mutant virus triggered increased IFN-β induction, we also demonstrated that this led to the destruction of positive-strand viral RNAs and the shutdown of viral replication in primary MEFs (Figure 5). Thus, IFN-induction is undoubtedly important for limiting SFV infection. Here, we delineate the mechanism that mediates this important antiviral response.
To understand the mechanism by which alphaviruses trigger potent IFN-β production, we dissected the replication of SFV by decoupling the expression of the replicases of the wild-type and mutant viruses from their viral replication-competent RNA templates. First, it was shown that the replication of the short virus-like RNA template by these replicases triggered the production of IFN-β by COP-5 cells (Figure 1C). However, it was subsequently found that these replicases were fully capable of IFN-β induction even when expressed in the absence of any virus or virus-like RNA template (Figure 1D). In both cases, the mutant replicase induced increased IFN-β production, as compared to the wild-type replicase, and we observed that the viral replicase RdRp activity was responsible for IFN-β induction (Figures 1D–E). Moreover, we found that only the viral replicase with intact RdRp activity generated IFN-inducing RNA, which triggered the IFN-β production (Figure 3A). Our results strongly suggest that during SFV infection, two parallel processes are driven by the viral replicase; in the first process, the viral replicase drives the replication and transcription of the replication competent viral genome, whereas in the second process, the viral replicase drives the transcription of non-viral host cell RNA templates (Figure 9).
The IFN-inducing activity of the viral replicase-generated RNAs was resistant to RNase A and RNase T1, although this activity was sensitive to RNase III and alkaline phosphatase, which indicated that this ligand is a dsRNA containing at least one terminal phosphate (Figure 3D). Our interference experiments indicated that RIG-I was the primary sensor detecting IFN-inducing RNA produced by the SFV replicase (Figure 2B). This result also strongly suggests that this RNA contains a 5′-ppp. The immunofluorescence results indicated that IFN-inducing RNA generated by the SFV replicase contained duplex RNA regions exceeding 40 bp in length (Figures 4B and 4C). Thus, SFV replicase transforms host cell RNA into PAMPs that are recognized and counteracted as non-self entities. This strategy of innate immune recognition of a viral replicase is different from pattern recognition as the replicase generates novel PAMP structures not uniquely associated with a pathogen. Detection of viral replicase RdRp activity by vertebrate host cells may represent an ancient mechanism for the recognition of non-self enzymatic function.
The majority (∼90%) of the IFN-β-inducing RNA generated by the SFV replicase was non-polyadenylated (Figure 3B). In this regard, it is interesting to note that non-polyadenylated transcripts comprise almost half of the human and mouse transcriptomes, although the biological function of these transcripts remains unclear [73]. Therefore, it will be of significant interest to identify which non-polyadenylated transcripts encoded by the host genome are involved in viral replicase detection. Our results also exclude the possibility that SFV replicase-triggered IFN-β induction is mediated by the antiviral endoribonuclease RNase L, which has been reported to cleave host cell RNA to generate small RNAs with lengths less than 200 nt upon viral infection [45]. In fact, we found that host cell RNAs longer than 200 nt are modified by SFV replicase to induce IFN-β (Figure 3A). Subsequently, identical characteristics were observed for the PAMPs present in the non-polyadenylated RNA fractions isolated from SFV-infected MEF cells (Figure 7D,E).
Consistent with previously published results obtained using bone marrow-derived dendritic cells [32], we demonstrated that MEFs detect SFV using the MDA-5 and RIG-I sensors (Figure 2C). We observed that detection of SFV4-Rluc-RDR heavily relied on both MDA-5 and RIG-I, whereas for detection of the viral replicase, RIG-I was the primary sensor (Figure 2B). This discrepancy is likely due to the fact that long viral dsRNA replication intermediates are generated during SFV infection. It was previously demonstrated that MDA-5 and RIG-I detect long and short dsRNAs, respectively [5], and these findings further support the idea that the viral replicase drives both viral RNA replication and transcription using host cell RNA templates. In both cases the predominant type of RNA ligand produced by SFV replicase dictates the primary type of sensor responsible for ligand detection and IFN induction.
Due to codon-optimization, the mRNA encoded by pRep and pRep-RDR lacked any cis-sequences that are known to participate in SFV RNA replication; hence, the PAMPs generated in the pRep-RDR transfected cells were non-viral. More importantly, our study reveals that host cell RNAs were used as alternative templates by the viral replicase during virus infection, demonstrating for the first time that during positive-strand RNA virus infection, both viral RNAs and cellular RNAs that are modified by the viral replicase contribute to IFN-β induction. The latter conclusion is based on the following data. Fractionation of total RNA from infected primary MEF cells by oligo(dT)-affinity chromatography led to a 15-fold enrichment of the major viral RNA species in the polyadenylated RNA fraction compared with the non-polyadenylated RNA fraction (Figure 7A,B). Consequently, if the viral RNA is the only PAMP present in infected cells, then the polyadenylated RNA fraction should induce a 15-fold higher level of IFN-β compared with an equal amount of the corresponding non-polyadenylated RNA fraction. Remarkably, this was not the case, as polyadenylated RNA fractions from wild type and mutant virus infected cells induced only approximately 2-fold and 6-fold higher levels of IFN-β respectively (Figure 7C). These findings strongly suggest that non-polyadenylated RNA fractions isolated from infected MEF cells contain potent IFN-β inducers, most of which are not viral RNAs and were derived from the host cell RNA. Strikingly, this finding was especially compelling for the non-polyadenylated RNA species extracted from the MEF cells infected with the wild type virus. These results cannot be explained based on the current paradigm that type I IFN is triggered exclusively by viral dsRNA or viral RNA containing a 5′-ppp that are produced during the course of viral genome replication or transcription by viral replicases [5], [6], [12], [74], [75], [76].
Our results indicate that wild type and mutant SFV generate roughly the same amount of PAMPs. However, the amount of IFN-β that is induced by these viruses in MEF cells differs drastically (Figure 6) and, in the case of mutant virus, leads to the inhibition of viral replication (Figure 5). Obvious explanation to this conundrum is the presence of a single mutation in the nsP2 subunit of the viral replicase of SFV4-Rluc-RDR, which does not allow nsP2 to enter the nucleus of infected cell. It has been reported that the nuclear localization of nsP2 (for both SFV and SIN) is absolutely required for the suppression of the host cell antiviral response [20], [21], [27], [28]. Consequently, it is thought that nsP2 must suppress the innate immune system of the host cell to inhibit the sensing of the viral RNA PAMP(s) and/or response to the PAMPs. However, there are several arguments that challenge this hypothesis. First, cytosol-accessible SFV genomic (42S [+] strand) and subgenomic (26 [+] strand) viral RNAs, which direct the synthesis of viral replicase and viral structural proteins, are not PAMPs because their 5′-ppp is shielded by a cap-structure [55]. Second, the PAMP represented by the dsRNA replicative form (42S [±]) resides in a membrane-bound replication complex [48], [77]. Third, the 42S (−) strand, which does not have cap-structure [78] and, thus, may represent a PAMP, most likely does not exist as a free molecule (Figure 7B); instead, this strand is confined to the replication complex as part of the dsRNA. Indeed, our data clearly indicate that this RNA does not exist as a free molecule, as it was fully protected in the SFV4-Rluc-RDR infected MEF cells and, unlike free positive-strand RNA, was not degraded during the restriction of viral replication (Figure 5J). Thus, it appears that RLR sensors may have extremely limited access to viral PAMPs in infected cells. The only species that are presumably accessible to RLRs are DI-RNAs, which have a negative polarity and are double-stranded. However, although DI-RNA is presumably localized to the spherules, this has not been experimentally verified. DI-RNA may also contribute to IFN induction, but this contribution is likely to be minor, as their amount was found to be negligible compared with the full-length viral RNA species (Figure 7B). Moreover, the amounts of SFV DI-RNA produced during both wild type and mutant virus infection are very similar (Figure S4) and cannot explain the dramatic difference in IFN-β induction (Figures 5B and 6B). Still, a mutation in nsP2 results in the rapid detection of SFV4-Rluc-RDR infection (Figure 6B) and the restriction of viral replication by the host cell (Figure 5). Consequently, there are PAMPs that are easily accessible to RLR sensors in infected MEF cells. Our study strongly suggests that these easily accessible PAMPs are generated by viral replicase unspecific activity on host cell RNAs.
It is known that enzymatically active nonstructural viral proteins that participate in the formation of the viral replicase complex are produced in much larger amounts than are needed for virus RNA replication. This has been demonstrated for alphaviruses [36], [47] and HCV [79]. It has been proposed that these viral proteins play roles in processes other than viral RNA replication [36]. For example, the NS3/4A protease of HCV cleaves mitochondrial antiviral protein, precluding HCV detection by the host cell [80]. Proteases of picornaviruses cleave multiple host factors, including eIF4G, precluding the translation of host cell capped mRNAs [81]. Though the list of the enzymatic activities of viral proteins that target host cell components is long and rapidly growing, it is clearly biased towards the activities that are beneficial for virus infection. At the same time, it is hard to believe that this process is one-sided. The host cells have also evolved to take advantage of some of these enzymatic functions specific to viruses. In this regard, the use of these activities to detect the presence of a pathogen is a clear possibility that has not been considered. There are many benefits to recognizing RdRp activity of viral replicases rather than viral RNA PAMPs. First, especially in the case of in vivo infections, not all of the cells are permissive for viral RNA replication. However, entry of viral RNA into such cells will trigger the production of viral replicase, the RdRp activity of which then can be detected by the host cell. Second, all RNA viruses inevitably produce divergent progeny, including many viruses that are replication defective, which are also packed into virions. As long as the defective RNAs are capable of expressing a functional replicase, the synthesis of nonviral PAMP RNAs will occur.
In this study, we have developed a procedure that allowed us to selectively amplify the RNAs produced by the SFV replicase in the context of SFV4-Rluc-RDR infection (Figure 8C). However, because of the inefficient conventional cloning approach only a single non-polyadenylated RNA used by SFV replicase was identified (Figure 8D). It has recently been demonstrated by deep sequencing analysis that heterogeneous host cell RNAs (22 mRNAs, 3 non-coding RNAs, and 2 pseudogenes) are modified by the HCV RdRp when expressed alone in the mouse liver [41]. Our data strongly suggests that non-polyadenylated RNAs co-migrating with 28S and 18S rRNAs are the primary targets for the SFV replicase (Figures 3B, 7C, 7E, 8A, and 8B). We demonstrate that PAMPs produced from host cell RNAs are an important byproduct of alphavirus replicase (nsP1/nsP2/nsP3/nsP4) RdRp activity during infection and that the nsP2 protease subunit is used to counteract the consequences of this activity. For HCV, the activity of NS5B RdRp is counteracted by NS3/4A protease [13], [82], [83]. To date, the replicases of HCV, norovirus, and TMEV have been shown to be capable of inducing IFN without the requirement of viral RNA replication [13], [14], [15], [16]. Direct comparison of the IFN induction by SFV replicase and HCV RdRp showed that both replicases are very efficient PAMP inducers (Figure 1F). Therefore, together with our study, these findings suggest that the activation of the innate immune response by the replicases of positive-strand RNA viruses may be a general property used by the host cells to counteract viral invasion. If so, the question of why viruses have not evolved a way for their replicase to be more specific for their own RNAs remains. The likely reason is that the increased specificity of the replicase is not beneficial for the survival and/or evolution of the virus. Alphavirus replicases possess an outstanding ability to recognize defective cis-elements, repair such defects and rescue the replication in their genomes [57]. Thus, even if it were possible to obtain a polymerase with higher template specificity, the cost would be too high. This theory parallels reports studying the error prone nature of the same viral RNA replicases. As it was elegantly demonstrated, the fidelity of the picornavirus RNA polymerase can be easily increased [84], but this increased fidelity results in reduced virus fitness and virulence [85]. Therefore, as discussed above, viruses have instead armed themselves with safeguards that are effective against innate immune recognition irrespective of the origin of the PAMPs.
pRep was generated by inserting codon-optimized coding sequence of SFV replicase (EMBL-Bank: HC198689) into either KS plasmid (Stratagene) or GTU eukaryotic plasmid expression vector [86]. To generate pRep-RDR, we introduced mutations C4129G, G4130A, and G4131C (nucleotide numbering as in HC198689) leading to RRR→RDR mutation in nsP2 nuclear localization sequence of SFV replicase. To obtain pRep-RDR/GAA plasmid DNA, we introduced additional mutations A7424C and A7427C, inactivating the nsP4 RdRp catalytic centre (GDD→GAA). pRep, pRep-RDR, and pRep-RDR-GAA contained intron sequence from rabbit beta-globin gene (GenBank: V00882.1), which was inserted into SFV replicase coding sequence. The expression of SFV replicase polyprotein was under control of either hEF1α/HTLV composite promoter (increased expression) or Rous sarcoma virus long terminal repeat (RSV LTR). For the generation of pNS5B and pNS5B-GND, the coding sequence of NS5B Con1 genotype 1b (GenBank: AJ238799) was used. To keep total DNA amount constant during dose-dependent IFN-induction experiments, transfections were performed with “stuffer” DNA plasmid vector encoding d1EGFP.
Plasmid pSFVmin was constructed from two fragments. First fragment of SFV cDNA corresponding to the nucleotides 1–274 of SFV genome was placed under control of CMV promoter and cloned into KS plasmid (Stratagene). Second fragment consisting form polylinker (BstB1-PmeI-BglII-SpeI-NotI), complete cDNA copy of 3′ UTR of SFV4 with polyA sequence of 60 residues, hepatitis delta virus negative strand ribozyme and SV40 early transcription terminator was cloned immediately downstream of the first fragment. To obtain pSFVminRluc the coding sequence of Rluc was amplified by PCR and inserted to pSFVmin using SpeI-NotI restriction sites. In the resulting construct pSFVminRluc Rluc reporter is expressed in form of fusion protein containing 78 foreign aa residues at its N-terminus, first 63 of them representing the N-terminal region of nsP1 of SFV.
The antibodies for mouse RIG-I (R37) and LGP2 were purchased from Immuno-Biological Laboratories Co., Ltd. The antibody for mouse MDA-5 (AL180) was purchased from Alexis Biochemicals. The antibody for dsRNA (J2) was purchased from Scicons. Antibody for β-actin (C4) was purchased from Santa Cruz Biotechnology, Inc. The antibody for LAMP2 (H4B4) was purchased from Abcam. SFV nsP2 and nsP4 antibodies were kindly provided by Dr. Tero Ahola, antibodies against nsP1 and nsP3 of SFV were made in-house. Secondary antibodies used in immunofluorescence were purchased from Life Technologies.
Poly(I:C) and chloroquine were purchased from Sigma-Aldrich. RNase A, RNase T1, and RNase III were purchased from Ambion. Alkaline phosphatase and DNase I were purchased from Roche Applied Science and Promega respectively. T4 RNA Ligase I and β-Agarase I were purchased from New England Biolabs. RNA 5′ polyphosphatase was purchased from Epicentre (Illumina).
COP-5 and RD cells were maintained in L-glutamine-containing IMDM medium supplemented with 10% fetal bovine serum (FBS), and antibiotics. Primary MEF cells were purchased from Millipore (EmbryoMax Primary Mouse Embryo Fibroblasts, Not Mytomycin C Treated, Strain CF1, passage 3; Catalogue Number: PMEF-CFL) and cultured in L-glutamine-, sodium pyruvate-, and high glucose-containing DMEM medium supplemented with 15% FBS, 0.1 mM β-mercaptoethanol, and antibiotics up to passage 6. Transfection of DNA (0.2 µg/ml) and RNA (0.2–0.4 µg/ml) into COP-5 cells (0.25–1.0×106 per 60-mm dish) was carried out using Lipofectamine 2000 (Invitrogen). RD cells (0.2×106 per 60-mm dish) were transfected with DNA (0.2 µg/ml) by electroporation in GenePulser Xcell (Bio-Rad) instrument (settings: Exponential Wave, 190 V; 975 microfarads [µF]) in 400 µl OptiMEM medium (Invitrogen). RNA (2 µg/ml) and DNA (10 µg/ml) into MEF cells (0.25×106 per 60-mm dish) were transfected by electroporation (Exponential Wave, 240 V; 975 µF) in 400 µl OptiMEM medium. All electroporations were performed in 4-mm cuvettes (Thermo Fisher Scientific).
SFV4-Rluc and SFV4-Rluc-RDR were constructed, rescued and propagated as previously described [44]. The viral stocks were titrated using plaque titration on baby hamster kidney (BHK)-21 cells. For the infection of MEFs, which are considerably less susceptible to SFV4 infection than BHK-21 cells, the relative titers and MEF-specific MOIs were re-calculated based on the infectivity of the recombinant viruses, as measured by immunostaining of infected cells. The viral infection of primary MEFs was performed in OptiMEM supplemented with 0.1% FBS for 1 hr. The virus was subsequently aspirated, and fresh medium was added.
An infectious center assay was performed essentially as previously described with minor modifications [87]. Briefly, 1 µg of RNA fractions extracted from infected MEF cells were electroporated (two pulses 850 V, 25 µF, in 800 µl) into BHK-21 (baby hamster kidney) cells. Tenfold dilutions of electroporated cells were seeded into six-well plates containing monolayers of naïve BHK-21 cells. After a 2 hr incubation at 37°C, the cell culture medium was aspirated, and the wells were overlaid with a medium containing carboxymethyl cellulose. After 2–3 days, the plaques were visualized by crystal violet staining and counted.
The amount of IFN-β secreted into the cell culture medium was measured using a commercial Verikine Mouse Interferon-Beta ELISA kit (PBL InterferonSource), according to the manufacturer's instructions.
Cells were harvested (scraped with a rubber policeman) at different time points in a phosphate-buffered saline (PBS) on ice. Subsequently, cells were lysed with Renilla Luciferase Assay Lysis Buffer according to the manufacturer's instructions (Promega). Lysate was mixed with Renilla Luciferase Assay Substrate diluted 100-fold in Renilla Luciferase Assay Buffer, and luminescence was measured on the GloMAX 20/20 Luminometer (Promega).
We used the algorithm developed in-house to design 21-nt siRNA oligos, which had 19-bp perfect match duplex and 2-nt 3′-overhangs. The sequences of siRNA oligonucleotides used in the study are as follows (siRNA duplex name, guide strand [5′→3′], passenger strand [5′→3′]): dhx58_mus_2304, UUCUUAGAACAUCAUGGCAUA, UGCCAUGAUGUUCUAAGAACU; ifih1_mus_3004, AUUGACAUGAUGCAUCUUCUC, GAAGAUGCAUCAUGUCAAUAU; ddx58_mus_2678, AUAUCUUCCACGACGAAACUU, GUUUCGUCGUGGAAGAUAUUG. siRNA duplexes were synthesized and annealed by Proligo (Sigma-Genosys), whereas negative control non-targeting siRNA #4611 and #4635 was purchased from Ambion.
siRNA oligonucleotides at a final concentration of 20 nM were reverse-transfected into COP-5 cells (0.25–1.0×106 cells per 60-mm dish) using 5 µl Lipofectamine RNAiMAX (Invitrogen). For primary MEFs (0.25×106 cells per 60-mm dish), siRNA duplexes at a final concentration of 100 nM were transfected by electroporation in 4-mm cuvettes (Thermo Fisher Scientific) using a GenePulser Xcell (Bio-Rad) instrument (settings: Square Wave, 1000 V; 2 pulses, 0.5 ms) in 100 µl OptiMEM medium. Unless otherwise indicated, on the third day of culture, cells were transfected with 0.2 µg/ml of plasmid DNA or RNA using 5 µl Lipofectamine 2000 (Invitrogen) or infected by SFV4-Rluc-RDR. The amount of IFN-β in the cell culture medium was measured, and cells were harvested for the immunoblotting analysis on either the fourth or fifth day of culture.
Proteins in cell extracts were resolved on 10% polyacrylamide/SDS gels in Mini PROTEAN Tetra Cell systems (Bio-Rad). Subsequently, proteins were transferred to Immobilon-P (Millipore) polyvinylidene fluoride microporous 0.45 µm membranes using Trans-Blot Semi-Dry Transfer Cell apparatus (Bio-Rad). Blots were incubated with various primary antibodies. Secondary goat anti-rabbit and anti-mouse antibodies conjugated with horseradish peroxidase were from LabAs Ltd. Immunoreactive bands were detected by enhanced chemiluminescence (ECL) (GE Healthcare) and subsequent exposure to X-ray film (SuperRX, Fuji).
Total RNAs, large RNAs (>200 nt), and small RNAs (<200 nt) were extracted from cells and purified using the mirVana miRNA Isolation kit (Ambion) or TRIzol reagent (Invitrogen), according to the manufacturer's instructions. Oligo(dT)-Cellulose Type 7 (GE Healthcare) was used for affinity-chromatography fractionation of total RNA into polyadenylated (polyA+) and non-polyadenylated (polyA−) RNA species. For RNA extraction from native low melting agarose gel, β-Agarase I enzyme was used accordingly to manufacturer's protocol (New England Biolabs). When required, PD-10 desalting columns (GE Healthcare) containing Sephadex G25 were used for buffer exchange. The resulting OD260/OD280 and OD260/OD230 ratios for all RNA preparations exceeded 2.1, as determined by measurements obtained using a ND-1000 spectrophotometer (NanoDrop Technologies, Inc.). The integrity of the RNA was confirmed by denaturing formaldehyde agarose gel electrophoresis. Alternatively, RNA was resolved on non-denaturing agarose gel electrophoresis, stained with ethidium bromide, its image was recorded and analyzed by NIH ImageJ 1.46 software (http://rsb.info.nih.gov/ij/download.html).
Two micrograms of nucleic acid were treated with DNase I (0.1 U/µl) and alkaline phosphatase (0.1 U/µl) at 37°C and 50°C, respectively, for 1 hr in a volume of 20 µl. For RNase digestion experiments, 2 µg of RNA was digested with RNase A, RNase III, or RNase T1 at the specified concentrations at 37°C for 1 hr in a volume of 20 µl. The undiluted (1×) RNase concentrations used in the reactions were 1 µg/ml (RNase A), 1 U/µl (RNase III), and 1 U/µl (RNase T1). Enzyme-treated RNAs were precipitated with ethanol in the presence of sodium acetate and glycogen prior to transfection.
Sub-cellular fractionation was performed as previously described [88]. In brief, pRep-RDR or pRep-RDR/GAA transfected COP-5 cells (4–6×107) were harvested, washed, and resuspended in 800 µl of HB buffer (8.6% sucrose, 3 mM Imidazole, pH 7.4) supplemented with protease inhibitors (Roche). Subsequently, cells were homogenized by passing homogenate through 22G1 ¼ needle mounted onto 1-ml syringe until the ratio of unbroken cells to free nuclei was 10% to 90%, as examined under microscope. Unbroken cells and nuclei were pelleted by centrifugation and post-nuclear supernatant (PNS) collected. Concentration of sucrose in the PNS was adjusted to 40.6% using 62% sucrose solution and refractometer. PNS was loaded in the bottom of an SW41 centrifuge tube and overlaid with 4.5 ml of 35% sucrose, 3 ml of 25% sucrose, and 3 ml of 8.6% sucrose cushions. All sucrose cushions also contained 3 mM Imidazole pH 7.4 and 1 mM EDTA. Tubes were centrifuged for 1.5 hr at 35000 rpm in a Beckman Optima L-90 K ultracentrifuge at 4°C. After centrifugation the pellet, containing cytosolic ribonucleoprotein complexes and whitish bands of membrane particles at every interphase between sucrose cushions were collected. Subsequently, RNA was extracted from each fraction with TRIzol Reagent (Invitrogen). Before isopropanol precipitation 40 µg of RNA-grade glycogen was added to each sample for maximal recovery of RNA.
COP-5 or RD cells transfected with pRep, pRep-RDR, or pRep-RDR/GAA were washed twice with phosphate-buffered saline (PBS), fixed with 4% paraformaldehyde in PBS for 10 min at 22°C, and permeabilized with 0.5% Triton X-100 in PBS for 5 min at 22°C. Blocking and antibody binding was performed in two different ways. First, samples were treated with block buffer (10% goat serum and 1% BSA in PBS) for 1 hr at 22°C, and then incubated for 1 hr at 22°C with antibodies against LAMP2 (H4B4, Abcam) and nsP1 diluted in antibody-binding buffer (3% BSA and 0.05% Tween 20 in PBS). Second, samples were incubated in block buffer (10% goat serum, 1% BSA, and 0.2% Triton X-100 in PBS), and consequently with antibodies against dsRNA (J2) and nsP1 in 3% BSA, 0.2% Triton X-100, and 10 mM MgCl2 antibody-binding buffer. Antibody binding was detected using appropriate antibodies conjugated with Alexa fluor 488 and 568 (Invitrogen). Specifically, for H4B4 and J2 binding detection, secondary antibodies purchased from Life Technologies and reacting with the Fc portion of the heavy chain of mouse IgG1 (A-21121) and IgG2a (A-21134) were used respectively. SlowFade Gold antifade reagent with DAPI (Invitrogen) was used for counterstaining of cells nuclei. Samples were imaged on a Nikon ECLIPSE TE2000-U inverted microscope and recorded with Nikon DXM1200C Digital Camera. Images were collected using 60× immersion objective and processed with Nikon Capture NX2 and ACT-1C software.
Oligo(dT)-Cellulose type 7 powder (GE Healthcare) was suspended in sterile water and resulting gel was used for the preparation of gravity-flow chromatography columns. Subsequently, oligo(dT) columns were washed with 10 volumes of water and 5 volumes of 100 mM sodium hydroxide (pH∼10). The pH of the oligo(dT)-columns was brought to 7.5 by equilibrating it with 10 volumes of TEN0 buffer (10 mM TrisHCl, 1 mM EDTA, pH 7.5) and subsequently with 10 volumes of TEN500 (10 mM TrisHCl, 500 mM sodium chloride, 1 mM EDTA, pH 7.5). Total RNA samples were heat-denatured (70°C, 10 min) in water, chilled on ice and loaded onto columns in TEN500 buffer. Then, unbound nonpolyadenylated RNAs were collected. Oligo(dT) columns with bound polyadenylated RNAs were extensively washed with 30 volumes of TEN500 buffer. Finally, bound RNA was eluted with 10 volumes of TEN0 and RNA-containing fractions were pooled.
RNA samples were denatured in loading buffer (1×MOPS, 50% formamide, and 6% formaldehyde) for 5 min at 100°C, chilled on ice and separated on 1% agarose 6% formaldehyde-containing denaturing gel using 1×MOPS buffer system (5 V/cm, 5 hr, 4°C). Consequently, samples were transferred to nylon membranes (Hybond N+, GE Healthcare) in 10X SSC using capillary blotting technique for 16 hr. After UV cross-linking at 0.12 J/cm2 the membrane was blocked in hybridization solution (water-reconstituted DIG Easy Hyb Granules [Roche]) at 65°C for 30 min in the hybridization vessels. The RNA probes were generated with phage T7- and T3-RNA polymerases run-off in vitro transcription using linearized DNA as templates and labeled with digoxigenin-UTP-containing RNA labeling mix (Roche). Consequently, RNA probes were treated with DNase I to remove template DNA, purified on Illustra S-300HR Columns (GE Healthcare), denatured and added to the hybridization vessels at 100 ng/ml for hybridization at 65°C for 18 hr. The membranes were washed twice in 2X SSC, 0.1% SDS for 10 min at 25°C and then twice in 0.1X SSC, 0.1% SDS for 15 min at 65°C. Consequently, membranes were washed, blocked, and incubated with Fab fragments against digoxigenin, conjugated to alkaline phosphatase for detection of hybridized signals using CDP-Star (DIG Luminescent Detection Kit, Roche Applied Science) according to manufacturer's instructions. Finally, membranes were exposed to SuperRX X-ray films (Fuji) or for longer exposure times in ImageQuant RT ECL system (GE Healthcare).
Model dsRNA for RNase digestion experiment was generated by annealing of two single-stranded RNAs generated by T7 and SP6 RNA polymerases via run-off transcription from the plasmid DNA containing hepatitis C virus subgenomic replicon sequence fragment (∼4000 bp) flanked by corresponding promoters. RNAs were dissolved in RNA annealing buffer (10 mM TrisHCl, 20 mM NaCl, pH 7.5), denatured for 1 min at 98°C, then incubated at 75°C for 10 min, and finally cooled to room temperature during 1 hr.
DIG-labeled full length SFV4-Rluc RNA was incubated in 1× alkaline hydrolysis buffer (50 mM NaHCO3/Na2CO3, 1 mM EDTA, pH 9.2) at 95°C for 6 min and rapidly cooled to 4°C. Subsequently, fragmented RNA was purified using RNeasy Mini Kit (QIAGEN).
Small RNA species (∼20–23 nt) were depleted from polyA− RNA samples by three rounds of purification using RNeasy Mini Kit (QIAGEN). Subsequently RNAs were incubated in the reaction buffer either in the absence (negative control) or presence of the RNA 5′ polyphosphatase (Epicentre, Illumina). After purification using RNeasy Mini Kit (QIAGEN), the RNAs were subjected to sequential ligation with pre-adenylated and blocked at its 3′-terminus full-DNA 3′Linker (IDT Linker-1, 5′-rApp-CTG TAG GCA CCA TCA AT-ddC-3′, Integrated DNA Technologies) and full-RNA 5′Linker (5′-GCC ACC UCG AGU CAC ACC GUA AGU UUC-3′ [89]) essentially as described previously with minor modifications [61]. First, we used T4 RNA Ligase 1 (New England Biolabs) for both steps. Second, both ligation reactions were performed in a single tube (“one-pot synthesis”). Third, after second denaturation step additional ligase was added. Ligation mixtures were used directly for reverse transcription and PCR.
RNA (100 ng) was reverse-transcribed using the SuperScript III First-strand Synthesis System for RT-PCR (Life Technologies) in a final volume of 20 µl according to the manufacturer's instructions with minor procedure modifications. Briefly, the RNA, the primer, and dNTPs were incubated for 2 minutes at 95°C and cooled quickly on ice. The remaining components (reverse transcriptase buffer, reverse transcriptase, MgCl2, dithiothreitol, RNase inhibitor, and water) were mixed and added on ice. The reactions were initiated by shifting the temperature to 50°C for 2 hr and stopped by heating at 85°C for 5 minutes. Subsequently, RNAs were removed by RNase H treatment. Two microliters of the RT reaction mixture was used for subsequent PCR analysis, which was performed either with the Phusion or with the Dynazyme II polymerases (Thermo Scientific) according to the manufacturer's instructions. For strand-specific RT, we used the HPLC-purified primers (Microsynth) 5′SFV (5′-ATG GCG GAT GTG TGA CAT ACA CGA C-3′) and 3′SFV (5′-GGA AAT ATT AAA AAC CAA TTG CAA AAT AAA ATA-3′) as previously described to efficiently amplify SFV DI-RNA [59] for negative and positive strand detection, respectively. Both primers were used for PCR amplification of the RT products. For the amplification of cDNA, corresponding to tagged SFV replicase generated products, we used HPLC-purified primers Y-adaptor-a (5′-GCC ACC TCG AGT CAC ACC GTA-3′) [89] and AF-JIG-37 (5′-CAA GCA GAA GAC GGC ATA CGA ATT GAT GGT GCC TAC AG-3′) [90], the latter primer was also used for RT.
|
10.1371/journal.pcbi.1003689 | Stability Curve Prediction of Homologous Proteins Using Temperature-Dependent Statistical Potentials | The unraveling and control of protein stability at different temperatures is a fundamental problem in biophysics that is substantially far from being quantitatively and accurately solved, as it requires a precise knowledge of the temperature dependence of amino acid interactions. In this paper we attempt to gain insight into the thermal stability of proteins by designing a tool to predict the full stability curve as a function of the temperature for a set of 45 proteins belonging to 11 homologous families, given their sequence and structure, as well as the melting temperature () and the change in heat capacity () of proteins belonging to the same family. Stability curves constitute a fundamental instrument to analyze in detail the thermal stability and its relation to the thermodynamic stability, and to estimate the enthalpic and entropic contributions to the folding free energy. In summary, our approach for predicting the protein stability curves relies on temperature-dependent statistical potentials derived from three datasets of protein structures with targeted thermal stability properties. Using these potentials, the folding free energies () at three different temperatures were computed for each protein. The Gibbs-Helmholtz equation was then used to predict the protein's stability curve as the curve that best fits these three points. The results are quite encouraging: the standard deviations between the experimental and predicted 's, 's and folding free energies at room temperature () are equal to 13 , 1.3 ) and 4.1 , respectively, in cross-validation. The main sources of error and some further improvements and perspectives are briefly discussed.
| The prediction of protein stability remains one of the key goals of protein science. Despite the significant efforts of the last decades, faster and more accurate stability predictors on the proteomic-wide scale are currently demanded. The determination and control of protein stability are indeed fundamental steps on the path towards de novo design. In this paper we develop a method for predicting the stability curve of proteins. This curve encodes the temperature dependence of the folding free energy (). Its knowledge is important in the study of protein stability since all the thermodynamic parameters characterizing the folding transition can be extracted from it. Our prediction method is based on temperature-dependent mean force potentials and uses the tertiary structure of the target protein as well as the melting temperature () and the heat capacity change () of some other proteins belonging to the same family. From the predicted stability curves, the , the and the at room temperature can be inferred. The predictions obtained are compared with experimental data and show reasonable performances.
| The understanding of the mechanisms used by nature to stabilize proteins against thermal inactivation is still an open issue of primary importance. From a theoretical perspective, such comprehension is fundamental in the study of the adaptive strategies used by the organisms to inhabit extreme environments. Due to evolution, such organisms are not only able to tolerate extreme temperature conditions, that range from less than ten degree Celsius to more than 120 , but require these conditions for their survival. The control of the thermal resistance is also important from an applicative perspective, as it would allow the optimization of a wide series of industrial, bioanalytical and pharmaceutical bioprocesses through the design and manufacture of new and more efficient enzymes [1]–[3].
In the last decades, different attempts and methods have been developed to obtain proteins of increased thermal stability. Protein engineering methods that include directed evolution methods [4]–[6] have been quite successful even if their applicability remains limited due to the intensive work required. In silico engineering approaches based on sequence conservation or free energy calculation methods have also been developed but with only partial success [7]–[12].
Recently, we developed a thermal stability prediction tool based on (melting)-temperature dependent statistical potentials that are derived from datasets in which only proteins with given thermostability properties are included [13]–[15]. The introduction of such potentials in the thermal stability framework is motivated by the fact that the amino acid pair interactions are temperature dependent, which means that some of them are more stabilizing than others in the high temperature regime and less stabilizing at lower temperatures (and vice versa) [16]–[29]. This peculiar approach allowed us to study the thermal properties of proteins without detour through their thermodynamical stability, which is advantageous since it is well known that the two types of stability are poorly correlated.
Proteins use different ways to promote their thermoresistance, which can – in a first approximation – be classified in three main strategies according to the Nojima analysis [30] (for a more recent review see also [31]). Let us start by introducing the stability curve of a protein, which can be described by the Gibbs-Helmholtz equation:(1)where is the free energy change associated to the folding transition from the unfolded to the native state, and the change in enthalpy and entropy measured at the reference temperature , and the change of the heat capacity across the transition. To obtain this equation, one has to fix the pressure of the system, to consider two-state transitions only, and to take as temperature independent. Usually, the melting temperature , which is the midpoint of the thermal denaturation, is chosen as the reference temperature. Eq.(1) can then be rewritten as:(2)where is the enthalpy measured at . Sometimes, the reference temperature is taken equal to , the temperature of maximal stability, which yields the equation: (3)
The first strategy that a protein can use to increase its thermostability [30] is to make the enthalpy change () measured at more negative. This yields an overall decrease of for all temperatures as we can see from Eq.(3) (Figure 1.a). In the second strategy, becomes less negative, which leads to an increase of through a modification of the shape of the curve (see Eq.(2) and Figure 1.b). The last strategy consists in an increase of the maximum stability temperature, , defined at the minimum of the curve, where the transition is purely enthalpic. This shifts the curve towards the high temperature region (see Figure 1.c).
It is, in general, not obvious to determine which type of strategy is adopted by a given protein; often several strategies are used in combination [31]. A realistic example of stability curve is depicted in Figure 1.d: the value of the folding free energy is plotted both for a thermostable protein, the -methyl-guanine-DNA methyltransferase from Thermococcus kodakaraensis (Tk-MGMT) with = 98.6 , and for its mesostable counterpart, the C-terminal Ada protein from Escherichia coli (Ec-AdaC) with = 54.8 , as determined experimentally in [32]. We can clearly see that in this case the three strategies are used simultaneously in the achievement of a higher thermal stability.
The strategies for improving the thermal resistance of a protein sometimes also improve the thermodynamic stability, defined by the folding free energy at room temperature (25 ), and sometimes not. The first strategy clearly does; for the other two strategies, it depends on the relative values of and (see Figures 1a–c).
It is unfortunately quite difficult to get accurate predictions of thermal stability. The results described in the literature are in general family-dependent and sometimes even contradictory [16]–[29]. Indeed, the temperature-dependent nature of the amino acid interactions makes the thermal stability analyses quite intricate and the mechanism behind it difficult to unravel. Predicting the thermodynamic stability is not easy either. There are no methods for predicting the thermodynamic stability of a given protein, with the notable exception of molecular dynamic simulations, which are however very time-consuming and not applicable on a large or medium scale. Only methods for predicting thermodynamic stability changes upon point mutations () have been developed and reach good scores [33]–[43]. No predictions of the enthalpy or entropy do exist either. In contrast, the prediction of is relatively easy since it is strongly correlated to the change of accessible surface area upon unfolding [44]–[46].
In this paper we go a step further than previous analyses aiming at evaluating either , or . We indeed present a method for predicting the whole stability curve of a protein from its sequence and structure, in the temperature range that is relevant for such systems (), using as main tool the temperature-dependent statistical potentials developed and tested in [13]. We would like to emphasize that this is, to our knowledge, the first prediction method that outputs the complete stability curve. To get a satisfactory performance, we used in the predictions some information about proteins belonging to the same homologous family, and more precisely their and . The predicted stability curve yields an estimation of the melting temperature , the thermodynamic stability , the temperature of optimal stability , the , as well as the enthalpy and the entropy at certain temperatures. We present our results in cross validation for a set of 45 proteins belonging to eleven homologous families (for the list of their PDB codes [47] and their characteristics, see Table S1 of Supporting Material). The predicted values are compared with the experimentally determined values when available, and the different strategies used by the proteins for thermal stabilization are investigated and discussed.
In this section we describe the main tools used in this analysis, namely the statistical potentials, and how they have been optimized for the current investigation. The main steps of our approach are schematically illustrated in Figure 2.
The statistical potentials are well known since some seminal papers [48]–[50]. They are derived from the frequency of associations between certain sequence and structure elements in a dataset of experimentally determined native protein structures. Even though such potentials have been extensively and successfully used in the analysis of the thermodynamic stability of proteins, they have only recently been applied in the thermal stability context, where the temperature dependence of the amino acid interactions must be taken into account [13]–[15]. To deal with this, potentials that depend on the melting temperature were derived from different datasets in which only proteins with given thermal properties were included. Three such datasets were considered [15]: a set containing only mesostable proteins, denoted and characterized by a mean value of the melting temperature of its entries () of about , a thermostable ensemble, denoted , with , and a reference set containing both mesostable and thermostable proteins, denoted , with . The list of proteins belonging to these datasets are given in Table S0–S11 and Table S13 of the Supplementary Material of [15].
From these different datasets, statistical potentials were derived using the standard formalism of the inverse Boltzmann law [13], [14]:(4)where is the relative frequency of observation of the sequence element associated to the structure element , and and are the frequencies of observation of the sequence element and of the structure element , respectively. In this computation, corresponds either to the amino acid type of residue along the polypeptide chain, or to the amino acid types of residues and , while is either the backbone torsion angle domain of residue , as defined in [51], or the spatial distance between the residues and . The former are called torsion potentials and the latter distance potentials.
While the torsion potentials describe local interactions along the chain and are a measure of the propensity of a given amino acid to adopt certain backbone torsion angles, the distance potentials describe the tertiary interactions and measure the propensity of amino acids to be separated by a certain spatial distance . The values of the distance between two residues, defined as the distance between the geometrical centers of the heavy side chain atoms, range between 3.0 and 8.0 and were grouped into 25 bins of 0.2 width, with two additional bins that contain distances larger than 8.0 and smaller than 3.0 , respectively.
Note that we have made the -dependence of the frequencies explicit to stress that these are computed from a dataset associated with specific thermal properties, characterized by . As a consequence, the potentials are -dependent and reflect the thermal characteristics of the dataset from which they are derived.
Due to the smallness of the dataset, some techniques are required to smooth the potentials and improve their performances. A first modification that has been performed is a correction for sparse data consisting in rewriting the frequencies as [52]:(5)where is an adjustable parameter chosen to be equal to 10 for the distance potentials and to 20 for the torsion potentials (based on preliminary tests), and where is equal to . This correction ensures that the potentials tend to zero when the number of observations in the data set is too small. A second trick that has been used consists, for a given bin , in summing the number of occurrences of the neighboring bins giving them a decreasing weight: (6)where is the number of occurrences in bin .
Predicting the stability curve of proteins from their sequence and structure alone is quite a difficult task. To slightly simplify the problem, we focused on families of homologous proteins, and make predictions that take into account some informations from the other family members. We therefore searched the full protein set for families of homologous proteins with at least three members of known . We found 11 such families containing both mesostable and thermostable proteins. They are: -amylase, acylphosphatase, lysozyme, myoglobin, -lactamase, -lactalbumin, adenylate kinase, cell 12A endoglucanase, cold shock protein, cytochrome P450 and ribonuclease. The complete list of the 45 proteins belonging to these families is given in Table S1 of Supporting Material.
Some quantities (such as the number of residues, , etc.) remain approximately constant inside a given family. This obviously makes the prediction method simpler to build. Such family-dependent analysis remains nevertheless quite intricate, since the thermostability properties of the proteins of a given family are sometimes very different.
In order to improve the performance of our method, the datasets , and have been further enlarged by adding proteins that belong to the protein family considered but whose was estimated from their environmental temperature instead of being experimentally determined; note that the pairwise sequence identity within each set was kept below 25% to avoid biasing the potentials (see [15] for details about the dataset construction procedure). Strictly speaking, this modification makes the datasets and the corresponding potentials family dependent.
The folding free energy of a given protein is computed at the temperatures , and from the (melting-)temperature dependent potentials defined in the previous subsections. More precisely, we have:(7)(8)(9)where for the distance potentials, for the torsion potentials, and the parameters are positive real numbers. The normalization coefficient is defined as:(10)
The temperatures (, , ) correspond to the average melting temperatures of the mesostable, thermostable and average datasets. The real -dependence of the folding free energies is obviously related to these melting temperatures. However, it would be a very strong (and obviously wrong) assumption to suppose that the average melting temperatures and the real temperatures are equal. Rather, as will be seen in the next subsection, a scale parameter must be introduced to relate the 's to the real .
The strategy for identifying the parameter values consists in maximizing the anticorrelation between the melting temperature and the difference in free energies . Indeed, has been shown to be much more correlated to the melting temperature than the folding free energy [15]. The optimization is performed on all proteins with known (listed in Table S1), excluding those of the protein family that we want to predict:(11)
The subscript indicates the family-dependent nature of the coefficients since their optimization is performed without the proteins of . This avoids the overestimation of the performance, and amounts to cross validation. All the optimizations described in this paper are performed using the ordinary least square regression method implemented in Mathemetica 7.0.
In the next steps of the computation, we estimate the full stability curve given by Eq.(2) from the three values of the folding free energies given by Eqs(7–9), for the set of 45 proteins from the 11 protein families. Let us assume for the moment that the -dependence is the true -dependence of the potentials. Under this assumption, the stability curve can easily be obtained: it is has the form (2) and depends on the thermodynamic quantities (, and ), viewed as parameters, which are identified to best fit the three data points:(12)
However, this simple approach does not give accurate predictions, both because the - and -dependences differ and because the error on these three points, which are moreover quite close along the -axis, leads to large errors on the whole curve. Three different issues must be solved to get reasonable stability curves.
The first issue concerns the sign of the second derivative of the curve. In a few cases (less than 10%), this sign is wrong, which implies that the curve is upside-down and the protein seems unfolded in the physiological temperature range. This error is related to the fact that the three points given in Eq.(12) are too close along the axis; this is due to the limited number of known proteins with very low or very high . The shape of the curve depends thus strongly on the relative position of the average point relative to the mesostable and thermostable points and . Sometimes even a small variation of these values can lead to the inversion of the shape of the curve.
To overcome this problem, we imposed a fourth point in the fitting procedure, in addition to those given in Eq.(12). This point is taken at a temperature of 0°K, where we impose to be equal to the average of the 's of the other proteins that belong to the same family. This quantity has no physical interpretation, as the inverse bell shape of the stability curve may not be extrapolated to zero temperature; indeed, we have in reality . This trick is however quite useful to impose the correct sign of the second derivative of the curve in the physiological temperature range.
This procedure has been applied when the predicted curve is upside-down, but also when the value of deviates by more than one standard deviation from the mean computed inside the family . This leads to an overall improvement of the results since it smooths out possible errors on the average point , which is amplified in the curve derivation procedure.
The second issue is the determination of the overall scaling factor of the curve. When more than one value of was experimentally determined within the considered family , we fix for the protein in the family as the ratio:(13)where is extracted from the predicted curves as the coefficient of the term, is the experimental value and the sum is over the proteins belonging to excluding ; this again amounts to obtain predictions in cross validation. If only one or no values were available for the family, we took as normalization factor the mean of the values found for the other families, excluding the largest and smallest values. This is a rough approximation since this quantity is expected to be strongly family dependent. However, despite the crude approximations made, the final result shows a fair performance that will certainly improve when more data or an independent determination will be available.
The last issue concerns the real temperature dependence of the potentials. Strictly speaking, the -dependence of the potentials is different from the real -dependence, even though they are obviously related. Indeed, the temperature resistant interactions can be expected to play a fundamental role in the stabilization in the high temperature regime and vice versa in the low temperature region (see [16]–[20] for the temperature dependence of the amino acid interactions). The assumption that we made is that the real value at which the potentials are calculated is related to the value of by a multiplicative factor that we call , which is assumed to be different for each protein. The strategy for fixing it is the following: once the function has been estimated for all the proteins of a given family , we determined the temperature at which it is zero. We identified for a protein so as to minimize the cost function:(14)
Since we are working in cross validation, the sum is over the proteins that belong to family . For a given protein , the folding free energy is thus given as .
The prediction of the mechanisms used by proteins to enhance their thermoresistance is a highly non-trivial issue. The principal mechanisms of this stabilization can be schematically described in terms of three strategies (see Figures 1a–c). The first consists in a global decrease of the folding free energy at all temperatures, which automatically implies an increase of the melting temperature. The second strategy consists of less negative values of , which broadens the stability curve. In the third strategy the temperature of maximal stability undergoes a shift towards the high temperature region. It is not simple to understand which mechanism is used by each protein and if it is used alone or in combination [31]. Moreover, different proteins of the same family can reach higher thermostability through completely different mechanisms.
In order to gain understanding into the thermal stability enhancement strategies and to obtain some quantitative predictions, we designed a method to predict the full stability curve of 45 proteins that belong to 11 homologous families (see Methods section). The results are the 45 stability curves given explicitly in Table S3 and plotted in Figure 3.
To make the analysis quantitative, we extracted from these predicted stability curves three independent thermodynamic parameters that define the transition, namely , and at 25 , and compared them with the experimental values. For the melting temperature, the experimental values are known for all 45 entries while for the other two quantities, they are known for 17 and 16 proteins, respectively (see Table S2). We report in Table 1 the standard deviation between the computed and the experimental values, as well as the correlation coefficient between the two quantities with the corresponding P-values.
Let us start with the analysis of the melting temperature whose values are simply extracted from the protein stability curves by looking for the zero of Eq. (2), since by definition:(15)
The value of the standard deviation between the experimental and the so computed 's is, in cross validation, equal to about 13 and reduces to 10 when the 10% worst predicted entries are excluded (Table 1). This value is comparable with the one found previously with a different method [15], with the notable difference that we predict here simultaneously the whole stability curve. In Figure 4.a, the predicted versus the experimental 's are plotted; the corresponding correlation coefficient is found to be equal to 0.69 (P-value ), and to increase to 0.76 upon exclusion of the 10% worst predicted proteins.
We also computed the for all the proteins belonging to the eleven homologous families. In this prediction, the identification of the normalization factor defined in Eq. (13) is fundamental. Unfortunately, we do not have enough input data, i.e. experimental 's, to identify this parameter inside each family: only for 17 entries is the known, with moreover often quite large experimental errors (of the order of 10–20%). When performing predictions in cross-validation, we have thus to fix independently of the other proteins of the family (using the procedure explained in Methods) for more than half of the entries, which inevitably gives rise the errors.
The standard deviation between the experimental and the predicted values of is reported in Table 1. It is equal to 1.3 ) and reduces to 0.8 when the two worst predicted proteins are excluded. The experimental and predicted values are plotted in Figure 4.b; the correlation between the two quantities is equal to 0.92 (P-value ), but falls down to 0.41 upon exclusion of the two worst predictions.
We chose as last independent quantity that can be extracted from the predicted curves the folding free energy at 25 (). The considerations made in the previous paragraph about the normalization factor are valid for this quantity too and thus we cannot expect a perfect correlation between the predicted and experimental values due to the lack of data. We found indeed a standard deviation of 4.1 between predicted and measured 's, which reduces to 2.6 when the two worst predicted proteins are excluded. The correlation coefficient between the experimental and the predicted values is 0.4 (P-value 0.05) and 0.7 upon exclusion of the two worst predictions. These results are shown in Table 1 and plotted in Figure 4.c. A list of values of , , and predicted from the 45 stability curves, as well as the corresponding experimental values where available, are reported in Table S2 of Supporting Material.
A further outcome that can be derived from the predicted stability curves is a better understanding of the strategies used within each protein family to reach a higher thermal stability. In particular, we can evaluate quantitatively the correlation between the thermodynamic and thermal stabilities: the linear anticorrelation between and (usually taken as the descriptor of the thermodynamic stability) is relatively high and is of the order of 0.7 when the two worst predicted families are excluded. The increase of the thermodynamic stability thus remains the principal mechanism for the thermal stability enhancement. The reason for this is that single amino acid substitutions can cause much easier an increase of the number of thermodynamically stabilizing interactions, such as hydrogen bonds and hydrophobic interactions, than for example a shift of the optimal stability temperature towards higher , for which more complex amino acid substitutions are in general necessary. This result, which has already been obtained on the basis of experimental data [31], [53], is here derived purely on the basis of our predictions.
The other two mechanisms for enhancing the thermostability, discussed in the previous sections, turn out to be important too even though they show a lower correlation with the melting temperature. In particular, the shift of the maximum stability temperature has a linear correlation coefficient of about 0.5 with and the change in heat capacity an anticorrelation coefficient of about 0.3, when excluding the two worst predicted families.
These predicted values can be compared with experimental data for the few proteins for which the full stability curve has been determined and thus similar correlation coefficients between and , and between and can be computed (see for example [53]). Notably, the experimental correlation coefficients and are equal to 0.6 and 0.2, respectively, and are thus quite close to the correlation coefficient predicted by our method. The shift of towards higher appears thus to be a preferred method for enhancing the thermostability compared to the change in . In other words, the reduction of the conformational entropy in the denaturated state or its increase in the native state seems easier to achieve compared to a change of .
The full understanding of protein thermal stability remains a challenge in protein science despite the large amount of research on this topic the last decades. As a matter of fact, it is globally more intricate to understand than the thermodynamic stability. Indeed, besides the problem due to the marginal stabilization achieved by a delicate balance of opposite forces, it poses the additional – and not the least – issue of the temperature dependence of the amino acid interactions, which is barely known.
We have designed a method based on (melting)temperature-dependent statistical potentials to deepen the thermal stability investigation. The basic idea behind this approach is simple and consists in constructing different datasets in which only proteins with given thermal properties were considered. Mean force potentials were extracted from sequence-structure frequencies computed from these datasets, following the standard statistical potential formalism, and hence reflect their thermal characteristics. They actually represent the amino acid interactions at some temperature that is related to the average of the proteins in the dataset. The folding free energy of a given protein at a given temperature was estimated on the basis of these -dependent potentials. More precisely, three different datasets with different average 's were constructed, from which three folding free energies at these 's were computed for each protein. The identification of the protein's full stability curve was accomplished by the identification of the modified Gibbs-Helmholtz equation (2) that best fits these three points.
Before concluding with future perspectives, let us summarize briefly the performance of the method and the main errors that affect it. The standard deviations between the experimental and computed quantities are equal, in cross-validation, to 13 , 1.3 ) and 4.0 for the melting temperature, the and the folding free energy at 25 , respectively. These results can be considered as rather good especially if one considers the three main sources of error that we have encountered. The first source is certainly the lack of data. As already stressed in the main text and in [15], we do not have enough experimentally resolved proteins with known to build larger datasets and thus more accurate potentials, even though we introduced some tricks to partly overcome this problem. This issue will certainly be improved when more experimental data will be available. The second source of error is related to the presence of ligands in some of the analyzed families, which contribute strongly to the protein stabilization but which we unfortunately cannot take into account with our statistical potentials. Finally, the measurement errors are sometimes quite significant, especially due to the fact that the experiments are not performed exactly in the same environmental conditions. These different issues taken together significantly increase the error on the predictions.
A noteworthy result that can be deduced from our predictive approach is that the preferred mechanism for enhancing the thermostability is an increase of the thermodynamic stability, in agreement with previous results based on experimental data [31]. Unfortunately, this does not allow us to construct an accurate predictor for the melting temperature on the basis of the thermodynamic stability only [15], since the other thermostabilizing mechanisms turned out to be important too – although to a lesser extent. Taking these other mechanisms into consideration as we did in this paper led us to a prediction method with much better performances, which we moreover hope to further improve in the near future. Furthermore, the analysis of the thermal stability optimization strategies has also shown that it is not possible to determine a unique molecular cause or a thermodynamic effect that explains the complexity of the thermal resistance modulation for the different families, since different strategies are used in combination.
We would like to underline the main strength of our approach that is the possibility to predict at once all the thermodynamic parameters that characterize the protein folding transition. We can indeed predict with our method both the thermodynamic and thermal stabilities in a large temperature range. As far as we know this is the only method that is able to do that, and moreover it does so in a fast and relatively accurate way. Neither the standard statistical potential formalism nor the molecular dynamics simulations or the coarse-grained computational approaches to protein folding are able to consider explicitly the temperature dependence of the amino acid interactions and give predictions for both kinds of stabilities.
However, some points of the present analysis can still be improved, and we plan to do so in a future investigation. In particular, we will try to supply to the lack of data by enlarging the dataset of proteins whose thermal properties have been measured experimentally and subdivide it in more than three subsets so as to be able to get more reliable fits of the stability curves.
Two different ways can be explored to enlarge the datasets. The first consists in adding proteins with known structure but unknown melting temperature. To decide to which of the thermal ensembles these additional proteins belong, one could estimate their from the method presented in this paper or from the environmental temperature of their host organism. The other strategy consists in the use of proteins with known melting temperature, whose structures are unknown but could be obtained by comparative modeling techniques. This approach is motivated by earlier analyses that tested modeled structures for the prediction of thermodynamic stability changes upon point mutations on the basis of standard statistical potentials [54]. Indeed, predictions applied on modeled structures have been shown to undergo a surprisingly small accuracy loss compared to experimental structures owing to the coarse-grained structural representation on which the potentials are based. This finding lets foresee an increase of the overall accuracy of our prediction method due to the enrichment of the datasets with modeled structures. But it also foreshadows the applicability of the resulting prediction method to low-resolution or modeled structures, with good performances. This undoubtedly increases the potentialities and interest of our approach.
We expect the enlargement of the datasets to play an important role in the reduction of the prediction errors, since it will allow us to define more than three datasets and thus to compute the folding free energies of a target protein at more than three different temperatures. This should definitely reduce the consequence of the errors on the predicted points in the -plane when fitting the stability curve through those points. Moreover, larger datasets will allow us to consider more types of statistical potentials (for example potentials that depend simultaneously on amino acid types, interresidue distances and backbone torsion angle domains [52]), which are now forbidden for statistical significance reasons.
Note finally that the current version of our prediction method is family-dependent, as the datasets vary slightly from one family to another and the optimization of some parameters is performed inside the families (see Methods section). We would like to stress that this procedure does in no way bias the predictions. All our tests are indeed performed in pure cross validation. Rather, this procedure improves the predictions by exploiting relevant information that characterizes the homologous families. Another promising improvement of our prediction method, which would make it applicable to any target protein of known structure, consists in extending the current version without too much accuracy loss to the more general case that ignores any reference to homologous proteins.
In conclusion, although there is still room for improvements and generalizations, our approach has opened a novel and original way for designing fast and accurate predictors of protein stability at different temperatures.
|
10.1371/journal.pgen.1003118 | A Histone Deacetylase Adjusts Transcription Kinetics at Coding Sequences during Candida albicans Morphogenesis | Despite their classical role as transcriptional repressors, several histone deacetylases, including the baker's yeast Set3/Hos2 complex (Set3C), facilitate gene expression. In the dimorphic human pathogen Candida albicans, the homologue of the Set3C inhibits the yeast-to-filament transition, but the precise molecular details of this function have remained elusive. Here, we use a combination of ChIP–Seq and RNA–Seq to show that the Set3C acts as a transcriptional co-factor of metabolic and morphogenesis-related genes in C. albicans. Binding of the Set3C correlates with gene expression during fungal morphogenesis; yet, surprisingly, deletion of SET3 leaves the steady-state expression level of most genes unchanged, both during exponential yeast-phase growth and during the yeast-filament transition. Fine temporal resolution of transcription in cells undergoing this transition revealed that the Set3C modulates transient expression changes of key morphogenesis-related genes. These include a transcription factor cluster comprising of NRG1, EFG1, BRG1, and TEC1, which form a regulatory circuit controlling hyphal differentiation. Set3C appears to restrict the factors by modulating their transcription kinetics, and the hyperfilamentous phenotype of SET3-deficient cells can be reverted by mutating the circuit factors. These results indicate that the chromatin status at coding regions represents a dynamic platform influencing transcription kinetics. Moreover, we suggest that transcription at the coding sequence can be transiently decoupled from potentially conflicting promoter information in dynamic environments.
| Many human pathogenic fungi are able to change their morphological properties, including their size and shape, in response to their outside environment. This ability, which is key for infection, is not completely understood on the molecular level. We have previously shown that not just DNA–binding transcription factors, but also chromatin-modifying enzymes that interact with DNA–binding proteins, are important regulators of morphogenesis in the model fungus C. albicans. In this work we dissect how such a chromatin-modifying enzyme regulates fungal morphogenesis. We surprisingly found that perturbation of chromatin has little influence on steady-state transcription, but modulates transient gene expression changes in differentiating C. albicans cells. Altered transcription kinetics affects a group of transcription factor genes that determine morphology. We thus identified a chromatin modifier that exerts kinetic control of transcription factor genes to control fungal morphogenesis. The results highlight the importance of chromatin to determine the kinetics of transcription changes rather than the steady-state transcript levels.
| Cells with identical genomes can adopt various phenotypes, which is a central feature during differentiation of multicellular organisms. During differentiation, the outside stimuli and the cellular machineries that process them are thought to determine the resulting cell types. Often, both the original and the resulting cell types are stable and display defined morphologies and gene expression programs. Understanding how such transitions between two stable cellular states are controlled and achieved on the molecular level is of central importance for understanding development and the relationship of organisms with their environment.
Unicellular species such as simple fungi can also undergo differentiation. For instance, pleiomorphic fungi, such as the opportunistic human pathogen Candida albicans display diverse morphologies ranging from unicellular yeast-like, to multicellular pseudohyphal, and hyphal structures [1], [2]. The ability to undergo reversible transitions between the distinct morphologies is a key virulence factor of C. albicans, shared by many other pathogenic fungi of even distantly related taxa such as Histoplasma and Cryptococcus [2], [3]. Consequently, blocking fungal morphogenesis represents a plausible antifungal therapeutic strategy [4].
Hyphal differentiation of C. albicans is responsive to environmental and host stimuli and is controlled by several signal transduction cascades and over thirty transcriptional regulators [5], [6], [7], [8]. Nevertheless, how the individual factors interact and how they integrate information from upstream signaling cascades is poorly understood. Recently, several studies have implicated chromatin and chromatin-modifying enzymes in the signal integration process. For example, the NuA4 histone acetyltransferase (HAT) and the Hda1 histone deacetylase (HDAC) mediate dynamic acetylation and deacetylation of histones at promoter regions of hypha-specific genes, and their proper function is required for the establishment of a normal filamentation expression program [9]. In yeast-phase cells, hyphal-specific genes are repressed by the transcription factor Nrg1 [10], [11]. During hyphal initiation, cyclic adenosine monophosphate (cAMP)/protein kinase A (PKA) signaling drives eviction of Nrg1 from its target promoters, where Hda1 is recruited by another transcription factor, Brg1 [10], [12]. Hda1 activity subsequently results in the eviction of NuA4 from the target promoters, which prevents Nrg1 rebinding [10]. In this model, promoter chromatin is perceived as a platform for temporally regulated transcription changes in morphogenesis. Consistent with this notion, several other chromatin modifier mutants, including the histone methyltransferase Set1 [13], the HAT Rtt109 [14], and the Set3 HDAC complex [15] display morphogenesis-related phenotypes.
The Set3 Complex (Set3C) was first identified as a repressor of sporulation in Saccharomyces cerevisiae, and sequence homology suggests that it is evolutionarily conserved from fungi to mammals [16]. The catalytic subunit Hos2 was the first identified non-canonical HDAC required for gene activity [17]. In S. cerevisiae, the Set3C occupies the coding sequence of highly transcribed genes and is required for full expression of the galactose-inducible gene cluster [17]. Deacetylation of nucleosomes within the coding regions is thought to reset chromatin to a permissive state facilitating repeated cycles of transcription [17], [18]. Homologues of the Set3C in other fungal species have been implicated in the regulation of morphogenesis and virulence [19], [20]. We have recently shown that the C. albicans Set3C acts as a repressor of hyphal differentiation and its function requires functional cAMP/PKA signaling [15].
In this study, we set out to identify the molecular mechanism through which the C. albicans Set3C controls morphogenesis. First, we used a combination of chromatin immunoprecipitation followed by sequencing (ChIP-Seq) and RNA sequencing (RNA-Seq) to define the genome-wide regulatory target genes of the Set3C in yeast and hyphal cells. We found that similar to its S. cerevisiae homologue, the CaSet3C exclusively decorates coding regions and is associated with high transcriptional activity. However, during transient hyphal-inducing conditions, the Set3C delays the establishment of the hyphal-specific gene program by modulating the transcript levels of four phase-specific transcription factors (BRG1, TEC1, NRG1, EFG1), suggesting that the Set3C can act both as a transcriptional activator and repressor. We also demonstrate that these four factors form a core transcriptional circuit underlying morphogenesis, whose output is restricted by the Set3C. The results provide comprehensive insights into the mechanisms whereby the chromatin layer of regulation superimposes on a core transcriptional factor circuit that controls cellular morphogenesis in C. albicans and possibly in other fungal pathogens.
In S. cerevisiae, the Set3C is composed of seven different subunits, which show sequence conservation in C. albicans and mammals [15], [16]. In S. cerevisiae, four of the subunits form a core complex (Set3, Hos2, Snt1, Sif2) and are required for structural integrity, while three additional subunits are peripheral (Hst1, Cpr1, Hos4) [16]. To enable biochemical investigation of the CaSet3C, we constructed a series of C. albicans strains carrying epitope-tagged alleles of the Set3 and Hos2 subunits. C. albicans is diploid, thus the second alleles in these strains were deleted. Since deletion of the Set3C causes hyperfilamentation [15], phenotypic analysis ensured that the tagged alleles were functional (Figure S1A). To probe the conservation of the complex architecture, we immunoprecipitated Set3 and Hos2 from whole cell extracts, and identified their interaction partners by mass spectrometry. Both subunits co-purified with the homologues of all S. cerevisiae core complex subunits (Set3, Hos2, Snt1, Sif2) (Figure S1B), indicating that the core complex is conserved in C. albicans (Figure 1A). In addition, we verified the Set3-Hos2 interaction by immunoprecipitation and immunoblotting (Figure 1B) and confirmed nuclear localization of the complex with a Hos2-GFP construct (Figure S1C).
To obtain a genome-wide binding profile of the Set3C and to determine its regulatory targets, we performed chromatin immunoprecipitation followed by sequencing (ChIP-Seq) of Set3 and Hos2 in exponentially growing yeast-phase cultures. We first identified binding peaks using Model-based Analysis of ChIP-Seq (MACS) [21], and found that 90% of all peak summits and around 85% of positions within peak regions fall within annotated coding regions (Figure S3A). Read density profiles averaged across all genes (meta-gene analyses) confirmed the strict localization of binding to coding regions (Figure 1C), and suggested that binding of Set3 and Hos2 to each gene across conditions could be assessed quantitatively by read-coverage of the coding regions (in reads-per-kilobase-per-million-reads [RPKM]) similar to RNA-Seq analyses; see Materials and Methods). Set3 and Hos2 enrichments showed a strong correlation genome-wide both for RNA Polymerase II (RNAPII)-transcribed genes (r = 0.80, Pearson's correlation) and RNAPIII-transcribed tRNA loci (r = 0.52, Pearson's correlation), arguing that both subunits co-localize on the chromosome (Figure 1D).
To test whether the Set3C indeed functions as a histone deacetylase in vivo, we performed ChIP experiments of acetylated histone H4. The ratio of acetylated H4 to total histone H3 was increased in a set3Δ/Δ strain when compared to wild type at the Set3C-bound positions (Figure 1E). Taken together, these results demonstrate that the C. albicans Set3C has histone deacetylase activity and localizes to coding regions of its target genes and tRNA loci.
To dissect how the Set3C regulates gene transcription, we performed RNA-Sequencing (RNA-Seq) of exponentially growing yeast-phase cultures. We found that C. albicans Set3C target genes were on average 8.7-fold more highly expressed than all genes (Figure 2A), and that binding correlated with RNA expression, as described for the S. cerevisiae homologue (Figure S4A) [17]. In particular, hexose catabolism genes and nucleosomal histone genes (both enriched among Set3C targets; P = 1.3×10−10, P = 8×10−8, respectively) are both highly occupied and expressed (Figure 2A). This implies that Set3C might be involved in enhancing gene expression of its target genes. However, surprisingly, when we performed RNA-Seq of a SET3-deletion mutant, the expression levels of only few target genes were affected, including the hexose catabolic genes (Figure 2B, Figure S4B). In contrast, histone genes and transcription factors, another functional category enriched among the targets (P = 2×10−3) remained unaffected. In wild type cells, acetylation level of histone H4 at the coding sequences of selected Set3C-targets correlated with RNA expression (r = 0.88, Pearson's correlation, Figure S4D). Moreover, increased acetylation of histone H4 was detectable at almost all tested loci in set3Δ/Δ cells (Figure S4C). These data argue that the occupancy of Set3C is linked to active transcription, but the presence of Set3C at transcribed gene bodies influences the steady-state transcript levels of only a small subset of targets.
The yeast-to-hypha transition in C. albicans, which is repressed by the Set3C [15], involves transcriptional changes that affect around 600 genes, corresponding to roughly 10% of the genome [11], [22]. To further dissect how the Set3C represses this transition, we indentified the binding targets of the Set3C in cells exposed to filament-inducing conditions (Figure 3A). In total, we detected 237 Set3C targets in hyphae, after a 30 minute induction. The ratio of the hypha/yeast ChIP-Seq values was used to classify a hypha-specific target gene set (127), a yeast-specific target set (85) or genes constitutively bound (110) (Table S5). We also performed RNA-Seq during filament formation and found that hypha-specific Set3C targets were on average induced whereas yeast-specific target genes were repressed upon filament-induction (Figure 3B). In fact, the differential RNA expression values and differential ChIP enrichment signals showed a strong correlation both for Set3 (r = 0.69, Pearson's correlation) and Hos2 (r = 0.8, Pearson's correlation) (Figure 3C, 3D; data not shown). However, when set3Δ/Δ cells were induced to form hyphae, their transcript induction profile was virtually identical to that of wild type cells (r = 0.9, Pearson's correlation, Figure 3E). These results clearly demonstrate that the Set3C is recruited to induced genes while it is depleted from repressed genes upon hyphal induction. Notably, set3Δ/Δ cells are still able to efficiently execute initiation of the hypha-specific transcriptional program.
Though the Set3C appears to be selectively recruited to a subset of transcribed genes, the differences in steady-state transcription and gene induction patterns in set3Δ/Δ cells are most surprisingly only minimal, and cannot explain why set3Δ/Δ cells are hyperfilamentous. To identify such potentially misregulated loci, we decided to identify which transcription factor(s) are responsible for Set3C recruitment, and collected the target lists of all transcription factors (TFs) whose genome-wide binding has been analyzed in C. albicans. These candidate TFs have been implicated in several cellular processes, including morphogenesis and biofilm formation [23], biofilm matrix regulation [24], carbohydrate metabolism [25], ribosome biogenesis [26], telomere control [26] and metabolic pathways [26]. Interestingly, the P-values of the overlaps showed an around 10 order of magnitude difference for TFs involved in morphogenesis (Efg1, Ndt80, Rob1, Brg1, Tec1) and carbohydrate metabolism (Gal4, Tye7) compared to other TFs (Figure 4). The fact that glycolytic genes were enriched in the GO-term analysis (see above) is in agreement with the finding that the Set3C target list shows a significant overlap with the targetome of Gal4 and Tye7, the major activators of the glycolytic gene cluster [25]. To our surprise, we also found that TFs whose target set showed a significant overlap with the Set3C set are themselves Set3C targets (blue boxes, Figure 4). Taken together, these results indicate that Set3C is a transcriptional co-factor of morphogenesis and glycolysis regulators. The fact that the regulators themselves are Set3C targets suggests that misregulation of rather the TF genes and not their targets could be the cause of the hyperfilamentous phenotype displayed by set3Δ/Δ cells.
While morphogenesis regulators are major Set3C targets, and their target sets show significant overlap with the Set3C targetome, steady-state transcript levels of TFs were unaltered in set3Δ/Δ yeast phase cells (Figure 2B). Thus, we hypothesized that if altered transcription of TFs is responsible for the hyperfilamentous phenotype of set3Δ/Δ cells, this effect could be transient. Consequently, we analyzed the transcript level changes of the Set3C-target TFs at a high kinetic resolution around the induction stimulus. We found that the transcript levels of EFG1, NRG1, TYE7, which were identified as hyphal-enriched Set3C targets, undergo a rapid decrease following hyphal induction. By contrast, a rapid 50–100-fold induction of transcripts of the hyphal-specific targets BRG1 and TEC1 were visible after only 10 minutes (Figure 5A). The TFs bound constitutively by the Set3C did not show significant transcript changes around the induction point. This pattern was qualitatively identical in differentiating set3Δ/Δ cells. However, a 1.5–2-fold quantitative difference was observed between wild type and set3Δ/Δ cells at several time points that mostly affected the phase-specific TFs (Figure 5A, 5B). For instance, BRG1 transcript level was about 1.5-fold higher in set3Δ/Δ cells at all time points following induction, and TEC1 transcript level was 2-fold higher in set3Δ/Δ cells at 20 and 30 minutes post induction. On the other hand, EFG1 and NRG1 transcript levels showed an approximate 1.5–2-fold decrease in set3Δ/Δ cells at 20 and 30 minutes, respectively (Figure 5A, 5B). This suggested that the hypha-specific factors reach higher transcript levels in set3Δ/Δ cells, while the yeast-specific factors are repressed more upon hyphal differentiation. This effect was not a result of more cells responding to serum in the set3Δ/Δ culture, as the non-target IHD1 gene showed a quantitatively identical induction pattern, and the number of filamenting cells was scored above 90% percent after a 60 minute induction in both genotypes (Figure 5B; additional supporting qRT-PCR data in Figure S5).
If the transient transcript level differences of the phase specific TFs cause the hyperfilamentous phenotype of set3Δ/Δ cells, then manipulation of the levels of the TFs should be epistatic to the lack of SET3. Indeed, we found that removal of one BRG1 allele almost completely reverted the hyperfilamentous phenotype of set3Δ/Δ cells under intermediate conditions, while deletion of one TEC1 allele did not (Figure 5C), which is probably explained by the facts that TEC1 expression shows a burst upon induction as opposed to BRG1 that stays stably high. These data demonstrate that the Set3C adjusts transient expression of phase-specific morphogenesis regulators during hyphal differentiation, and that genetic interference with the BRG1 regulator is sufficient to partially revert Set3C-deficiency.
Biofilm formation in C. albicans is controlled by a transcriptional circuit comprising of six core regulators, including BCR1, BRG1, NDT80, EFG1, ROB1 and TEC1 [23]. Strikingly, five out of the six factors (all except BCR1) are Set3C targets, and three of the factors (BRG1, EFG1 and TEC1) showed altered transcription kinetics in set3Δ/Δ cells (Figure 5A). We therefore tested if the four phase-specific regulators that display Set3C-dependent transcription kinetics (BRG1, EFG1, TEC1 and NRG1) also form a regulatory circuit. Remarkably, ChIP experiments revealed that Nrg1 bound its own promoter and the promoters of the other three TFs in yeast-phase cells (Figure 6A). Tec1, Brg1 and Efg1 also bound the promoter regions of all four factors in hyphae (Figure 6A). Together with a recent genome-wide binding map of Efg1 in yeast phase [27], these data allow for the reconstruction of a partial map of the transcriptional circuit that underlies hyphal differentiation in C. albicans (Figure 6B).
In this study, we set out to dissect how the conserved Set3C histone deacetylase complex regulates hyphal differentiation in the human fungal pathogen C. albicans. Our data not only provide mechanistic and evolutionary insights into how chromatin deacetylation at the coding sequences regulates gene expression, but also allow for the characterization of the transcriptional circuitries underlying fungal morphogenesis. Remarkably, our data show that the activity of the transcriptional layer of gene regulation is fine-tuned by a second layer requiring chromatin modification.
To date, over a hundred genes have been implicated in C. albicans morphogenesis, including numerous conserved signaling cascades and transcription factors [6], [7], [8], [28]. However, a comprehensive understanding of how the underlying genetic circuit is organized and how it integrates outside stimuli is still lacking. Our data has led to surprising insights into the architecture of the hyphal regulatory circuit, whose simplified model is shown in Figure 6B. In this model, a core circuit is formed by four TFs, two of which are enriched in the yeast phase (NRG1 and EFG1) and two are enriched in hyphae (TEC1 and BRG1). The four factors form an interwoven network, whereby all four factors regulate themselves, as well as the other three factors. For practical reasons, we consider here the yeast-specific phase of the circuit as the ground state. In yeast cells, Nrg1 represses hyphal-specific genes, and cAMP/PKA signaling-dependent removal of Nrg1 is required for the induction of hyphal genes [10], [11], [12]. The relief of Nrg1 repression enables expression of hyphal-specific genes such as the TFs Brg1 and Tec1, which once expressed, repress the promoters of the yeast-specific regulators and reinforce their own expression (Figure 6A). This “excited” state of the circuit is subsequently responsible for the establishment of a normal hyphal transcription program. Indeed, all four TF genes are strong binding targets of the Set3C in their respective phases. In set3Δ/Δ cells, the circuit remains intact, but shows a “hyper-excited” state shortly after hyphal induction, that is reflected in the transcript levels of all four regulators (Figure 5, Figure 6B; Table S5). (It is, however, possible that the early surplus of BRG1 is at least partially accountable for the transcriptional effect at the other three TF loci.) Further evidence for the “hyper-excited” state includes the elevated expression of verified Brg1 target transcripts such as UME6 and HGC1 (Figure S5). This model predicts that in set3Δ/Δ cells the circuit responds to weaker hypha-inducing stimuli than in wild type cells, which explains why set3Δ/Δ cells are hyperfilamentous under “intermediate” conditions (Figure S1, Figure 5C; [15]). The model receives further support by the finding that removal of one BRG1 allele partially reverts the phenotype of set3Δ/Δ cells (Figure 5C). Thus, though many additional TFs are implicated in morphogenesis, we believe this simple circuit architecture is consistent with several key features of the differentiation process.
Recently, a similar, small transcription factor network composed of six core TFs was identified as the master network controlling biofilm formation of C. albicans [23]. Biofilms are formed on solid surfaces and indwelling medical devices by a population of diverse cell morphologies (including yeast, pseudohyphal and hyphal forms) that are embedded in an extracellular matrix. Remarkably, five out of the six biofilm regulators are Set3C targets (TEC1, BRG1, EFG1, NDT80 and ROB1), and three factors are found in the hyphal regulatory circuit (TEC1, BRG1 and EFG1). Hence, if the Set3C restricts the excitation state of the hyphal regulatory network, and there exists at least a partial overlap between the two regulatory circuits, the Set3C may also affect the output of the biofilm regulatory circuit and thus modulate biofilm formation. Strikingly, set3Δ/Δ cells produce a strong “rubbery” biofilm with visibly stronger mechanical properties when compared to wild type cells (Figure S6, Nobile and Johnson, unpublished data). These results strongly suggest that a kinetic control of transcriptional circuit components by the Set3C modulates biofilm formation, and likely constitutes a conserved regulatory mechanism underlying morphogenetic processes in C. albicans.
In S. cerevisiae, the Set3C binds to gene bodies where it deacetylates various residues of histones H3 and H4 [17]. The occupancy of the complex correlates with expression, and in set3Δ/Δ cells, galactose metabolic genes are turned on with a slower kinetics when compared to wild type after galactose induction, which suggests that the complex is required for gene induction [17]. In addition, Hos2, the catalytic subunit of the SetC3 binds at a few tRNA loci where it is necessary for efficient integration of Ty1 retrotransposons [29]. The C. albicans Set3C also decorates coding regions and tRNA loci, and its presence correlates with transcriptional activity, supporting the notion of conservation of Set3C localization being linked to active transcription. However, the genome-wide association analysis does not reveal how the recruitment of the complex affects transcription itself. In S. cerevisiae, the genetic removal of Set3 has only a marginal effect on genome-wide transcription [30]. Recently, a systematic study of chromatin-modifier mutants in S. cerevisiae revealed that several enzymes modulate the induction kinetics of their target genes rather than steady-state transcript levels [31]. Similarly, we detected transcript level differences at only 19% of the Set3C binding targets in set3Δ/Δ C. albicans cells. These genes mostly include the glycolytic genes that were all downregulated in the set3Δ/Δ mutant, arguing for a positive role of the Set3C in transcription (Figure 2), but the exact role of the Set3C is possibly determined by the context it is recruited. Thus, such a static view of the steady-state transcript levels is insufficient to predict precise function. Indeed, we detected 50% more BRG1 mRNA in set3Δ/Δ cells than in wild type cells already after 10 minutes of hyphal induction, which implies that the Set3C can also exert repressive functions. This is in agreement with an early report where set3Δ/Δ diploid S. cerevisiae cells experienced premature activation of meiotic genes upon induction of sporulation [16], and indicates that both the context and timing of histone modifications contribute to the net transcriptional output [32]. Since in set3Δ/Δ cells, the TF-cluster genes were either “hyper-induced” or “hyper-repressed” during transient hyphal induction (Figure 5B), and several other Set3C targets showed altered transcription kinetics (Figure S5), we propose that the complex is also part of a conserved mechanism that creates a transient “transcriptional memory” at coding regions to buffer fast promoter changes. The mechanistic basis of opposing functions (repression and activation) could be that dynamic histone (de)acetylation affects nucleosome density at coding regions, while nucleosomes carry other modifications that directly affect transcription [33]. Indeed, we observed a slightly reduced nucleosome density at several target loci in set3Δ/Δ cells (data not shown).
Why would the chromatin at coding regions be used as such a relay platform? A detailed resolution of the transcriptomic response to several stresses in S. cerevisiae recently revealed that histone deacetylation at stress-responsive promoters by Rpd3 is required for normal induction and repression kinetics [34]. In our view, the utilization of coding region chromatin to adjust transcription kinetics is advantageous when several outside stimuli transmit potentially conflicting information to target promoters. In such cases, a transient decoupling of transcription from promoter input may provide sufficient time to shape the proper response to the environment or even prevent overshooting responses. In the cases of transcriptional circuits, the time needed to establish the new circuit output may be used to “decide” if the morphological conversion is in fact favorable under the given circumstances. Indeed, such a “test the waters” strategy was recently hypothesized to underlie white-opaque switching, another morphological switching process of C. albicans [35].
One of the most intriguing questions that arise based on these results is how the Set3C is selectively recruited to the coding sequences of its target genes. Localization is in fact so exclusive to gene bodies, that genomic segments encoding long 5′ untranslated regions (UTR) are completely devoid of ChIP-Seq signal (Figure S3C). In S. cerevisiae, recruitment of the Set3C has been linked to the recognition of H3K4me2 which decorates mainly coding regions [18], but most surprisingly, we found that genome-wide localization the C. albicans Set3C does not depend on dimethylation of H3K4 (Figure S7). Notably, despite the fact that Set3C occupancy correlates with expression of the corresponding locus, not all highly transcribed genes, for instance ribosomal protein genes, are Set3C targets (Figure 2). To rule out that this is caused by differences in RNA stability of the gene clusters, we examined RNAPII density [36], and found a similar correlation of Set3C enrichment with RNAPII density as with RNA level genome-wide (data not shown). This indicates that though Set3C recruitment is linked to active transcription by RNAPII, the recruitment signal stems from sequence-specific binding of transcription factors at the gene promoters. Our preliminary experiments indeed suggest that the promoter of a Set3C target gene is sufficient to direct the complex to exogenous gene bodies (data not shown). Although further experiments are necessary to dissect the mechanistic basis of the transmission of the recruitment information, we believe the transcriptional regulators either alone or more likely, in combination may direct posttranslational modification(s) of RNAPII, that contribute to Set3C recruitment once the polymerase reaches the coding region. Indeed, the plethora of modifications of the RNAPII C-terminal domain (CTD) has been postulated to constitute a “CTD code”, whereby specific CTD modifications orchestrate the binding of protein factors that affect RNA processing, RNAPII termination or histone modifications [37], [38], [39]. Thus, it is tantalizing to speculate that the Set3C and its homologues could serve not only as erasers of histone acetylation but also as readers of CTD modifications.
During its coevolution with the human host, C. albicans had to adapt to various niches representing a wide spectrum of physical, chemical and immunological parameters. One of its adaptive strategies appears to be controlling phenotypic transitions through small, evolvable transcriptional circuits that are responsive to outside stimuli of broad amplitudes. Our model proposes that the switch between distinct circuit phases proceeds through an intermediate stage characterized by chromatin changes not just at promoters, but also at gene bodies. We postulate that such chromatin-overlayed transcriptional regulatory circuits underlie the morphological diversity of C. albicans, and most likely many other pleiomorphic fungal pathogens.
We wish to point out that during the revision of this work, a report was published, demonstrating that the Set3C modulates transcription kinetics in response to carbon source shifts in S. cerevisiae [40]. This work shows that the yeast Set3C HDAC acts as an active repressor. Notably, indirect regulatory functions of Set3C can lead to positive regulation of target genes through a mechanism involving repression of overlapping non-coding RNAs [40].
C. albicans strains were routinely cultured in YPD (2% Bacto Peptone, 1% yeast extract, 2% Dextrose). For solid media 2% agar was added. Yeast phase cultures were propagated at 30°C. To obtain exponentially growing yeast cells, single colonies were grown in YPD at 30°C overnight, diluted the next day to an optical density at 600 nm (OD600) of 0.1, and incubated on a rotary shaker for exactly five hours after which cultures reached OD600 = 0.8±0.05. For hyphal induction the cultures were split, one aliquot was washed once with distilled water and resuspended in prewarmed (37°C) YPD+20% fetal calf serum (FCS) of an equal volume of the starting culture. Induced cells were shaken at 37°C for 30 minutes, which provides enough time to induce several hypha-specific transcripts and it is before the first nuclear division occurring around 60 minutes after induction (data not shown). Cultures were optionally snap-frozen in liquid N2 after washing steps and stored at −80°C.
The complete list of C. albicans strains, primers and plasmids used in this study are listed in Tables S1, 2 and S3, respectively. All strains were derived from SN152 [41]. The wild type (CAIF-100), SET3/set3Δ (DHCA401), set3Δ/Δ (DHCA402), HOS2/hos2Δ (DHCA405) and hos2Δ/Δ (DHCA406) strains were previously described [15], [42]. SET1, BRG1 and TEC1 alleles were deleted using the fusion PCR strategy with the C.d.ARG4 and NAT1 markers [15], [41]. NRG1 and SET3 alleles were deleted using the SAT1 flipping strategy [15], [43]. Epitope-tagging constructs were created using the fusion PCR strategy [44] with the “tag-marker” donor plasmids described in Table S3. The Efg1-myc (CJN1781), Brg1-myc (CJN1734) and tec1Δ/Δ (CJN2320) strains were previously described [23]. The construct for targeted replacement of NRG1 with GFP was also created by fusion PCR. Transformation was performed via electroporation as described [43]. Correct genomic integration was verified with PCR and immunoblotting (for epitope tags).
IP was performed essentially as described [18] with modifications. Exponentially growing cells were harvested by centrifugation and washed twice with distilled water. For Set3-HA and Hos2-3HA IPs (Figure 1B and Figure S1B) cells were resuspended in 300 µl stringent lysis buffer (phosphate-buffered saline (PBS) with 300 mM NaCl, 1% Triton X-100, 2 mM DTT and protease inhibitors). After addition of around 200 µl glass beads (425–600 µm, Sigma), cells were lysed at 5 m/s for 45 seconds 5 times on a FastPrep instrument (MP Biomedicals). Samples were chilled on ice for five minutes between cycles. Tubes were then centrifuged at 14000 g for 5 min at 4°C, the supernatant (around 300 µl) was transferred to a fresh tube and diluted to 1 ml with stringent lysis buffer. Whole cell extracts were incubated at 4°C overnight with agarose beads covalently bound to the 3F10 clone anti-HA antibody (Roche) with the amount of beads previously titrated to the amount of antigen in the respective amounts of extracts. The next day beads were washed six times with stringent lysis buffer and bound complexes were resolved by SDS-PAGE. For the Set3-3HA-histone co-IPs (Figure S7) the same protocol was followed, except that a mild lysis buffer (10 mM Tris-HCl pH 8.0, 150 mM NaCl, 0.1% Nonidet P-40 and protease inhibitors) was used throughout the whole procedure and beads were washed only four times after immunoprecipitation. The antibodies used in the Western blot detection included anti-HA (3F10, Roche), anti-myc (ab32, Abcam) anti-H3 (ab1791, Abcam), anti-H4 (ab10158, Abcam) and anti-tubulin (DM1A, Sigma).
RNA was isolated by the hot phenol method [44] after which RNA was further purified by the SV Total RNA isolation system (Promega) according to the manufacturer's instructions. For qPCR, 500 ng total RNA was reverse-transcribed using the AMV reverse transcription system (Promega), and the diluted first strand cDNA was directly used as template. For RNA-Seq, 5 µg total RNA was depleted of rRNAs using the RiboMinus kit (Invitrogen) and cDNA was synthesized and amplified with the SMARTer cDNA synthesis kit (Clonetech) according to the manufacturer's instructions. 5 µg cDNA was purified with phenol∶chloroform∶isoamylalcohol extraction (PCI), precipitated in 70% ethanol at −20°C overnight, washed once with 70% ethanol, dissolved in distilled water, and was used as a template for library preparation. RNA quality was controlled with the BioAnalyzer (Agilent) during the whole procedure.
ChIP was performed essentially as described [44] with several modifications. Cultures were crosslinked by the addition of formaldehyde at a final concentration of 1% for 15 minutes at room temperature. Crosslinking was quenched by the addition of 125 mM glycine for 5 minutes. Cells were washed twice with ice-cold Tris-buffered saline (TBS) and pellets were frozen in liquid N2. Pellets corresponding to 80 OD600 cells were resuspended in 480 µl ice-cold ChIP lysis buffer (50 mM HEPES/KOH pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% Na-deoxycholate, protease inhibitors). After addition of around 500 µl glass beads (425–600 µm, Sigma), cells were broken at 6 m/s for 60 seconds 8 times on a FastPrep instrument (MP Biomedicals). Samples were cooled on ice for five minutes between cycles. The bottom of the tubes was punctured with a 273/4G needle and lysates were collected by centrifugation at 1500 g for 1 min at 4°C into a fresh tube. The lysates were diluted to 2.4 ml with ChIP lysis buffer. Aliquots of 300 µl were sonicated 15 times with 30s ON/30s OFF cycles at high setting on a Bioruptor (Diagenode). After sonication the aliquots were combined and centrifuged once at 14000 g for 5 min at 4°C. Supernatants were collected and used as input chromatin lysates. Set3-9myc and Hos2-9myc ChIPs were performed using EZ-View anti-myc coupled agarose beads (Sigma). After overnight incubation at 4°C washing and DNA purification was performed exactly as described [44]. For histone ChIPs (on Figure 1E, Figure S4C, Figure S7F) anti-acetyl H4 (06-598, Millipore), anti-H3 (ab1791, Abcam) and anti-H3K4me2 (07-030, Millipore) were used. Brg1-myc and Efg1-myc ChIP was performed with a commercial anti-myc antibody exactly as described [23]. Nrg1-3HA ChIP was performed with an anti-HA antibody (ab9110, Abcam). Tec1 ChIP was performed with a custom anti-Tec1 antibody [23]. After overnight incubation at 4°C, ProteinG-coupled Dynabeads (Invitrogen) were added for 2 h at 4°C, and subsequent washing and DNA purification steps were performed exactly as described [44]. The sonication settings typically resulted in fragments sizes mostly around 200–300 bp, which was controlled by agarose gelelectrophoresis of purified input samples. All ChIP experiments were carried out at least with three biological replicate cultures.
Sequencing of fragmented cDNA was carried out on a HiSeq 2000 instrument (Illumina) at GATC Biotech AG (Konstanz, Germany). Three biological replicates of wild type and set3Δ/Δ cells in both yeast and hyphal phases were included. The resulting 51 base reads were mapped onto the Assembly 21 of the C. albicans genome containing only the coding sequences using TopHat with default parameters, and allowing only for uniquely mapping reads [45]. Fragment Per Kilobase in a Million mapped reads (FPKM) values were calculated with Cufflinks, including quartile normalization (removing top 25% of genes from the FPKM denominator) and bias correction [46]. The transcript coordinates were fixed as in the annotation of the coding sequence assembly. For transcriptome analyses on Figure 2B and Figure 3B and the right panel of Figure 2A, only genes with a mean coverage of at least five nucleotides per base were included. Mean coverage was calculated as an average of the three biological replicates for each genotype or phase. For the transcriptome analysis on Figure 3E the same mean coverage cutoff (>5 nt/base) was used, but only for the wild type yeast and hyphal RNA-Seq samples. The complete RNA-Seq dataset is found in Table S6.
ChIP libraries were sequenced on a GAIIx platform (Illumina). Three biological replicates of the Set3-9myc ChIP in wild type and set1Δ/Δ backgrounds in yeast phase cells were sequenced, with two biological replicates of Set3-9myc ChIP and Hos2-9myc ChIP in hyphal phase cells and two biological replicates of Hos2-9myc ChIP in yeast phase cells. ChIP material of six biological replicates of the untagged control strain in the yeast phase was pooled prior to library preparation. One input sample of each genotype in each morphological phase was also sequenced. Samples were multiplexed with custom adapters. Reads (31–32 base) were mapped onto the chromosomal Assembly 21 of the C. albicans genome using Bowtie, allowing only for uniquely mapping reads [47] (Figure S2). Peak calling was performed by MACS with mfold = 2 and effective genome size = 14.324.316 bp parameters [21]. With a peak calling method such as MACS, detected peak coordinates rely on arbitrary thresholds and can vary even between replicate samples, and peak positions cannot be adjusted. Therefore, we developed a Read Per Kilobase in a Million mapped reads (RPKM) pipeline for quantifying ChIP-Seq enrichment within gene bodies. RPKM measures the enrichment of sequence reads within fixed chromosomal coordinates and is generally used to quantify RNA species in RNA-Seq experiments when the precise transcript coordinates are known [48]. The RPKM values of all open reading frames (ORFs) were calculated using their default chromosomal coordinates in ORF Assembly 21. Overlapping genes were removed from further analyses. Since tRNA genes are relatively short (80–120 bp), 20 bases up- and downstream of their chromosomal coordinates were added to define tRNA-cluster positions, which were used in the RPKM calculation. If two adjacent tRNA genes were closer than 20 bases, they were included in the same cluster. Overall, RPKM(ChIP)/RPKM(input) values correlated with the MACS fold enrichment parameter in the same sample (Figure S3B). Hence, RPKM is a suitable quantitative readout of our peak signals. Enriched genes were initially defined as having an RPKM(ChIP)/RPKM(input) ≥2 with both RPKM values ≥10 (“targets” on Figure 1C). Since a few ORFs showed minor enrichment in the untagged control ChIP-Seq sample (Figure 1C and Figure S2), the RPKM ChIP values were subsequently normalized against the RPKM ChIP values of the untagged control instead of the respective input samples. Further normalization against the input samples for the comparison of genotypes was not necessary, because RPKM values of all input samples showed a high pairwise correlation (r>0.9, Pearson's correlation, not shown). We defined the Set3C target set as an ORF having at least 2-fold Set3 or Hos2 enrichment and at least 1.5-fold enrichment of the other (blue box, Figure 1D), or a tRNA having at least 1.5-fold enrichment of both (Figure 1D). For the genome-wide enrichment analyses on Figure 1D, Figure 3C, and Figure S7 only genes with an average RPKM≥5 in both genotypes or phases were included. On Figure 3C only two biological replicates (R2 and R3) of the Set3-9myc yeast phase ChIP-Seq samples were included. Mapped reads were extended with the length of the MACS d parameter (∼150 bp) prior to visualization. The complete ChIP-Seq dataset including chromosomal coordinates and RPKM values of all ORFs and tRNA clusters is found in Table S4.
Data have been deposited at the Gene Expression Omnibus (GEO) under accession number GSE38427.
|
10.1371/journal.pcbi.1003237 | Network Signatures of Survival in Glioblastoma Multiforme | To determine a molecular basis for prognostic differences in glioblastoma multiforme (GBM), we employed a combinatorial network analysis framework to exhaustively search for molecular patterns in protein-protein interaction (PPI) networks. We identified a dysregulated molecular signature distinguishing short-term (survival<225 days) from long-term (survival>635 days) survivors of GBM using whole genome expression data from The Cancer Genome Atlas (TCGA). A 50-gene subnetwork signature achieved 80% prediction accuracy when tested against an independent gene expression dataset. Functional annotations for the subnetwork signature included “protein kinase cascade,” “IκB kinase/NFκB cascade,” and “regulation of programmed cell death” – all of which were not significant in signatures of existing subtypes. Finally, we used label-free proteomics to examine how our subnetwork signature predicted protein level expression differences in an independent GBM cohort of 16 patients. We found that the genes discovered using network biology had a higher probability of dysregulated protein expression than either genes exhibiting individual differential expression or genes derived from known GBM subtypes. In particular, the long-term survivor subtype was characterized by increased protein expression of DNM1 and MAPK1 and decreased expression of HSPA9, PSMD3, and CANX. Overall, we demonstrate that the combinatorial analysis of gene expression data constrained by PPIs outlines an approach for the discovery of robust and translatable molecular signatures in GBM.
| Glioblastoma multiforme (GBM) is the most common and aggressive brain tumor in adults, and, while the median survival time for treated patients is approximately one year, subgroups of patients respond differently to the same treatments, with some patients showing little improvement and other patients living far longer than expected. These differences in treatment response indicate that the tumors may show molecular differences that we can harness to tailor cancer therapy. To this end, we sought to identify biomarkers of patient survival in GBM. To improve the applicability of our molecular markers to other patient groups, we constrained our markers using maps of protein-protein interactions, and we also employed a unique computational strategy that incorporates patient-to-patient molecular variability into the results. We identified a set of 50 genes comprising a subnetwork signature that successfully separated GBM patients by their survival times. Our approach to identifying this subnetwork signature also improved our ability to identify its protein products in an independent cohort of patients. In the ongoing search to improve cancer detection and treatment, our work represents a successful strategy for identifying reproducible biomarkers that can more efficiently lead to the discovery of druggable protein targets.
| Glioblastoma multiforme is the most common primary brain tumor in adults and, unfortunately, also the most fatal. While GBMs are categorized histologically, the nature of the disease leads to significant variability in both tumor classification and patient outcome. To more specifically define the disease and simultaneously reveal the etiology, an unbiased search for “molecular signatures” of GBM has been undertaken by several groups [1], [2], resulting in a variety of GBM markers which, unfortunately, have modest overlap. Given the large degree of molecular heterogeneity of GBMs, analysis of thousands of patient samples may be required to identify comprehensive gene sets by conventional statistical approaches [3]. However, suggestions that these myriad lists can be integrated via a systems-level analysis, e.g. using molecular networks to find consensus marker sets [4], may help to simplify the observed heterogeneity. In such an approach, an individual gene can affect the algorithmic contribution of a neighboring gene when they coexist in pathways or networks that act to integrate molecular heterogeneity.
While approaches measuring gene expression across a group can capture gene interaction effects, they often employ summary measures, e.g. averaging, that omit valuable information regarding inter- and intra- patient differences. In this work, we hypothesize that the considerable patient-to-patient variability of GBM can be simplified into molecular networks by identifying molecular “state functions” using the computational method, CRANE (for Combinatorially dysRegulAted subNEtworks) [5]. The use of molecular states – where the binary expression pattern of a gene set is considered as a whole – allows us to identify subsets of genes whose configuration (i.e. the expression pattern rather than expression level alone) distinguishes between the two phenotypes of interest. In this approach, we do not assign a single expression state to a phenotype, but, rather, we search for the set of all states matching a particular phenotype. These expression states are grounded in well-known sets of biological interaction data, as defined by curated protein-protein interaction (PPI) networks.
We applied CRANE to the gene expression data collected by The Cancer Genome Atlas [6] for patients with primary (de novo) GBM. We identified novel subnetwork signatures of survival, which we then tested against an independent gene expression dataset. We also hypothesized that mRNA dysregulation analyzed in the context of PPI subnetworks more efficiently translates to detectible dysregulation at the protein level. To test this, we examined protein expression of selected targets using label-free proteomics in a retrospectively selected set of GBM tumor samples. The workflow presented here is a prototype for identifying manageable subsets of genomic and proteomic targets to ultimately drive the design of cost-effective clinical assays for predicting patient survival – a much desired endpoint for clinicians and patients alike.
We began by using GBM patient information and microarray data from The Cancer Genome Atlas [6] (TCGA) as compiled by Verhaak et al. [7]. CRANE, an established method for mining molecular networks [5] (illustrated in Figure 1), successfully identified several subnetworks that were informative in separating short-term (STS) from long-term survivors (LTS) using TCGA mRNA data. The expression patterns for individual genes comprising the top ten states within subnetwork 1 are shown in Figure 1, illustrating how the varied configurations of an individual subnetwork drive the identification of specific subgroups of patients. As an example, note that subnetwork state 3 (LHLHLHLLLL) occurs in two short-term survivors, whereas subnetwork state 4 (LHLHLLLLLL) occurs in two long-term survivors; though state 3 and state 4 differ only in the switch of one gene from H to L, they predict opposite outcomes. Also, note that the top ten states using these 10 targets only capture 39% of the total patients, reflecting the significant heterogeneity at the patient level. The complete list of subnetwork signature genes can be found in Table S1.
To investigate the reproducibility of CRANE subnetworks in predicting survival, we tested the TCGA-discovered subnetworks' classification performance on an independent GBM dataset published by Lee et al. [8]. In this analysis, subnetwork discovery and training of the classifier was done on the TCGA data, and testing of the classifier was done on the Lee et al. data. In this test of the TCGA training set, the targets were fixed by the training data (Table S1), and classification accuracy on the Lee et al. data was incrementally calculated for each 10-gene subnetwork (see Methods). We achieved a maximum classification accuracy of 80% when using the top 5 subnetworks generated by CRANE from TCGA data (Figure 2, further details in Table S2); we henceforth refer to this 50 gene set as the subnetwork signature. With only 1 subnetwork, or 10 genes, the positive predictive value (PPV) of short-term survival is slightly better than random chance (57%) while the PPV for long-term survival is 74%. The PPV for short-term survival reaches 90% with 5 subnetworks while the maximum observed PPV for long-term survival was 85% with four subnetworks. The cumulative value of using multiple networks – each with a defined set of states – is illustrated in Figure 1. For example, state 1 in sub-network 1 (LLLLLLLLLL) is seen in 21% of patients, while the next 9 states cover only an additional 18% of patients (total of 39% for the top ten states). Thus, the heterogeneity of the patient population cannot be captured even with 10 binarized states from a single subnetwork; multiple subnetworks (each with multiple states) are needed to provide adequate patient coverage and clinically useful prediction accuracy.
Known molecular subclasses of GBM exhibit differences in survival [9], [10], and we examined whether our subnetwork signature was acting as a surrogate for known subtypes. A well-accepted basis for the molecular subtyping of GBMs was recently established by Verhaak et al. using an 840-gene signature [7]. Only four of our top 50 CRANE targets – phospholipase C (PLCG1), paxillin (PXN), transforming growth factor beta 3 (TGFB3), and topoisomerase (TOP1) – overlap with this list, strongly suggesting that our subnetwork signature is not classifying patients by these existing subtypes. Since the CRANE targets may be acting as proxies for the 840 genes, we also checked for an association between our predefined survival groups and molecular subtype using the molecular subtype calls made by Verhaak et al. for the 173 “core” TCGA samples (i.e. those samples most representative of a molecular subtype). When using the 50-gene subnetwork signature for classification, our LTS group consisted of 21% Classical, 35% Mesenchymal, 38% Proneural, and 5% Neural samples while our STS group consisted of 13% Classical, 37% Mesenchymal, 42% Proneural, and 8% Neural samples. Using a chi-square test of independence, we found that these molecular subtypes are not significantly associated (p-value>0.05) with membership in our survivor groups in the TCGA data.
To examine the extent to which our subnetworks were capturing true differences in survival, we investigated the concordance between the predictions of the network-based classifier and the survival times of the 166 patients in the Lee et al. dataset. As seen in Figure 3, significant differences in survival are apparent between patient groups predicted by the 50-gene subnetwork signature (p-value<1e-6, logrank test), indicating the expected performance of the CRANE classifiers within the test dataset. We then compared CRANE's performance against the four subtypes proposed by Verhaak et al. Though the Verhaak signatures were not designed to segregate patients by survival, the Proneural subtype has slightly longer survival than the other subtypes (Figure 3). By the logrank test, there is no significant difference among the four Verhaak subtype survival curves; the four subtypes track the survival curve of the CRANE long-term survivors while the curve for the CRANE short-term survivors is quite distinct.
Given that younger patients tend to have better prognosis [11], we also tested for differences in the age distributions of the two CRANE predicted groups of patients. The age distributions of patients classified by the 50-gene subnetwork signature were similar (Figure S3), and a logrank test indicated that there is insufficient evidence to conclude that the age distributions differ (p-value = 0.14). Overall, the above tests show that our CRANE gene expression subtypes are distinct from the Verhaak subtypes and represent novel, age-independent subtype classifications for GBM.
CRANE examines heterogeneity at the mRNA level to produce state-based classifiers, and we hypothesized that the identified subnetworks transduce this heterogeneity into protein-level differential expression. We tested this hypothesis by examining protein-level changes in an independent cohort of 16 patients from the Ohio Brain Tumor Study, 10 of which were STS and 6 were LTS based on the criteria outlined above. We employed a label-free proteomic approach using ultra-long chromatographic gradients, which permitted the accurate identification and quantification of 5019 peptides from 1491 proteins across the patient samples. Differential expression of proteomic targets was defined using a mixed model of peptides, and we report p-values for the differential expression of each protein. Using this model, 338 proteins were significantly up- or down-regulated at a p-value≤0.05 (Table S3). We did not make false-discovery rate corrections for these p-values as this is not an unbiased discovery experiment. Instead, we were interested in modeling how proteomic expression varied for pre-specified subsets of genes. Although proteomics has less dynamic range than gene expression analysis, the above method permitted the confident identification and quantification of over one-third of the CRANE subnetwork signature (17/50 targets). Of the 17 targets of interest that were identified and measured (see Table 1), five proteins were significantly down-regulated and two were significantly up-regulated in LTS. Interestingly, these 7 proteomic targets have modest classification potential at the level of individual gene expression, as illustrated by the irregularity of their gene expression patterns in the TCGA dataset (Figure S2).
To explore the prognostic potential of the proteomic targets, we used classification and regression trees (CART) to identify patterns of proteins that would robustly classify STS from LTS using the significant proteomic targets; for classification, we used the 7 significantly differentially expressed proteomic targets, as well as YWHAQ, which was of borderline significance. This yielded a simple 2-gene protein-level classifier, illustrated in Figure S4. Using only CANX and MAPK1, the classifier is able to correctly identify 100% of long-term survivors in the group and 90% of short-term survivors. For example, when CANX has a normalized value greater than −1.05 and MAPK1 is less than 0.50, we can identify 9 of our short-term survivors, though such high sensitivity and specificity are likely indicative of over-fitting.
To explore our hypothesis that the use of network topology improves our ability to detect targets at the protein level, we compared the performance of CRANE-identified targets versus that of individual gene markers in identifying dysregulated proteins. The “individual gene markers” refers to a set of the most differentially expressed genes selected without respect to any underlying interaction structure. Specifically, we identified all genes with a fold-change ≥2 between the 86 LTS and STS survivors in the TCGA data and then ranked these genes according to their absolute t-statistic (i.e. the difference in group means divided by the pooled standard deviation). Of the top 200 individual gene markers, only one – ACTG1 – overlapped with the 50-gene subnetwork signature. Thus, 49/50 genes identified using a network-based classifier could not be discovered based on conventional analysis of individual gene markers.
As seen in Figure 4A, the use of an interaction network in an mRNA-based classifier markedly improves our ability to identify targets differentially expressed at the protein level compared to examination of individually dysregulated genes. Specifically, CRANE identified dysregulated subnetworks that were better represented in the proteomic data, and these subnetworks included more differentially expressed proteins when compared to dysregulated individual gene markers. When interrogating the proteomics data for the top 200 network-based genes (i.e. the top 20 CRANE subnetworks), over 50 proteins were identified (25%) and 18 of these subnetwork proteins showed differential expression (36% differentially expressed among those identified). In contrast, when using the top 200 differentially expressed individual genes, 21 were identified via proteomics (10%) and only 3 showed significant changes (14% differentially expressed among those identified). Fitting a linear regression model to the data, we find that individual gene markers yield differentially expressed proteins at a rate of 1.5% (relative to the number of genes used), whereas the network-based approach has a rate of return of 9.8% - a 6.5-fold improvement in the yield of our proteomics validation experiment.
We also explored the proteomic yield of the four Verhaak et al. subtypes. As shown in Figure 4B, the 210-gene Neural subtype had the best yield in the proteomics experiment, with 41 targets identified via proteomics (20% of all targets identified) and 20 showing significant changes (49% differentially expressed among those identified). However, the number of proteomic targets identified by the Proneural, Classical, and Mesenchymal subtypes was considerably lower. While the rate of return for these three subtypes (ranging from 1%–2.7%) was comparable to that of the individual gene markers, the rate of return for the Neural subtype was 9.9%.
In this work, we analyzed the mRNA-level heterogeneity of GBMs using protein interaction networks, arriving at a succinct list of 50 genes that predicts patient survival at 80% accuracy. Not only does the unique subnetwork signature show reproducible prediction of patient survival at the mRNA level, it also exhibits protein-level dysregulation that segregates short-term from long-term survivors of glioblastoma – a valuable characteristic in light of recent evidence suggesting that many mRNA-level signatures have questionable classification power and modest biological significance [12]. Additionally, the 50-gene subnetwork signature indentified here represents an experimentally tractable number of targets – measurable in a streamlined proteomics experiment – while previously discovered target lists are not likely to be translated into clinical assays due to their large size [7]. While past work on unsupervised classification of high-grade gliomas was complicated by the use of mixed WHO grade III and grade IV patient samples [1], [13]–[16], we herein develop a molecular signature based solely on primary, untreated grade IV tumors from the TCGA database. We note the caveat that the number of subnetworks included in the signature was selected based on the classification performance on the test (Lee et al.) data and, thus, requires further validation to be useful as a standalone classifier of gene expression data. In this work, we choose, instead, to explore how this 50-gene subnetwork signature behaves at the protein level.
Building upon the success of gene pair classifiers [17], the network analysis framework presented here identifies multigene subnetworks based on mRNA state functions – series of 1's and 0's – allowing us to account for patient-level heterogeneity in expression profiles. While binarization of continuous expression data certainly involves a loss of information, this concept lends itself to the design of therapeutic interventions, where targeted molecular therapies inhibit or activate key “switches” in the circuits of distinct patient subtypes. For instance, upregulation of insulin-like growth factor receptor (IGF1R), seen in subnetwork 3, has been identified in a wide variety of human cancers [18], and in vitro evidence suggests that this upregulation contributes to resistance against EGFR inhibitors [19]. Our results suggest that IGF1R has variable expression – on, or 1, in some tumors and off, or 0, in others – in patients within the same GBM survival class, indicating that experimental IGF1R monotherapies [20], [21], while inappropriate as a population-level intervention, may be highly effective in precisely selected individuals. A binary model of expression-activation is an oversimplification in some instances, however, where protein activity does not necessarily correlate with expression levels, e.g. in the case of kinases.
In contrast to proteomic approaches, several groups have worked on classifying the genomic alterations underlying GBM [22], [23]. Of the 309 unique, validated mutations identified through sequencing of the TCGA GBM tumor samples, CTNNB1, EP300, STAT3, and TOP1 also appear in the 50-gene subnetwork signature. These genomic alterations are likely to play causative roles in establishing the global state function of the subnetwork signature. β-catenin (CTNNB1), for instance, complexes with N-cadherin to coordinate tumor invasiveness [24] and shows some promise as a prognostic marker [25]. Additionally, TOP1 is targeted by topoisomerase inhibitors to treat a wide variety of cancers [26], [27]. CRANE identifies these key genes not simply because they show consistent expression across a group, but, rather, because their expression levels form a distinct pattern when viewed in conjunction with the 46 other genes in the milieu. This is in line with the known patient-to-patient variability in the mutational landscape of cancer [28]. In this light, the presence or absence of common mutations in patient subgroups differentially disrupts network state functions, and a single chemotherapeutic agent is unlikely to be effective in every patient.
We hypothesized that the underlying network structure would ultimately lead to differences in protein expression between survival groups. Using a mixed model accounting for inter-peptide dependencies within a protein, we identified 7 dysregulated proteins out of a total of 17 detected in the proteomics experiment from the 50-gene subnetwork signature. Though the stochastic nature of proteomics workflows may have discouraged their use as validation platforms, we demonstrate that ultra-long chromatographic gradients coupled with high-resolution mass spectrometers allow us to probe the signaling networks of interest in a high-throughput fashion, with chromatographic reproducibility (Figure S1) sufficient for the development of targeted assays (i.e. using pre-specified lists of M/Z values to measure daughter peptides of network targets).
To gauge how the interaction network influenced our success in identifying dysregulated protein targets, we compared the proteomic performance of CRANE against that of a signature based on differentially expressed individual genes. We found a marked improvement in our ability to detect protein-level changes in identified markers when a network-guided combinatorial algorithm is used to detect mRNA-level dysregulation signatures (see Figure 4), and the improved representation of subnetwork targets in the proteomic data can be attributed, in part, to the use of the PPI network. Sources of experimental bias in the measurement of protein expression can be similar to those in the identification of PPIs (i.e. more abundant proteins are more easily identified). However, when we consider the fraction of differentially expressed proteins among all proteins identified, the top 200 CRANE targets always deliver more than 30% precision in identifying differentially expressed proteins, reaching a maximum of 43% when 150 targets are evaluated. In contrast, when we consider the products of the top 200 individual gene markers (i.e. those having significant mRNA differential expression), the fraction of differentially expressed proteins reaches a maximum of only 14%. Assuming the trend in discovery is linear, the network-based approach affords a nearly 7-fold improvement in the rate of discovery of differentially expressed proteins. As a testament to the combinatorial aspect of our analysis, our seven differentially expressed proteomic targets (in Table 1) would not have been discovered if we had based our classifier on individually differentially expressed genes, for these proteins did not exhibit consistent mRNA expression across survival groups in the TCGA data (Figure S2). While it is well known that dysregulation at the level of individual gene expression does not necessarily correlate with protein expression (the mRNA-to-protein correlation is 0.43 for humans [29]), our observations clearly suggest that combinatorial, network-based mRNA-signatures serve as better indicators of post-transcriptional dysregulation when compared to sets of differentially expressed single genes. This result speaks to the ability of network-based algorithms to reproducibly detect dysregulated proteins at the population level, as opposed to uncovering the relationship between mRNA expression and protein expression within a single sample. As an alternative explanation, the network-based targets may point to proteins that are more abundantly expressed and for which dysregulation can be more efficiently measured.
Given that the Verhaak et al. subtypes were constructed through hierarchical clustering of gene expression data, we expected that their yield in a proteomics experiment would largely compare to the performance of individual gene markers (which were constructed based on ranked differential expression). While this was the case for Proneural, Classical, and Mesenchymal subtypes, the Neural subtype performed relatively well in predicting differentially expressed proteins, yielding proteomic targets at a rate comparable to the CRANE signatures. This suggests that the Neural subtype contains hidden network structure that boosts the visibility of the group at the protein level and/or that both the CRANE signature and the Neural subtype contain classes of proteins (e.g. structural and metabolic proteins) that are more amenable to proteomic measurement. In support of the latter hypothesis, the top gene ontology (GO) term in the Neural subtype was nucleotide metabolic process (GO:0009117, p-value = 4.72e-5) [7], and metabolic enzymes are typically well-represented in proteomic experiments [30].
We also examined gene ontology (GO) term enrichment of our CRANE signature using DAVID [31], and we compared the results to the enrichment of the Verhaak et al. subtypes. Of the CRANE GO terms significant at the 0.01 level, only 6 overlapped and were significant (p-value≤0.01) in the Verhaak et al. dataset, including terms such as “regulation of transcription,” “regulation of cell proliferation,” and “cytoskeletal organization” (see Table S4 for the complete list of significant overlapping terms). The most significant and informative GO terms found in the CRANE signature included items such as “protein kinase cascade” (GO:0007243, p-value = 3.98e-8), “I-kappaB kinase/NF-kappaB cascade” (GO:0007249, p-value = 6.56e-5), and “regulation of programmed cell death” (GO:0043067, p-value = 8.08e-5), all of which were absent or not significant in the Verhaak et al. subtypes (see Table S5 for the complete list of terms significant in the CRANE signature). These results indicate that the CRANE subnetwork signature emphasizes kinase cascades and the NF-κB pathway. NF-κB expression has been shown to be positively correlated with astrocytoma grade and inversely correlated with patient survival [32]. Importantly, deletions of NF-κB inhibitor α (NFKBIA) and amplifications of EGFR have been shown to be mutually exclusive events in GBM [33], suggestive of underlying genomic subtypes. Our work recapitulates the importance of understanding patient-to-patient variability in NFKB signaling to better direct therapeutic decisions.
Seven subnetwork targets were validated using proteomics, and these proteins have interesting connections to both glioma and cancer. For example, HSPA9 is not only upregulated in a variety of cancers [34], [35], but its expression also correlates with glioma grade and the proliferative potential of cells [36]. In our data, HSPA9 is strongly (fold change = 0.80) and significantly (p-value = 1.34e-5) downregulated in the tumors of long-term survivors, suggesting that, even between tumors of the same grade, HSPA9 biology may differentially affect patient survival. Similarly, we found that calnexin (CANX) has 0.74-fold diminished protein expression in long-term survivors, and this result is in line with the observation that CANX expression is significantly correlated with the transition from angiogenesis-independent to angiogenesis-dependent (i.e. more invasive) tumor growth in xenografts [30]. In turn, PSMD3, a subunit of the 26S proteasome, was also found to be downregulated in the tumors of long-term survivors, which is in line with the promising results of proteasome inhibitors in pre-clinical studies [37], [38]. More recently, a novel role for PSMD3 was proposed by Okada et al., who identified a SNP near the gene associated with the regulation of neutrophil count by both GWAS and eQTL analysis [39]. It has long been recognized that cancer and inflammation are synergistic processes [40], and it appears that increased neutrophil activity is associated with highly infiltrative gliomas [41], [42]. Given the potential role of PSMD3 in neutrophil recruitment in GBMs, our data are consistent with a hypothesis that downregulation of PSMD3 leads to less neutrophil-mediated inflammation and longer survival.
In assessing patient outcomes of GBM, we argue that the most informative prediction is whether or not a patient has a poor prognosis, i.e. is a “short-term survivor,” as this prognosis identifies patients who are poor candidates for the standard of care and for whom more aggressive therapies may be beneficial. To demonstrate the therapeutic potential of proteomic targets, we used CART to identify a decision tree useful in classifying our proteomic cohort. We found that two proteins could effectively classify our cohort of 16 patients with near perfect sensitivity and specificity, though this result may be due to overfitting in our cohort. Nonetheless, this result illustrates how gene expression targets may be translated into clinical proteomics biomarkers.
We note that the many of the GBM patients with a poor prognosis in our proteomic validation cohort did not receive the full standard of care: surgery, radiation, and chemotherapy. Consequently, survival classification in our study is not a proxy for response to the standard of care. In future clinical work, efforts should be directed to identifying cancer survivors matched on treatment protocols to allow for the identification of molecular features that render them susceptible to various therapies. While our 50-gene network signature is currently useful for prognostication, analysis of a treatment-matched cohort would potentially allow for the identification of targets to guide therapeutic decision making.
The results published here are in part based upon data generated by The Cancer Genome Atlas (TCGA) pilot project established by the NCI and NHGRI. Information about TCGA and the investigators and institutions who constitute the TCGA research network can be found at http://cancergenome.nih.gov/. Patient data was obtained from TCGA, where clinical data and corresponding microarray data were available for 200 glioblastoma patients [6]. Samples run on three different array platforms – the Affymetrix U133A GeneChip, the Affymetrix Human Exon GeneChip, and a custom-made Agilent array – were pooled into a composite dataset by Verhaak et al. [7], and these data were used for further analysis. To select only de novo GBM, we removed those patients with a pretreatment history, a histologic classification of “treated primary GBM”, or a prior history of glioma. We also excluded patients whose final vital status (living vs dead) was unknown. The remaining patients were separated into two groups based on survival, taking the top 25% (43 patients, surviving>635 days, ages 11–83) as long-term survivors and the bottom 25% (43 patients, surviving<225 days, ages 39–85) as short-term survivors.
CRANE [5] was employed to discover subnetworks of proteins coordinately dysregulated at the level of mRNA; the MATLAB code is available. The global human protein-protein interaction network was compiled from publicly available interactions in the Human Protein Reference Database [43], and the CRANE search algorithm was constrained to subnetworks of consisting of at most proteins. We binarized gene expression data by setting the genes in the top quartile of expression intensity to H (high expression) and all others (bottom 75%) to L (low expression). This threshold for high expression (25%) was previously shown to be most effective in identifying discriminative subnetworks using a range of datasets [5]. After binarizing the data, we were interested in identifying subnetworks whose “state” – the binary sequence of H's and L's – was informative in regards to the phenotype (STS vs LTS). This is formulated as an optimization problem, where the objective function to be maximized is the mutual information between phenotype and expression state, the J-value. Mutual information is a measure of the reduction in our uncertainty of a patient's phenotype, given observations of the subnetwork's expression state. More precisely, denoting the phenotype random variable with and letting denote the k-dimensional binary random variable representing the expression state of a subnetwork of size k, the mutual information between the expression state and phenotype is defined as . Here, denotes the entropy of the phenotype random variable, and denotes the entropy of the phenotype given the expression state of the subnetwork, . The entropy of a random variable X is defined as , where A denotes the set of all possible values of X and px denotes .
We refer to a particular expression state of a particular subnetwork as a “state function.” For a state function, the J-value is defined as the amount of information provided by that particular state on the phenotype, i.e. its contribution to the mutual information between phenotype and the state of the corresponding subnetwork. Namely, for a given state function f for a subnetwork composed of k proteins (i.e., f is an observation of random variable ), the J-value is defined as . Here, denotes , and denotes . It can be shown that .
In this analysis, we first identified high-scoring subnetworks according to their J-values and then sorted these high-scoring subnetworks according to their mutual information for survival. Additional parameters used to assess a network's prediction accuracy are the support (the fraction of samples containing a particular subnetwork state, ); the confidence (the fraction of long-term survivors possessing a particular subnetwork state, ); and the anti-confidence (the fraction of short-term survivors possessing a particular subnetwork state, ). A subnetwork and an associated state function have a high J-value if the state function provides high support, high confidence, and low anti-confidence (or, symmetrically, high anti-confidence and low confidence).
To test the network features discovered using TCGA, we explored their prediction accuracy using an independent GBM microarray dataset, GSE13041, available via the Gene Expression Omnibus [8]. After removing patients known to have received prior radiotherapy, chemotherapy, and/or temozolomide treatment, a total of 166 patients remained; using the survival time cut-offs as before, the short-term survivor group consisted of 41 patients (ages 34–86), and the long-term survivor group consisted of 50 patients (ages 22–78).
A neural network (NN) was trained on the TCGA data using the top k subnetworks (ranked by mutual information, where k is a variable), and test performance was gauged using classification accuracy, calculated as , where S is the number of correctly predicted short-term survivors, L is the number of correctly predicted long-term survivors, and T is the total number of test samples in the test dataset. We calculated the cumulative classification accuracy for k ranging from 1 to 10, i.e. examining accuracy of the best performing network alone, and then examining the performance of the best two networks, and then the best three networks, etc. Overall classification accuracy reached a maximum of 80% when using k = 5 subnetworks, each composed of size d = 10 genes (Table S2).
For comparison, we assessed how the four GBM subtypes proposed by Verhaak et al. stratified patient survival in the testing dataset, GSE13041. We first removed pretreated patients from the testing dataset, and the data was then log transformed, median centered, and normalized by each array's standard deviation; gene expression was inferred by averaging probe-level expression. For the 840 genes in the Verhaak et al. GBM subtype classifier, we calculated the Spearman correlation coefficient between the centroid expression profiles (derived from the TCGA dataset) and each sample in the testing dataset, assigning each sample to the subtype with maximum correlation.
To identify statistically significant proteomic changes, missing values were imputed using the median intensity per peptide within each survival group, and the data was standard normalized for each peptide. We used a mixed model to compare the group-wise protein intensity differences of interest, with the survival group set as a fixed effect and the peptide set as a random effect, which allowed us to account for the within-protein correlation of the peptides inherent in mass spectrometry-based proteomic experiments [44]. In the results, we only compare differences between various prespecified protein sets observed in the data, namely the proteins coded by the following genes: the genes in the top-ranking subnetworks identified by CRANE (200 genes in total), the genes in the Verhaak molecular subtypes (840 genes in total), and the top 200 genes with the most significant individual differential expression. Using a likelihood ratio test, a p-value≤0.05 for the proteins of interest was considered significant and no correction for multiple hypothesis testing was performed. These statistical analyses were performed using R 2.13.2 and SAS version 9.2 (SAS Institute Inc., Cary, NC).
|
10.1371/journal.pgen.1000531 | Forging Links between Human Mental Retardation–Associated CNVs and Mouse Gene Knockout Models | Rare copy number variants (CNVs) are frequently associated with common neurological disorders such as mental retardation (MR; learning disability), autism, and schizophrenia. CNV screening in clinical practice is limited because pathological CNVs cannot be distinguished routinely from benign CNVs, and because genes underlying patients' phenotypes remain largely unknown. Here, we present a novel, statistically robust approach that forges links between 148 MR–associated CNVs and phenotypes from ∼5,000 mouse gene knockout experiments. These CNVs were found to be significantly enriched in two classes of genes, those whose mouse orthologues, when disrupted, result in either abnormal axon or dopaminergic neuron morphologies. Additional enrichments highlighted correspondences between relevant mouse phenotypes and secondary presentations such as brain abnormality, cleft palate, and seizures. The strength of these phenotype enrichments (>100% increases) greatly exceeded molecular annotations (<30% increases) and allowed the identification of 78 genes that may contribute to MR and associated phenotypes. This study is the first to demonstrate how the power of mouse knockout data can be systematically exploited to better understand genetically heterogeneous neurological disorders.
| Mental retardation (MR; also known as learning disability) affects 1%–3% of people and is often associated with the presence of genomic copy number variations (CNVs) such as deletions and duplications. Most of these CNVs are rare and they often involve tens, sometimes hundreds, of genes. Pinpointing exactly which particular gene or genes are responsible for MR in an individual patient is therefore challenging and limits diagnostic applications. In this study, the functions of genes present within a large collection of MR–associated CNVs were investigated by comparing them to data from large-scale mouse knock-out experiments. We found that MR–associated CNVs contain greater than expected numbers of genes that give specific nervous system phenotypes when disrupted in the mouse. Not only does this study confirm that CNVs frequently cause MR, but it narrows down the list of genes whose changes lead to this disorder from thousands to several dozen. This reduced list of genes brings wide-spread genetic testing for MR one step closer. It also provides a better understanding of the biology behind MR that could, eventually, yield medical treatments.
| Mental retardation (MR) is defined as an overall intelligence quotient lower than 70, and is associated with functional deficits in adaptive behaviour, such as daily-living skills, social skills and communication. This disorder affects 1%–3% of the population and results from extraordinarily heterogeneous environmental and genetic causes [1]. Genetic changes underlying MR are still poorly resolved, especially for the autosomes that provide the largest contribution to disease aetiology [2]. Microscopically visible chromosomal rearrangements detected by routine chromosome analysis are the cause for MR in ∼5%–10% of patients [3]. Such rearrangements represent gains or losses of more than 5–10 Mb of DNA and affect many genes thereby almost inevitably leading to developmental abnormalities during embryogenesis. The most common effect of these variants is cognitive impairment, but they can also be frequently associated with other abnormalities such as heart defects, seizures and dysmorphic features [4].
Many recent genomic microarray studies have indicated that smaller, submicroscopic rearrangements, such as copy number variations (CNVs), frequently underlie MR (Table S1). However, CNVs, defined as DNA deletions or duplications greater than 1 Kb [5], are also widespread in the general population which considerably hinders the clinical interpretation of patients' CNVs [6]. Until now, most clinical CNV studies have focused on the identification of rare de novo CNVs [7]–[9], as the rate of de novo large (>50 kb) CNVs in the general population is comparatively low [10],[11]. Nevertheless, discriminating between benign and pathogenic CNVs solely on the basis of size and lack of inheritance is crude and provides no insights into how CNVs exert their phenotypic effects.
Fortunately, the genomics era has amassed a wealth of data that have long promised to associate the disruption of a particular molecular function or cellular pathway with clinical observations; in short, to forge links between genotype and disease phenotype. These genomic data include behavioural, physiological and anatomical examinations following the disruption of more than 5000 individual mouse genes [12]–[14]. These mouse phenotypic measurements more closely resemble observations from human clinical examination than any other systematic genome-wide data source. They might be especially relevant to human gene deletion variants, which represent a large majority among the rare disease-associated CNVs considered here (Table 1 and Table S2). Available genomic data also include functional annotations such as from the Gene Ontology resource [15], tissue expression levels [16] and carefully curated pathway data such as the Kyoto Encyclopedia of Genes and Genomes (KEGG) [17].
Our approach was to test the null hypothesis that genes present in MR–associated CNVs randomly sample all human genes. In particular, are they a random sample of genes (i) that, when disrupted in mice, result in particular phenotypes, or (ii) that are predominantly expressed in the human brain, or (iii) that participate in specific human disease pathways? To ensure that we correctly account for the application of multiple tests, we have controlled the false discovery rate (FDR) [18] such that there is only a small 5% likelihood that any annotation term has been identified as over-represented in our tests simply by chance. Only if any particular set of genes present within MR–associated CNVs form a significantly (FDR<5%) non-random sample can we be truly justified in predicting single genes, among the dozens commonly overlapped by such CNVs, as contributing to MR disease aetiology. In this study, we show both significant and substantial enrichments in phenotypic annotations whose power in predicting pathoetiology greatly exceeds that of molecular annotations.
For this study, 148 MR–associated rare CNVs collated from a variety of sources (Table S1) were merged to obtain a set of 112 distinct non-overlapping CNV regions (CNVRs) and partitioned according to the direction of copy number change (Gain or Loss). We also collated a control set of 26,472 benign CNVs (1,388 CNVRs) from previous publications (see Materials and Methods). MR–associated CNVs are most obviously distinguished from benign CNVs by their large sizes and by their larger numbers of copy number losses (n = 111, 75%) relative to gains (n = 37, 25%) (Table 1). These differences remained even when comparing benign and MR CNVs detected by the same platform (tiling resolution 32 k BAC arrays): the median size of 40 MR CNVs is approximately twice that of benign CNVs (1.6 Mb versus 0.85 Mb) while 58.6% of benign CNVs on this platform are losses. This increased bias towards loss CNVs would be expected if the MR phenotypes considered here result either from haploinsufficiency or from recessive deleterious mutations being revealed in the remaining haplotype. There is only a small difference (17.6%) between the average gene densities of MR–associated and benign CNVs (Table 1). Consequently, we need to look to gene function, rather than gene numbers, when attempting to differentiate disease-associated from benign CNVs.
We first tested whether MR–associated CNVR genes were enriched in 33 major categories of mouse phenotypes (see Materials and Methods). Although for All MR–associated CNVRs none of these terms was significant, the set of Loss MR–associated CNVRs showed a strong and significant enrichment in genes whose knockouts in mice produced a nervous system phenotype (+13.6%, or 1.14-fold, enrichment, p = 3×10−3, FDR<5%; Figure 1). An enrichment of genes associated with nervous system phenotypes was not observed within the Gain CNVRs (+0.2%).
Given the significant enrichment within the Loss set, we then tested this set against each of 147 finer-scale mouse nervous system phenotypes. Two of these terms were significantly enriched (FDR<5%): abnormal axon morphology (obs = 19, exp = 7.1, +170% enrichment, p = 3×10−5), and abnormal dopaminergic neuron morphology (obs = 9, exp = 2.5, +260% enrichment, p = 3×10−4) (Figure 1). Both of these mouse neural phenotypes are relevant to human MR phenotypes owing to these mouse phenotype's abnormalities in neuronal and cerebral cortex morphologies (see Discussion). Within Gain CNVRs, we observe a non-significant enrichment of genes associated with abnormal axon morphology (obs = 6, exp = 2.7, +120% enrichment, p = 5×10−2) but a non-significant depletion of genes associated with abnormal dopaminergic neuron morphology (obs = 0, exp = 0.95, −100% deficit, p = 0.38).
The neurological phenotypes of MR patients suggested that MR–associated CNVs might contain an unusually high density of genes that, when mutated, are involved in human neurological disease. Considering those genes classified by KEGG to be involved in 6 neurodegenerative pathways, we indeed found MR–associated CNVRs to be significantly enriched in genes involved in the Parkinson's disease pathway (obs = 8, exp = 2.7, +196% enrichment, p = 3×10−3, FDR<5%; Figure 2). While enrichments of this pathway's genes were observed both for Loss CNVRs (obs = 7, exp = 2.1, +230% enrichment, p = 3×10−3, FDR<5%) and for Gain CNVRs (obs = 2, exp = 0.8, +151% enrichment, p = 0.19), significance was reached only for Loss CNVRs. As Parkinson's disease is a condition characterized by the degeneration and dysfunction of dopaminergic neurons [19], these enrichments corroborate our finding that orthologues of genes whose disruption in mouse gives rise to abnormal dopaminergic neuron morphology are enriched in MR–associated CNVRs (see above).
The allelic changes underlying MR phenotypes might also be expected to preferentially involve ‘brain-specific’ genes, those that are highly expressed in the human brain relative to other human tissues. Indeed, All MR–associated CNVRs were significantly enriched in brain-specific genes (+24% enrichment, p = 1×10−2; Figure 3), specifically for Loss (+31% enrichment, p = 8×10−3) but not for Gain CNVs (+4% enrichment, p = 0.45). The significant enrichments observed when testing mouse phenotypes are thus corroborated by enrichments in human gene expression.
These findings would have little or no predictive potential if apparently ‘benign’ CNVs (those present in the general human population) also exhibit such biases. However, in contrast to the above results, benign CNVs show no significant enrichments of (i) genes that are highly-expressed in the brain (−11% deficit, p = 0.2; Figure 3), (ii) genes present in neurodegenerative disease pathways (−32% deficit, p = 0.1; Figure 2), or (iii) genes with nervous system phenotypes when disrupted in mice (−11% deficit, p = 0.01; Figure 1). Instead, benign CNV genes show significant tendencies to encode proteins with roles in immunity and host defense [20],[21]. Each of these three features thus may be exploited to distinguish MR–associated CNVR genes from benign CNVR genes.
MR–associated and benign CNVs show no significant tendency to overlap (p = 0.1). Nevertheless, by excluding all genes in MR–associated CNVs whose gain/loss-matched copy number change is also seen in benign CNVs we enhanced the discrimination of genes whose copy number change is predicted to contribute to MR aetiology. This was specifically the case for mouse fine-scale nervous system phenotypes and human neurodegenerative disease pathways (Figure 1 and Figure 2). Moreover, after excluding benign CNV-overlapped genes, not only Parkinson's disease pathway genes, but genes from 5 other neurodegenerative disease pathways (namely, Alzheimer's disease, Amyotrophic Lateral Sclerosis, Huntington's disease, Dentatorubropallidoluysian atrophy and Prion Diseases) when considered together, became significantly enriched (+60% enrichment; p = 0.02) in this analysis. These results would be explained if MR-causative alleles segregate more with sequence that is copy number variable in MR individuals than with CNVs observed in the general population.
We considered whether our method could identify significant associations between mouse and human patient phenotypes other than MR. We investigated 7 clinical features that were present in our patient population in addition to the MR phenotype, namely brain-, cleft palate-, eye-, facial-, heart- or urogenital- abnormalities and seizures (see Materials and Methods). We tested whether CNVs from individuals with these specific clinical features were significantly enriched in genes associated with phenotypically-relevant mouse phenotypes. In order to limit the large number of statistical tests that could be performed we matched mouse phenotype categories (each containing between 129 and 220 terms) to each of the 7 clinical features based on clinical experience (see Materials and Methods) before performing the association tests. We found that 4 of the 7 additional clinical features were significantly associated (FDR<5%) with between 1 and 6 mouse phenotypic terms (Figure 4). For example, the CNVRs of the 8 MR patients presenting with cleft palate were significantly enriched with genes whose mouse orthologues, when disrupted, also exhibited cleft palate (Figure 4). Importantly, no significant associations were observed between CNVs from humans without a particular clinical feature apart from MR and any mouse phenotype category matched to patients with that clinical feature, with the notable exception of ‘abnormal axon morphology’ that thus appears to be a term of broad relevance to the primary MR presentation (Figure 4). These findings demonstrate the relevance of mouse gene knockout observations to both the MR phenotype and associated phenotypes in patients.
The distinctions between MR–associated and benign CNVR genes, described above, allowed the identification of genes whose copy number change may contribute to MR and associated phenotypes. To identify such candidate genes, we could not exploit Gene Ontology annotations (Figure S1) or brain expression enrichments (Figure 3) as these enrichments provide insufficient discriminatory power (<30% increase over expected). Of the 4,009 genes present in the 148 MR–associated CNVs, 55 are annotated with either a mouse knockout phenotype (n = 29) and/or a neurodegenerative disease pathway (n = 29) that was significantly over-represented in MR–associated Loss CNVRs (Table 2). 50 of the MR–associated CNVs (33%) contain at least 1 of these 55 candidate genes. We calculate that our list represents a ∼120% increase of likely phenotype-contributing genes over the random expectation (see Materials and Methods). Similarly, 34 genes were identified as potential candidates for additional clinical features such as cleft palate, facial or brain abnormalities, or seizures, 23 of which were not associated with MR itself (Table 2). We note that whilst some of these candidate genes might have been prioritized from among the 4,009 CNVRs genes using a priori subjective expectations, our method is the first to generate a candidate gene set on the basis of objective and statistically sound criteria.
If de novo MR–associated CNVs do not contribute to disease etiology their gene contents would not be expected to exhibit biases in gene function or expression. Instead, we demonstrate the first evidence for significant tendencies of MR–associated CNV genes to be brain-expressed, to belong to neurodegenerative pathways, and to present particular phenotypes when disrupted in mice, all of which validate the assumption that large de novo CNVs commonly underlie MR phenotypes. These results could not have been obtained without collating data from a number of sources. For example, essentially all (147 of 148) CNVs were required to obtain a significant enrichment of genes whose mouse orthologues' knockout produced a nervous system phenotype (Figure S2). It was only by harnessing the statistical power of a research community's large data set that this meta-analysis achieved significance of statistical associations (see Materials and Methods).
The significant signals seen in Loss CNVs, but not in Gain CNVs, imply that MR phenotypes commonly result from gene dosage sensitivity (haploinsufficency). However, we cannot discount that they may occur from the uncovering, by DNA loss, of rare recessive alleles. While we did not observe an enrichment within the Gain CNVRs of genes associated with abnormal dopaminergic neuron morphology or of genes that showed brain-specific expression, we did observe non-significant enrichments of genes associated with abnormal axon morphology and of Parkinson's disease pathway genes. Given that the Gain CNVRs overlap 38% of the number of genes overlapped by the Loss CNVRs (Table 1), it is plausible that these enrichments might reach significance as more Gain MR–associated CNVs are reported and analysed.
Our results are in contrast with previously-reported sporadic and familial cases of MR whose associated genes are enriched in both X-chromosome location and enzymatic function [22]. Nevertheless, this is explained by Wright's physiological theory of dominance: haplosufficient genes, such as those lying on the X chromosome, have an expected tendency to encode enzymes, whereas haploinsufficient genes, such as those expected to underlie our autosomal MR disorders, have an expected tendency to encode transcription regulatory genes [23]. Indeed, we do observe a significant enrichment of genes associated with transcriptional regulation within MR–associated CNVRs (Figure S1). In contrast to X-linked MR genes, of which approximately one quarter encode postsynaptic proteins [24], we observe a small and non-significant depletion (p = 0.39) of postsynaptic protein genes among our MR–associated CNVs.
None of the human CNVs recorded in this study represent homozygous losses. Thus it may initially appear problematic to compare human phenotypes directly with those from mice harbouring homozygous gene disruptions. Nevertheless, without sequence information confirming the genetic integrity of the surviving haplotype we cannot be certain that these human hemizygous loss CNVs do not contain independent disruptions of each allelic copy. To gain some insight into this issue we considered 21 of the 55 candidate genes that contribute to a significantly enriched mouse knock-out phenotype identified in our study (Table 2), and whose phenotype has been recorded in the MGI resource when in the hemizygous state. Of these 21, four (namely, En1, Mn1, Plp1 and Pmp22) also exhibit the phenotype of interest when hemizygously disrupted [25]–[28]. Of the remaining 17 genes, all exhibit abnormal phenotypes, and thus are haploinsufficient, with the exceptions of Mapt and Slc6a3 [29],[30]. Importantly, these mouse hemizygous phenotypes are often closely-related to the homozygous phenotypes, while some hemizygous phenotypes appear particularly relevant to the associated human phenotype. For example, Scn1a (which contributes to the tremors phenotypic enrichment we find to be associated with patients presenting with seizures) exhibits a seizures phenotype when in the hemizygous state in mice [31].
Does our analysis allow us to link particular mouse gene knockout phenotypes to human CNV phenotypes? Obviously, a direct comparison between mouse neural phenotypes and human MR phenotypes is hindered because the invasive procedures of brain biopsies in patients are unacceptable. Results from a limited number of post-mortem studies of MR patients suggest that abnormalities of dendritic spines are a general neuropathological feature of MR [32]. The mouse gene knockout phenotypes do provide a plausible explanation for the brain phenotypes observed in some patients as a consequence of the structural variation identified in their genomes. An example of this is the myelin-associated glycoprotein (MAG) gene that is deleted in one patient (case 123, Table S2) and duplicated in another (case 124), whilst the knockout of its orthologous gene in mice leads to both abnormal axon morphology and tremors phenotypes [33]. Underexpression of MAG in transfected Schwann cells is known to lead to hypomyelinisation [34]. Therefore, the delayed brain myelinisation observed in the patient with the MAG deletion could be caused by under-expression of MAG during brain development. By contrast, over-expression of MAG is known to lead to accelerated myelinisation [35]. Whether the macrocephaly in the patient with the MAG duplication is related to over-expression of MAG during brain development remains unknown.
Our enrichment analysis revealed 8 genes associated with cleft palate in humans, present in 6 different patients (cases 10, 13, 27, 48, 96, and 141). Seven of these genes were located in Loss CNVs on human chromosomes 1p31.1p31.3 (containing LHX8), 1q41q42.13 (DISP1), 2q24.3q31.1 (DLX1, DLX2 and GAD1), 4q31.21q31.23 (EDNRA) and 22q12.1 (MN1), and one with a Gain CNV on human chromosome 16p13.2–p13.3 9 (CREBBP). Except for DISP1, all these genes have been associated with cleft palate in mouse models [26], [36]–[39], whereas only LHX8 and GAD1 have been associated with cleft palate disorders in humans [40],[41]. This strongly suggests that our approach revealed 6 novel orofacial cleft (OFC) candidate genes in humans. Strikingly, the hemizygous loss of five of these OFC candidate genes may also contribute to MR. Absence of both Dlx1 and Dlx2 in mice results in abnormal differentiation within the forebrain [36],[42]. Both genes also regulate Arx, a homeobox transcription factor required for the migration of interneurons, whose human equivalent ARX, when mutated, is associated with X-linked MR and epilepsy [43]. In addition, mutations and deletions of CREBBP causes the Rubinstein-Taybi syndrome which is characterized by MR [44]. Ednra is involved in cranial neural crest cell migration from the posterior midbrain and hindbrain to the arches [45]. Lhx8 is required for the development of many cholinergic neurons in the mouse forebrain [46], whereas GAD1, which encodes the GABA-producing enzyme, may play a role in the development and plasticity of the central nervous system [39]. In conclusion, it appears that our approach identified a large number of interesting and plausible novel candidate genes for both MR and associated clinical phenotypes.
Mouse phenotype data have not previously been exploited in a systematic genome-wide analysis, and our results clearly show its utility in addressing a particularly difficult and contemporary challenge in the field of neurological genomic disorders. The functional biases we see for MR–associated CNV genes can now be exploited to prioritise genes for further investigation in MR individuals without large de novo CNVs (Table 2). We suggest that all human genes whose orthologues present specific phenotypes when disrupted in mice (Figure 1) deserve particular scrutiny for fine-scale insertion, deletion or point mutations contributing to MR. Mouse orthologue knockout data are available currently for only ∼25% of all human genes. More specifically, of the 4,009 genes overlapped by the MR–associated CNVs considered here, 830 (∼21%) have available phenotypic annotations. Thus, we would expect that many more candidate genes possessing these annotations will be discovered within MR–associated CNVs as further knockouts are generated. Furthermore, we consider all genes that are involved in the specific molecular pathways we have identified, such as Parkinson's disease and other neurodegenerative disorder pathways, to represent candidates for MR and/or associated phenotypes when hemizygous. We propose that the contribution of these candidate genes (Table 2) to many MR phenotypes can now be investigated thoroughly in mouse model systems: specifically, the 55 genes whose hemizygous deletions may be associated with MR are now amenable to study using hemizygous knockout mouse models.
Our study has exploited CNVs identified using several different platforms. As the identification technologies have improved, CNVs called using earlier technologies have been shown to over-estimate the true extent of a CNV's boundaries [47]. Thus, we expect enhanced resolution of pathogenic CNVs to also increase the power by which genic enrichments can be identified. However, it should also be noted that CNVs have been shown to affect the expression of neighbouring genes and it is possible that pathogenic CNVs may exert their genetic effect through outlying genes [48].
Finally, there is no reason why this approach can not be applied successfully to other complex neurological diseases, including schizophrenia and autism, which show a high frequency of rare de novo CNVs [8], [9], [49]–[51]. Many studies that are currently under-powered to demonstrate significance after correcting for multiple testing may yet prove informative of the genetic etiology of complex genomic disorders. For this, it will be crucial to collect large disease-associated CNV sets from well-phenotyped cohorts, as our analysis has shown that only then is there sufficient power to detect significant associations (Figure S2).
For this study we collected 148 rare structural variants associated with MR from the literature, the Decipher database (https://decipher.sanger.ac.uk/), as well as from our own in-house diagnostic microarray group [52] (Table S1). The majority of these CNVs (n = 135, 91%) were proved to have occurred de novo in the patient and all were independently validated. Thirteen rare autosomal CNVs for which parental samples were unavailable were included, as were seven rare maternally inherited CNVs on the X chromosome in male patients that are considered to be as clinically relevant as de novo CNVs on the autosomes. Importantly, at the point of discovery none of these CNVs were known to greatly (>50%) overlap with a collection of >15,000 CNVs identified in healthy individuals as collected in the Database of Genomic Variants version 3 (http://projects.tcag.ca/variation/). All CNVs were mapped to NCBI35 coordinates. The median number of Entrez genes within a CNV was 35. Overlapping CNVs were merged to obtain a non-redundant set of 112 CNV regions (CNVRs) totalling 440 Mb of unique sequence (14.3% of the total NCBI35 human genome assembly; Table 1). CNVR sets were also formed separately from Gain and from Loss CNVs (Table 1). For 121 of the 148 CNVs, information regarding distinct anatomical or physiological abnormalities presented by the patient in addition to MR was available (Table S2). These clinical features were used to form 7 non-exclusive groupings for additional tests.
We obtained 25,196 CNVs identified in 270 individuals from Redon et al. [11]. To these, we added 1,276 inherited CNVs identified in 494 individuals with a 32 k BAC tiling path array. This last set is described in Nguyen et al. [53] and, together with the Koolen et al. [52] MR–associated CNV data, are available from the Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) with accession number GSE7391. Combined, these apparently benign CNVs represent 430 Mb of unique sequence (14.0% of the total NCBI35 human genome assembly; Table 1). In the absence of information suggesting that any of the individuals present with MR, we conservatively assume that genes overlapped by these apparently benign CNVs do not contribute to the MR phenotypes.
Assignment of protein-coding genes depended upon the particular analysis performed: for protein-coding gene counts and the Gene Ontology analysis, we assigned genes to CNVs according to Ensembl [54] (Ensembl mart version 37), whereas for KEGG pathway and MGI analyses we assigned genes to CNVs according to Entrez genes [55].
Information on human NCBI genes whose mouse orthologues' disruption had been assayed were obtained from the Mouse Genome Informatics (MGI) resource (http://www.informatics.jax.org, version 3.54) [12]–[14]. We employed the MGI's human/mouse orthology and marker assignment to map MGI mouse marker phenotypes to Human Entrez genes [55]. We mapped, using unambiguous gene orthology relationships, 5,075 different MGI phenotypic annotation terms to 4,999 human genes. We considered all phenotypic annotations from all experimental methodologies described within the MGI resource. While the vast majority of these annotations are derived from the disruption of mouse genes, some phenotypes were derived from experiments in which mutant alleles are introduced into the mouse (e.g. [56]). Nonetheless, we regard the phenotypic information from these experiments as remaining informative of the biological functions or pathways to which the gene contributes. It is noted, however, that the phenotypes of all genes underlying the phenotypic enrichments we report in this work (Figure 1 and Figure 2; Table 2) were obtained through gene disruption experiments.
The MGI phenotypic annotations are categorised non-exclusively into 33 over-arching terms (Table S3). When examining finer phenotypic terms beneath an over-arching term(s) we considered only those finer terms that possessed at least 1% of the genes annotated with the over-arching term(s). This allowed a reduction in the number of tests performed thereby limiting spurious and uninformative results. The phenotypes associated with the Entrez genes overlapped by a given set of genomic regions were compared to the frequency of that phenotype across the whole genome. All p-values were obtained by application of the hypergeometric test and were subject to a false discovery rate (FDR) of <5% [18] (see below). Given the large number of phenotypic terms and the unrealistic assumption of terms' independence when applying an FDR, application of this significance threshold is likely to be conservative.
Many of the MR patients used in this study show additional clinical features. We tested for associations between commonly occurring non-MR clinical features in patients and a subset of MGI phenotypes. We scored patients for the presence of 7 common features derived from the London Dysmorphology Database [57]. These were: (i) seizures/abnormal EEG, (ii) facial dysmorphism, (iii) cleft palate, (iv) heart, general abnormalities, (v) eye abnormalities, (vi) brain, general abnormalities, and (vii) urogenital system abnormalities. Patients were excluded if specific phenotypic data were unavailable (all 19 cases from the Decipher database). As these secondary clinical feature-grouped CNVs were fewer in number than the entire set of MR–associated CNVs, and therefore relatively diminished in statistical power, the most relevant MGI phenotypic categories were selected (from a total of 33; Table S3) in order to reduce the number of tests. Two pairs of paralogous genes, DLX1 & DLX2 and SELE & SELP, contributed to the significant phenotypic enrichments reported within the secondary clinical feature grouped CNVs (Table 2). However, significant phenotypic enrichments that these pairs of paralogues contributed to all remained significant after removing one of the paralogous pairs (p<0.05; single test). Nevertheless, we note that an increased penetrance of a resulting phenotype might be expected if these pairs of paralogues provided a degree of redundancy to one another, and therefore the concurrent copy number variation of both paralogues may prove even more significant than variation involving only one [42].
Annotations of genes involved in neurodegenerative pathways were obtained from KEGG [17]. KEGG genes were collated if they belonged to KEGG Pathways section 5.3, namely Alzheimer's disease (KEGG pathway 05010), Parkinson's disease (KEGG pathway 05020), Amyotrophic Lateral Sclerosis (KEGG pathway 05030), Huntington's disease (KEGG pathway 05040), Dentatorubropallidoluysian atrophy (KEGG pathway 05050) and Prion Diseases (KEGG pathway 05060). KEGG genes were mapped to NCBI Entrez genes using associations provided by KEGG.
For human gene expression data, we used GNF's gene atlas data for the MAS5-condensed human U133A and GNF1H chips, considering all 74 non-cancer tissues [16]. Expression levels were mapped to LocusLink identifiers and to 11,594 Ensembl Ensmart 37 (NCBI35) genes using the annotation tables supplied by GNF. To identify genes that are highly expressed in the brain we selected those genes whose expression in the whole brain exceeded by 4-fold their median expression in all other non-brain tissues after excluding cancerous tissues. This resulted in 435 genes (3.75%) being classified as exhibiting strong expression in the brain relative to other tissues. However, the significant enrichments reported in the Results were also found when brain-specificity was redefined at 2-, 3-, 7-, 10-, 11-, 12-, 13-, and 14-fold expression in the brain above the median across all other tissues.
A set of postsynaptic protein genes was obtained from Collins et al. [58] and matched to human orthologues using Ensembl Compara [59]. Over- or under-representation of these genes within human CNVs was assessed using the hypergeometric distribution and all human Ensembl genes as the background set.
The significance of enrichments or deficits of genes associated with particular MGI knockout phenotypes, genes involved in KEGG neurodegenerative pathways, genes associated with particular GO terms and brain-specific genes were evaluated using hypergeometric tests. Where multiple tests were performed, a False Discovery Rate (FDR) multiple testing correction was applied to ensure a less than 5% likelihood of any significant term being a false-positive [18]. Explicitly, an FDR correction was applied when testing for enrichments of genes: (i) associated with MGI phenotypic terms, (ii) belonging to individual KEGG neurodegenerative pathways or (iii) annotated with Gene Ontology terms (Figure S1). All other tests performed were single tests.
Calculation of the fold-enrichment within MR–associated CNVs for the final set of 55 MR–associated candidate genes was performed by random sampling. 1000 gene sets, matched in gene number to that within the Loss MR–associated CNVRs, were obtained by random sampling and the median expected number of genes, 23 (std.dev. = 4.6), annotated with one or more significantly-enriched terms (Figure 1 and Figure 2) was recorded. Given the 50 candidate genes within the Loss CNVRs, we thus estimate a ∼2.2-fold enrichment over the number expected by chance.
|
10.1371/journal.pntd.0007661 | Vaccination coverage in the context of the emerging Yellow Fever threat in French Guiana | French Guiana, a French overseas department located in South America between Brazil and Surinam, is the only European territory geographically located in the Amazonian forest complex and is considered endemic for yellow fever (YF).
In the context of the emergent threat of YF in Latin America, we conducted a large household cross-sectional survey from June to October 2017 to estimate vaccination coverage in the population and to determine associations with sociodemographic and geographical characteristics.
In total, 1,415 households and 2,697 individuals were included from the 22 municipalities of French Guiana. YF vaccination coverage was estimated at 95.0% (95% CI: 93.4–96.2) in the entire territory but was spatially heterogeneous, with the lowest levels estimated in the western part of the territory along the Surinamese cross-border region, particularly in children under 16 years who were not enrolled in school, immigrant adults and disadvantaged populations with low socioeconomic indexes.
Despite the good vaccination coverage against YF in the general population of French Guiana resulting from the compulsory nature of YF vaccination for residents and travelers, there is an urgent need to improve vaccination coverage in vulnerable populations living in the northwestern part of the territory to limit the risk of transmission in the context of the emerging YF threat in South America.
Despite the relative rarity of YF and the significant number of infectious and tropical diseases in French Guiana, clinicians should adopt a high index of suspicion for YF, particularly in vulnerable and at-risk populations.
| Yellow fever (YF) is the most severe arbovirus to circulate in the Americas. French Guiana, a French overseas department located in South America between Brazil and Surinam, is the only European territory geographically located in the Amazonian forest complex and is considered endemic for YF. We conducted a large general population survey from June to October 2017 to estimate vaccination coverage in the population and to identify target vulnerable populations for catch-up vaccination strategies. In total, 1,415 households and 2,697 individuals were included from the 22 municipalities of French Guiana. YF vaccination coverage was estimated at 95.0% (95% CI: 93.4–96.2) in the entire territory but was spatially heterogeneous, with the lowest levels estimated in the western part of the territory along the Surinamese cross-border region, particularly in children under 16 years who were not enrolled in school, immigrant adults and disadvantaged groups of populations with low socioeconomic indexes. Our findings showed that vaccination campaigns should be prioritized and adapted to improve vaccination coverage among vulnerable populations living in the northwestern part of the territory to limit the risk of transmission in the context of the emerging YF threat in South America. Despite the relative rarity of YF and the significant number of infectious and tropical diseases in French Guiana, clinicians should adopt a high index of suspicion for YF, particularly in vulnerable and at-risk populations.
| Yellow fever (YF) is the most severe arbovirus to circulate in the Americas, with symptoms ranging from mild non-specific illness to hemorrhagic fever, a systemic illness characterized by high viremia, hepatic, renal and myocardial injury, hemorrhage, and high lethality [1]. A single-dose vaccine has existed since the 1940s and has helped to substantially control and reduce YF transmission [2–4]. However, complete eradication is prevented by the sylvatic cycle of the virus within nonhuman primary hosts, and Aedes aegypti mosquitoes are responsible for occasional transmission to people [5]. Recent important outbreaks of YF in Africa and South America have confirmed the potential of arthropod-borne viruses to emerge or reemerge in risk areas [6] and have highlighted the urgent need to assess vaccination coverage efforts in the most exposed countries.
Since November 2016, after decades of silence, Brazilian authorities and scientists have reported an outbreak of YF associated with an exponential increase in the number of confirmed cases and deaths in humans [7–9]. The YF virus has spread into the coastal Atlantic forest zones and moved rapidly into the southeast and south of the country in less than one year, reaching several populous Brazilian states whose residents had not been included in the YF vaccination program [9,10]. The majority of reported cases have occurred in rural areas, clearly reflecting a typical sylvatic transmission cycle occurring between forest mosquitoes and forest-dwelling nonhuman primates, with humans serving only as accidental hosts. This important alert has prompted the Brazilian Ministry of Health to conduct massive vaccination campaigns among unvaccinated residents of affected areas [11].
French Guiana, a French overseas department located in South America between Brazil and Surinam, is the only European territory geographically located in the Amazonian forest complex and is considered endemic for YF [12]. Since 1967, YF vaccination has been compulsory in French Guiana for all individuals older than 1 year of age (with a booster dose every 10 years). The vaccination is free of charge and widely accessible in public vaccination centers and by accredited private practitioners. In February 2016, according to the Strategic Advisory Group of Experts on Immunization and the modifications of the International Health Regulations [13], French health authorities adopted the use of only a single dose of vaccine for most residents and travelers.
Over the last decades, Ae. aegypti has been responsible for several major dengue fever outbreaks [14–18] and for the recent emergence of chikungunya in late 2013 [19–21] and Zika in 2016 [22,23]. Considering the large number of travelers moving through the Brazilian river border, the recent outbreak of YF has raised particular concern that an urban transmission could occur in French Guiana, specifically for nonvaccinated population subgroups, through the Ae. aegypti mosquitoes that are strongly represented in the territory.
Although vaccination coverage reported by recent assessments was better than that observed 20 years ago, some geographical areas may present unsatisfactory levels of coverage, particularly in forest and exposed environments [24–27]. While a survey conducted in 2000 estimated YF vaccine coverage of 80–90% in children under 15 years old [24], the overall vaccine coverage was estimated at 95.9% (95% CI 95·5–96·3) in 2009 in 9339 children from primary and secondary schools [25]. Lower coverage rates between 75% and 81% were observed in small municipalities located outside the urbanized and coastal areas.
Moreover, French Guiana is experiencing continuous major waves of immigration facilitated by the natural and uncontrollable quality of the river borders, particularly from countries where vaccination against YF is not mandatory.
In this context, two sporadic and fatal cases of YF were reported in French Guiana 1 year apart [28], confirming that sylvatic YF circulation is active in this territory, particularly among nonvaccinated populations involved in important activities in the forest environment. The first case was confirmed in August 2017 in a 43-year-old Brazilian woman with unknown vaccination status. Epidemiological investigations reported a history of stay in the forest, suggesting that the patient could have been contaminated either in the Amapá state in Brazil or in French Guiana. In August 2018, a second case was biologically confirmed in a non-vaccinated Swiss-citizen 47-year-old man who had entered through a river border and had lived in French Guiana for 4 months. Epidemiological investigations reported regular work activities on forest roads, suggesting autochthonous transmission in the forest environment.
The last autochthonous case was identified in 1998 in the southeast of the territory[29]. These two recent case reports illustrate that despite the compulsory nature of YF vaccination in this French overseas region since 1967, maintaining focus on the need for YF vaccination is important, especially in areas with favorable ecosystems for YF transmission.
In the context of the emergent threat of YF in Latin America and, consequently, in French Guiana, we conducted a general population cross-sectional study to estimate vaccination coverage in the population and to determine associations with sociodemographic and geographical characteristics.
The study was approved by the “Sud-Ouest & Outre-Mer IV” Ethical Research Committee (No. CPP17-007a/2017-A00514-49) and by the French Data Protection Authority (No. DR-2017-324), which is responsible for ethical issues and the protection of individual data collection.
Publicity and information about the survey were provided through the media and contact with local and national authorities. Fieldworker teams were trained to visit all households, explain the project objectives, and, when allowed, collect participants’ signatures on a free and informed consent form and conduct the interviews. All household members of selected households who were 2–75 years of age were invited to take part in the study. For all participants under 18 years of age, one or two responsible adults signed the informed consent form. Data were collected through a standardized questionnaire installed on tablets to register demographic, socioeconomic and household characteristics. Vaccination cards or any other proof-of-vaccination documents were requested from all participants. Vaccination status was based on presented vaccination cards and verbal reports of vaccination when vaccination cards were not available.
We employ the following notations to describe the study design:
We considered that in each municipality i, the probability of selecting a particular subject was equal to the probability of selecting the subject’s household and was (mi/Mi), corresponding to a statistical weight equal to (1/ mi/Mi) = (Mi/mi). This statistical weight indicates the number of people in the population represented by each subject in the sample.
We applied a post-stratification adjustment to each of these weights to arrive at the final statistical weight for each subject. This adjustment helped us to weight the age-sex groups within each municipality to match the distribution in the total population of French Guiana. Ten age groups ([2–5 years [, [5–10[, [10–15[, [15–20[, [20–25[, [25–35[, [35–45[, [45–55[, [55–65[, and ≥65 years) were defined within male and female groups. For each age-sex subgroup, we applied an adjustment factor cijk to obtain a final statistical weight:
wijk = (Mi/mi)*cijk, where i indexes municipalities, j indexes sex groups and k indexes age groups.
We constructed a household socioeconomic index combining a multiple correspondence analysis and a hierarchical cluster analysis based on household material possessions, socioprofessional category and household income.
The weighted vaccination coverage estimation for one dose of YF was based on doses recorded on vaccination documents and/or reported by participants.
Associated factors were identified using survey-weighted Poisson regression, and the strength of selected variables and vaccination coverage was estimated by raw and adjusted risk ratios (RR) and a 95% confidence interval (CI). All RRs excluding 1.0 were considered significant. Analyses were conducted using the survey capabilities of Stata version 15 statistical software (Stata Corp, College Station, TC, USA)[31]. French Guiana’s layers were drawn using geodata from OpenStreetMaps (http://www.openstreetmap.org), and mapping operations were performed using QGIS 2.18 software [32].
In total, 1,415 households and 2,697 individuals were included from 22 municipalities (Table 1), representing 58% of eligible household members. The mean household size was 1.9 individuals [range: 1 to 11]. The mean age was 30.5, ranging from 2 to 75 years old. Comparison of the sociodemographic characteristics of the study sample with census data demonstrated an overrepresentation of women (58.9% vs. 50.0% in the general population of French Guiana) and adults over 25 years (64% vs. 53% in French Guiana). These differences were accounted for in the analyses of vaccination coverage and risk factors by allocating a poststratification weight to each participant. Vaccination coverage for YF (at least one dose) was estimated at 95.0% (95% CI: 93.2–96.3) throughout the entire region of French Guiana (Table 1). Eighty percent of the respondents presented a vaccination certificate or an equivalent document as evidence of vaccination, while 15.6% reported that they had received the vaccination but had no hard evidence. The number of booster doses received by vaccinated individuals ranged from 1 to 6. The mean age at vaccination was 1.7 years among children under 10 born in French Guiana.
The coverage was spatially heterogeneous and decreased from the central coastal area to the western part of the territory along the Maroni River, which forms the border with Suriname (Fig 2). The highest vaccination coverage levels were observed in small and remote villages or in municipalities with fewer than 3,000 inhabitants (Antecume Pata, Twenke-Talhuen, Iracoubo, Roura, and Sinnamary), where all the respondents were vaccinated.
While the majority of French Guiana had coverage levels higher than 90%, three municipalities located in the west border area (Fig 2) had relatively low levels of vaccination coverage, including Grand-Santi (62.3%, 95% CI: 39.3%–80.9%), Saint Laurent (76.9%, 95%CI: 67.7%-84.1%) and Papaïchton (78.3%, 95% CI: 60.2%-89.6%).
More importantly, vaccination coverage was particularly low in children under 16 years in the municipalities of Saint-Laurent (40.1%, 95% CI: 21.6–61.9), Papaïchton (51.9%, 95% CI: 26.0–76.8) and Grand-Santi (52.7%, 95% CI: 25.1–78.7), while the coverage level was estimated at 97.1%, 95% CI: 94.5–98.5 in the other municipalities. More than 40% (N = 23/51) of unvaccinated children aged 3–16 years living in these municipalities were not enrolled in school. Most of them (91%) had parents born in Surinam or Guyana, and 72% of the children were born in French Guiana. Nearly 70% of the unvaccinated adults living in these municipalities were born outside of the country, and 30% had lived in French Guiana for less than 10 years. Disadvantaged groups benefiting from universal health coverage and state medical assistance schemes (specifically conceived for undocumented migrants who become eligible after 3 months of residency in the French territory) and those with low socioeconomic indexes were also associated with lower vaccination coverage in the entire territory (Table 2).
In 2017, epizootic and sporadic human cases were observed in the northern part of the state of Pará, Brazil, near French Guiana. A recent case in neighboring Suriname was also identified in the Brokopondo Lake area, less than 100 km from the river border with French Guiana [28,33]. These situations suggest ongoing viral circulation and an emerging threat in the wider Guiana Shield region. In this context, it was incumbent upon us to estimate YF vaccination coverage throughout the entire region of French Guiana.
Our results highlight good vaccination coverage against YF in the general population of French Guiana resulting from the compulsory nature of YF vaccination for residents and travelers [34].
However, vaccination rates appear to be insufficient in some western cross-border areas connected by river routes (outside of vaccination control) to countries potentially lacking sufficient vaccination coverage. This includes areas or countries where recent epidemics have occurred or where vaccination against YF is not mandatory [9,33].
While vaccination coverage estimates in French Guiana were the lowest in some western cross-border areas, including Grand Santi (62.3%), Papaïchton (78.3%), and Saint-Laurent (76.9%), the actual vaccination rate in these municipalities may be lower than the level recommended by the WHO to achieve and maintain a protective population-level immunity (assumed to be approximately 60–80%) [35]. This concern is based on the very low proportion of individuals who provided proof of vaccination in this part of French Guiana.
Importantly, a large number of unvaccinated individuals in these areas were out-of-school children who were consequently not involved in vaccination monitoring and catch-up strategies conducted in formal educational settings. This situation poses an additional challenge for health authorities and prevention operators to reach and include these populations in vaccination catch-up strategies. While we estimated the population of children aged 3–16 years not attending school in the entire territory to be 5.4%, 75% of them lived in these western cross-border municipalities.
Furthermore, it is possible that our study tends to overestimate vaccination coverage, particularly in specific areas associated with a low proportion of presentation of vaccination proof. Although a large majority of individuals who reported that they had received vaccination without hard evidence were able to provide the date and the service of vaccination, information on vaccination history without a card or hard evidence could be falsified.
Another limitation of our study is that irregular immigrants without health coverage were underrepresented in our sample. Given that individuals without health coverage could not be enrolled in our survey because of restrictions from French legislation, this population was underrepresented in our study. Although this population was very small in most of the main municipalities of the coastal areas, some households were excluded in the western part of the territory, which is known for high levels of immigration, because the adults and referents of the selected household did not have health insurance status.
Six individuals without health insurance status were included from households whose referents were eligible and enrolled in the survey. Only three of them had received single doses of vaccination, suggesting that recent immigrants, who are often in irregular situations and targeted by police operations, are at risk of being unvaccinated and are difficult for health professionals to reach. This situation may lead to YF cases or clusters in specific and unvaccinated population subgroups.
In this context, it should be a priority to focus vaccination campaigns in the northwestern part of the territory where vaccination coverage rates are the lowest and most likely overestimated. Vaccination strategies and campaigns should be adapted to continuously improve vaccination coverage in children who are not enrolled in school, migrant populations who have recently arrived in French Guiana and other vulnerable populations, particularly if they are involved in essential activities in forest areas. Although the Social Security Fund and local health authorities provide yellow fever vaccination free of charge in the entire territory, vulnerable and unvaccinated populations without health insurance status may have poor access to health care providers and limited opportunities to be considered in the vaccination campaigns that are usually conducted in school and health centers. Alternative community initiatives based on places of religious worship, citizen’s centres or other places for social gathering should be used to reach these populations to increase vaccination coverage in target populations. Moreover, it is essential to maintain vaccine strategies and policies related to airport vaccination status controls and to raise awareness among health-care providers regarding the importance of verifying the immunization status of patients at each encounter regardless of patient origin to contribute to vaccination catch-up efforts.
Despite the relative rarity of YF and the significant number of infectious and tropical diseases in French Guiana, clinicians should adopt a high index of suspicion for YF, particularly in unvaccinated travelers returning from affected regions.
Daily air travel exchanges with metropolitan France could be the basis for the introduction of the YF virus in Europe. Despite the low epidemic risk in temperate countries, the local cycle of YF transmissions in regions where competent Aedes albopictus populations are established becomes a plausible scenario [36]. The latest YF case, confirmed in a Swiss man who supposedly arrived in French Guiana by land in April 2018, illustrates this hypothesis. If the clinical symptoms of the patient had not developed while he was still in French Guiana, this could have led to an imported case in Europe and consequently to the occurrence of secondary cases, underlining the need for continued vigilance with respect to YF.
|
10.1371/journal.pgen.1002447 | DNA Methylation of the Gonadal Aromatase (cyp19a) Promoter Is Involved in Temperature-Dependent Sex Ratio Shifts in the European Sea Bass | Sex ratio shifts in response to temperature are common in fish and reptiles. However, the mechanism linking temperature during early development and sex ratios has remained elusive. We show in the European sea bass (sb), a fish in which temperature effects on sex ratios are maximal before the gonads form, that juvenile males have double the DNA methylation levels of females in the promoter of gonadal aromatase (cyp19a), the enzyme that converts androgens into estrogens. Exposure to high temperature increased the cyp19a promoter methylation levels of females, indicating that induced-masculinization involves DNA methylation-mediated control of aromatase gene expression, with an observed inverse relationship between methylation levels and expression. Although different CpGs within the sb cyp19a promoter exhibited different sensitivity to temperature, we show that the increased methylation of the sb cyp19a promoter, which occurs in the gonads but not in the brain, is not a generalized effect of temperature. Importantly, these effects were also observed in sexually undifferentiated fish and were not altered by estrogen treatment. Thus, methylation of the sb cyp19a promoter is the cause of the lower expression of cyp19a in temperature-masculinized fish. In vitro, induced methylation of the sb cyp19a promoter suppressed the ability of SF-1 and Foxl2 to stimulate transcription. Finally, a CpG differentially methylated by temperature and adjacent to a Sox transcription factor binding site is conserved across species. Thus, DNA methylation of the aromatase promoter may be an essential component of the long-sought-after mechanism connecting environmental temperature and sex ratios in vertebrate species with temperature-dependent sex determination.
| Temperature changes during early embryonic and/or larval stages are able to modify sex ratios in fish and reptiles. However, the underlying mechanism by which temperature is able to modify the molecular pathways that developing gonads follow to become ovaries or testes is still unknown. One of the most interesting questions raised from previous studies with our model species, the European sea bass, was how temperature could affect the developmental fate of the gonads at a time when they were not even formed in the most rudimentary manner. This was the telltale sign of an epigenetic mechanism. In this study, DNA methylation levels of the aromatase promoter were analyzed in European sea bass exposed to different temperatures during early developmental stages. Aromatase is the enzyme that converts androgens (male hormones) into estrogens (female hormones), which are essential for ovarian development in all non-mammalian vertebrates. We show that increased temperature during a critical period in early development is able to increase DNA methylation of the aromatase promoter, preventing aromatase gene expression. We conclude that gonadal aromatase promoter methylation is most likely part of the long-sought-after mechanism connecting temperature and environmental sex determination in vertebrates.
| The sex ratio is a crucial demographic parameter important for population viability that is established by the processes of sex determination and differentiation. The sex determination mechanisms in vertebrates include genotypic sex determination (GSD), temperature-dependent sex determination (TSD) or a combination of both. In TSD, the temperature experienced during a particular time during early development, referred to as the thermosensitive period (TSP), irreversibly determines gonadal sex. TSD is well established in reptiles and fish [1]. Regardless of the sex determining system, in non-mammalian vertebrates the androgen-to-estrogen ratio determines whether a sexually undifferentiated gonad sexually differentiates into a testis or ovary. This sex steroid ratio depends of the activity of the enzyme aromatase, Cyp19a, the product of the cyp19a gene, which irreversibly converts androgens into estrogens. Further, in ectothermic vertebrates, the effects of environmental temperature on sex ratios are mediated by changes in cyp19a expression. Thus, in reptiles with TSD, exposure to female-promoting temperatures is invariably associated with gonadal cyp19a upregulation, whereas exposure to male-producing temperatures is associated with cyp19a suppression [2], [3]. In all fish species analyzed so far, more males are produced with increasing temperatures [4]. The masculinizing effects of high temperature are also invariably caused by an inhibition of cyp19a expression and enzymatic activity [5]–[7]. Thus, regardless of the animal group and the sex determining mechanism considered, cyp19a regulation is a key player in the sex ratio response to temperature in vertebrates. Unfortunately, the molecular mechanism by which temperature affects cyp19a has remained elusive [1], [8], and this is most important since identifying environmental cues and their perception and transduction mechanisms is a central focus of eco-devo research [9].
Gorelick [10] hypothesized that different methylation patterns of virtually identical sex chromosomes in species with TSD could be altered by small environmental changes, hence determining the sex of individuals. He also proposed that sex differences are initially determined by different patterns of methylation on nuclear DNA of females and males. Recently, reviewing the evidence gathered so far on DNA methylation of four steroidogenic enzymes, it has been postulated [11] that epigenetics are the missing link between genetics, the environment and endocrine functions. Furthermore, other recent studies [12] have shown epigenetic regulation not only of the enzymes involved in the steroidogenic pathway but also of some transcription factors and nuclear receptors related to steroid biosynthesis and action. In mammals, Cyp19 is expressed in a tissue-specific manner and regulated by different tissue-specific promoters [13]. Recently, epigenetic regulation of Cyp19 gene expression in mammals has been demonstrated in humans [14], cattle and sheep [15], [16], and buffalo [17].
The European sea bass (sb), Dicentrarchus labrax, is a gonochoristic GSD+TE species, which means that sex determination can be controlled by both genetic (GSD) and temperature effects (TE). Temperature and genetics contribute approximately equally to sex determination [18]–[20]. Importantly, both GSD+TE and “pure” TSD fish species have exactly the same response to high temperature: inhibition of cyp19a expression [4]. Thus, the sea bass is a perfectly suited model and indeed one of the best documented fish species in terms of sex ratio shifts in response to temperature [20]–[25]. Exposure to high temperatures (>17°C) during the TSP, which covers the period between fertilization to ∼60 days post fertilization (dpf), results in male-biased sex ratios (Figure S1) [20]. During sex differentiation, the primordial gonads, starting from a common primordium, can take two mutually exclusive different developmental pathways towards the formation of an ovary or a testis. Previous studies demonstrated that temperature effects in the European sea bass are more pronounced during the first half of the TSP, when fish are about 30 mm [20]. Interestingly, this not only occurs before morphological sex differentiation takes place (>150 dpf; ∼120 mm fish) but also even before the formation of the gonadal ridges at ∼35 dpf [26]. This demonstrates that the time when the temperature influence takes place is well before the actual differentiation period of the gonads into either the male or the female pathway. For that reason, we hypothesized the existence of an epigenetic mechanism activated by temperature, which could result in different levels of DNA methylation in the gonadal cyp19a promoter, which in turn would affect gene expression, estrogen synthesis and hence sex ratios.
To test our hypothesis, the previously characterized sea bass aromatase (sb cyp19a) promoter [27] was examined and the CpG dinucleotides ∼500 bp upstream of the transcription start site were selected (see Materials and Methods and Figure S2). First, gonadal methylation levels of the sb cyp19a promoter were determined using bisulfite sequencing in one-year-old sea bass males and females (family 1) exposed to two different temperature regimes during the first 60 days of life (high temperature, HT group, and low temperature, LT group; see Materials and Methods and Figure S3). A two-way ANOVA indicated significant differences in average DNA methylation levels of the sb cyp19a promoter in one year old sea bass (∼160 mm length; ∼73 g weight) according to sex (F = 118.2; P = 0.001) and temperature treatment (F = 14.6; P = 0.000), but without interaction between the two variables (P = 0.703). Results showed that, overall, males had twice as much sb cyp19a promoter DNA methylation levels as females (mean ± S.E.M.: 81.15±2.54% vs. 45.5±3.47%; two-tailed Student's t-test; t = −9.591, P = 0.000; Figure S4). Furthermore, sex-related differences were also clearly evident by distinct frequency distributions, with values in the range 12.9–72.8% for females and 71.4–97.1% for males, with a threshold value of 67% (Figure S4). The most important finding, however, was that exposure to high temperature increased gonadal sb cyp19a methylation levels in females from 37.1±3.45 to 53.9±3.49% (two-tailed t-test; t = 3.186, P = 0.005), and from 77.0±1.81 to 85.3±3.28% in males (two-tailed t-test, t = 2.056, P = 0.062; Figure 1).
With seven CpGs analyzed in the sb cyp19a promoter, the maximum number of possible methylation patterns was 128, i.e., 27. Of these, 58 were observed, distributed with different absolute frequencies according to treatment (Figure S5). Based on the observed frequencies of each treatment, contingency analysis was applied to determine if there was any influence of sex and temperature in the distribution of methylation patterns. This, as statistical practice dictates, could be only done for those patterns in which the expected frequency for a given treatment was 5 or higher, which, with four groups means whose observed frequency was 20 or higher. This condition was satisfied by four patterns (Figure 2). These methylation patterns include one in which all seven positions were methylated, another in which all seven positions were unmethylated and two in which either the first or the last position was unmethylated and the remaining were methylated. Analysis of the presence of these four patterns among the four treatment groups showed highly significant differences (P<0.001) in three of them (Figure 2). This reinforces the observation that LT females had the lowest cyp19a promoter methylation levels since they have the highest frequency of the pattern with all positions unmethylated and, conversely, the patterns consisting of all positions methylated was most frequently present in the HT males (Figure 2). Technical error due to PCR bias was avoided since the number of mean methylation patterns was maintained through the different treatments in the range 5.6–7.1 out of a theoretically maximum range of 1–10. This indicates that the PCR reaction was able to amplify different methylation patterns or epialleles. (Figure S6A). Further still, the number of different methylation patterns was independent of the average cyp19a methylation levels (Figure S6B).
To examine the origin of these sex- and temperature-related differences, DNA methylation levels of the gonadal sb cyp19a promoter were assessed in much younger fish (94.8±0.08 mm) of an unrelated batch (family 2), in which biopsy confirmed that they were not sexually differentiated (sex differentiation in the European sea bass starts when fish are in the range ∼80–120 mm). The same temperature treatments were applied as explained above (high temperature, HT group and low temperature, LT group; see Materials and Methods and Figure S3). Since phenotypic sex was unknown in those individuals, a two-step unrestricted cluster analysis of cyp19a mRNA levels was used to classify fish as presumptive future males (low cyp19a mRNA levels) and presumptive future females (high cyp19a mRNA levels) within each temperature treatment (Figure 3), a procedure we had previously demonstrated to be reliable [21]. A two-way ANOVA analysis indicated significant differences in average sb cyp19a gene expression according to sex (F = 110.1; P = 0.000) but not to temperature treatment (F = 1.2; P = 0.277), but with interaction between the two variables (P = 0.013). Two-tailed t-tests also showed differences between cyp19a expression levels (RQ) of LT males vs. LT females (t = 8.427; P = 0.000), HT males vs. HT females (t = 6.251; P = 0.000), HT females vs. LT females (t = −2.516; P = 0.024), but no differences between the RQ values of HT males vs. LT males (P = 0.243) (Figure 3). Furthermore, methylation levels were also examined in 12 individuals who happened to be classified as presumptive females by the cluster analysis. In the HT group, cyp19a promoter DNA methylation levels were 79.0±3.43% (mean ± SEM, n = 3), with a coefficient of variation of 7.5% and a range of 77.2–85.7%. In the LT group, sb cyp19a promoter DNA methylation levels were 80.6±2.55% (two-tailed t-test, t = −0.373, P = 0.717), with a coefficient of variation of 9.5%. However, in one of the nine fish examined the value of cyp19a promoter methylation was 61.9%, i.e., below the threshold level of 67% calculated with the 95% confidence interval of DNA methylation levels observed in adult males vs. adult females (Figure S4), further suggesting that this fish was most likely a female.
In the brain, where cyp19a is only basally expressed, sb cyp19a promoter average methylation levels in one-year-old sea bass (family 1) were 85.1±0.91%, regardless of sex and/or temperature treatment. A two-way ANOVA showed lack of differences associated either with sex (F = 0.746; P = 0.405) or temperature (F = 0.013; P = 0.910) without significant interaction between the two factors (F = 0.307; P = 0.590) (Figure 4A). Furthermore, a similar analysis of DNA methylation of the promoter of the housekeeping gene β-actin in gonads of the same fish showed lack of differences associated either with sex (F = 0.191; P = 0.670) or temperature (F = 3.474; P = 0.087) without significant interaction between the two factors (F = 13.856; P = 0.603) (Figure 4B). Likewise, DNA methylation of the promoter of β-actin in the brain showed also lack of differences associated either with sex (F = 3.945; P = 0.068) or temperature (F = 0.038; P = 0.848) without significant interaction between the two factors (F = 0.029; P = 0.867) (Figure 4C). On average, gonad and brain β-actin promoter methylation levels (± SEM) were 13.8±2.76% and 19.63±4.21%, respectively.
When histologically sexed at about one year of age, the percentage of females in the HT group (family 1) was 56.0±11.3% (n = 40), a 15% decrease when compared to the 71.0±3.5% (n = 40) value of the LT group (Chi square = 7.28; P = 0.01). On the other hand, cyp19a mRNA expression levels in HT females was significantly lower than in LT females (t-test; F = 0.024; P = 0.003; Figure 5A). In addition, there was a significant inverse relationship between cyp19a expression and methylation levels in these one-year old females (r2 = 0.29; F = 7.84; P = 0.01) (Figure 5B).
We also conducted experiments to determine the possible influences of male vs. female differentiation pathways in the DNA methylation levels of the gonadal sb cyp19a promoter. When fish from a batch with a natural low incidence of females at LT (family 3) was treated with Estradiol-17ß (E2) the number of females increased from 2.5 to 90% (P<0.01). However, DNA methylation levels of the sb cyp19a promoter in the E2-treated female gonads were not statistically different from those of untreated LT females (mean ± S.E.M.: 32.9±6.66% vs. 41.2±5.13%; two-tailed t-test; t = 0.968, P = 0.356; Figure 6).
Analysis of Molecular Variance (AMOVA) was used to determine if the different CpGs positions found in the sb cyp19a promoter were differentially methylated between male and female gonads or between the gonads of animals of the same sex but reared at different temperatures. Results revealed that particular CpGs were differentially methylated according to sex and/or temperature (Figure 5C and 5D). Significant differences in all positions were detected between LT males and LT females (AMOVA, Fst = 0.361, P = 0.000) (Table 1). However, only CpG positions −431 and −13 showed significant differences (AMOVA, Fst = 0.052, P = 0.014) when LT females were compared with HT females, with position −13 presenting the highest significant differences (Fst = 0.083, P = 0.031 and Fst = 0.148, P = 0.005, respectively; Figure 5C and Table 1).
In a previous study [27], we characterized the sb cyp19a promoter using bioinformatic tools (MatInspector) and gel shift assays and identified putative transcription binding sites. From that study it was found that the CpG in position −431 of this promoter is located near a putative binding site for a transcription factor of the Fox family and that the CpG in position −13 is located near a putative binding site for a Sox family transcription factor (Figure S2). Furthermore, previous studies with other fish species have shown that co-transfection of sb cyp19a promoter constructs with either Foxl2, SF-1 or simultaneous co-transfection with both of these potent transcriptional activators of cyp19a significantly increased luciferase activity [28], [29]. Based on these previous findings, the sb cyp19a promoter activation was determined under methylated and control conditions by a luciferase reporter assay. As expected, SF-1 and Foxl2 were each capable of activating sb cyp19a expression in vitro. When combined, still higher expression was observed (Figure 7). Remarkably, induced hypermethylation of the sb cyp19a promoter completely suppressed promoter transcription stimulation in vitro by Foxl2 and SF-1 alone (two tailed t- test, P<0.01) or in combination (two tailed t- test, P<0.05).
To determine whether the number and position of CpGs was conserved in other fish species, 600 bp of the sb cyp19a promoter, including all CpG analyzed in the present study and 94 bp within the opening reading frame, were aligned with the promoter region of both phylogenetically-related and unrelated species. In addition and based on the previously characterization of the sb cyp19a promoter by Galay-Burgos and collaborators [27] some putative transcription binding sites for other species were identified based on sequence similarity with the sb cyp19a promoter. Overall sequence similarity was 53%, with a clear conservation of some of the transcription factor binding sites, including the TATA box (Figure 8). The CpG located at position −13 was the most conserved, with three species having a CpG dinucleotide at this position. On the other hand, position −431 was not found in the other species analyzed although two of them had a CpG position situated around 40–60 bp away.
The present study clearly shows that the DNA methylation levels of the sb cyp19a promoter were twice as high in the gonads of males as compared to females in one-year-old European sea bass. The origin of these sex-related differences is at present unknown. Throughout development, cell- and tissue-specific methylation patterns are the result of de novo methylation, maintenance of the existing pattern and demethylation processes [30]. It has been suggested that DNA methylation suppresses recombination, allowing Muller's ratchet to operate on differentially methylated sexes [10]. Differential methylation can also suppress transcription, contributing to accentuate sex-related differences [10]. Thus, sex differentiation of the gonads could be regulated through DNA methylation of key genes in a manner similar to that seen in other tissues. SRY, the sex-determining gene in mammals, is normally epigenetically silenced, activated during a specific window in development, and then silenced again [31]. The machinery needed for these changes is also present in fish, where up to eight different DNA methyltransferases (Dnmts), the enzymes responsible to transfer methyl groups to DNA, have been identified so far [32]. In medaka (Oryzias latipes) and Xiphophorus embryos, dnmt1 expression is spatially- and temporally-regulated, suggesting that it may play an important role during development. Changes in methylation levels of cyp19a and other gene promoters are then likely to be involved in the course of sexual differentiation in fish.
To elucidate the possible influences of male vs. female differentiation pathways, we measured DNA methylation levels of the sb cyp19a promoter in gonads of E2-treated vs. control females (both reared at LT to avoid confounding effects of masculinization by HT). No differences were found, indicating that E2 in this model does not affect gonadal sb cyp19a promoter DNA methylation levels. This is consistent with previous studies showing that cyp19a expression levels in gonads of both untreated and E2-treated females were similar [33]. Also, treatment of medaka with estrogen did not result in changes cyp19a methylation in the gonads [34]. In the present study, the lack of effects of estrogen treatment on gonadal sb cyp19a promoter methylation levels supports the idea that increased methylation is the cause of lower cyp19a expression and not the other way around.
Temperature effects on the methylation of certain promoters are well established in plants [35]. There is also evidence of environmental influences on phenotypic plasticity in animals mediated by epigenetic mechanisms [36]. In honeybees, nutrition load gives rise to two different adult female phenotypes: the fertile queens and the sterile workers [37]. Recently, it has been shown that the active ingredient in royal jelly involves an epidermal growth factor receptor-mediated signaling pathway, resulting in increased body size, ovarian development and shortened developmental time [38]. Interestingly, in many fishes including the sea bass, there is also an association between larger size and female development during sex differentiation [39], [40].
The European sea bass has a polygenic system of sex determination [18] with a strong environmental influence [4]. Sex ratios depend upon the broodstock used and the temperature experienced during the TSP [20]–[25]. Because of its sex determining mechanism, sea bass monosex populations are not available. To determine the influence of temperature in the methylation levels of the gonadal sb cyp19a promoter, we had deliberately chosen a family giving a high percentage of females when reared at the permissive LT (family 1). This allowed obtaining sufficient females even after rearing at masculinizing HT. The LT or control group had 71±3.5% females, whereas the HT group had 56±11.3% females, in agreement with the observation that rearing at HT during the TSP masculinizes ∼50% of the genotypic females into phenotypic males without negative effects on survival [20].
The main finding of this study is that exposure to high temperature during early development increased ∼1.5 times sb cyp19a promoter DNA methylation levels as evidenced in one-year-old female gonads. In contrast, HT did not significantly increase sb cyp19a promoter methylation levels in the gonads of males, which were already quite high (∼80%) in the LT group. In the present study, also a relationship between increased methylation and decreased female numbers was found since the reduction of 15% females in the HT group closely matched the 16.8% increase in sb cyp19a promoter methylation levels in the same group. Importantly, cyp19a expression was significantly lower in HT females. Thus, the present study demonstrates that high temperatures experienced during early life masculinize by increasing sb cyp19a promoter methylation levels and consequently decreasing cyp19a expression in the gonads. It could be argued that differential methylation of the sb cyp19a promoter could be part of the polygenic mode of sex determination by parental imprinting mechanisms. However, the existence of differences in methylation levels of the sb cyp19a promoter in the gonads of LT and HT females (originated from the same parents) suggests that high temperatures is able to override any possible parental imprinting. This indicates that the male-biased sex ratios found in stocks reared at HT is a temperature effect overriding the basic polygenic sex determination system.
Since genes that are not in use may become methylated, it could be argued that cyp19a suppression is the consequence rather than the cause of the suppression of female development, i.e., that methylation of the sb cyp19a promoter can be due to the initiation of the male pathway. Our results with sexually undifferentiated animals show differences on gonadal cyp19a gene expression levels between presumptive males and females, in agreement with the observation that differences in gonadal cyp19a gene expression levels can be detected prior to sex differentiation [41]. Methylation levels of the gonadal sb cyp19a promoter on presumptive females reared at LT or HT ranged from 61.9 to 83.7%. Interestingly, these values are closer to those of one-year-old males (range 71.4–97.1%) than those of one-year-old females (range 12.9–72.8%). This suggests that the methylation levels in presumptive females may reflect low gonadal cyp19a gene expression levels at that time of development. Further, although differences between putative females and males can be detected prior to sex differentiation [41], upregulation of cyp19a gene expression only occurs during female sex differentiation [39]. Together, these observations suggest that during fish sex differentiation DNA demethylation of the sb cyp19a promoter in the females could be required to enable cyp19a upregulation in their gonads. During mammalian gonadal development, genome-wide demethylation occurs during primordial germ cell migration in early development as observed in mice and pigs [42]–[44]. We hypothesize that high temperature during early development either immediately and irreversibly hypermethylates the cyp19a promoter or inhibits its demethylation later during sex differentiation. Further studies are needed to discern between these two possibilities.
We also checked whether the sex-and temperature-dependent changes on sb cyp19a promoter methylation in the gonads were promoter- and tissue-specific. It is known that the promoters of non-expressed genes are hypermethylated. The sb cyp19a promoter in the brain was consistently hypermethylated independent of sex and temperature, in accordance with the well-established fact that cyp19a is mainly expressed in gonads and only basally expressed in the fish brain [45]. In contrast, the β-actin promoter was hypomethylated both in brain and gonads and regardless of sex and temperature, reflecting the constitutively high expression of this housekeeping gene. Together, these results suggest that in the European sea bass sex- and temperature-dependent changes in DNA methylation of the cyp19a promoter are restricted to the gonads and are not a generalized effect of temperature, although other gene promoters could also be affected.
The mechanism by which temperature changes DNA methylation levels of the sb cyp19a promoter in the gonads is not known. In Xiphophorus, O6-methylguanine-dnmt activity is optimal at 23°C but is lost below 15°C [46]. Thus, it remains to be determined if temperature activation of the Dnmts is responsible for the resulting higher cyp19a promoter methylation seen at HT. Since cyp19a expression is essential for female differentiation in all non-mammalian vertebrates, resolving the link between temperature and cyp19a promoter methylation is a relevant issue. The results presented here, although mostly descriptive, represent a first step in this direction.
Foxl2 is one of the most potent transcriptional regulators of cyp19a in vertebrates, from fish to mammals [28], [47]; however, it appears that Foxl2 works best with the involvement of other cofactors. For example, co-transfection of cyp19a promoter constructs with either SF-1 or Foxl2 significantly increased luciferase activity [28], but simultaneous co-transfection of SF-1 and Foxl2 had an additive effect on cyp19a promoter activation [29]. The present study shows, as expected, that SF-1 and Foxl2 were each capable of activating cyp19a expression in HEK293T cultured cells, and when combined still higher expression was observed. Remarkably, induced hypermethylation of the sb cyp19a promoter completely suppressed both Foxl2- and/or SF-1-stimulated transcription in vitro. This strongly suggests that hypermethylation of the sb cyp19a promoter prevents binding of at least Foxl2 and SF-1 to their respective sites, thus blocking cyp19a transcriptional activation.
Sex- and tissue-specific differences on cyp19a methylation levels have also been found in the model fish medaka, a GSD species, where five CpGs located within a region of ∼300 bp of the cyp19a promoter were methylated mostly in testis and female brain, but unmethylated in ovary and male brain [34]. Further, in cattle and sheep cyp19a promoter regions P1.1, P1.5 and P2 were differentially methylated depending on sex and tissue [15]. Together, these results suggest that sex-related differences in DNA methylation levels of the cyp19a promoter maybe a generalized phenomenon present from fish to mammals. On the other hand, it has been suggested that the molecular mechanisms underlying sex ratio responses to temperature must be conserved throughout vertebrates [1], [8], [48]. Our results revealed that some of the transcription factors binding sites in the cyp19a promoter are conserved across species, being the CpG dinucleotide located at position −13 the most conserved one. Together, these observations suggest that the epigenetic mechanism described here may indeed be present at least in other fish species. Further, this study also shows that some sb cyp19a promoter CpG positions are differentially methylated by temperature, indicating that they can be important for cyp19a transcription. This is also the case of position −13, near a TATA box and Sox binding site, or position −431, which is close to a Fox binding site. Differential DNA methylation of gene promoters has been demonstrated to account for tissue-specific gene transcription through transcription factor binding inhibition [49]–[51]. Together the results presented here suggest that differential methylation of specific CpGs positions could significantly contribute to a transcription regulation of the sb cyp19a gene.
In summary, to our knowledge, this is the first report describing an epigenetic mechanism mediating temperature effects on sex ratios in any animal. Our data show that the sb cyp19a promoter is significantly more methylated in males than females, which is in agreement with the well-established constitutively lower levels of cyp19a expression in testes as compared to ovaries, and with the higher levels of estrogens seen in females [6]. Further, luciferase assays confirmed that DNA methylation of the sb cyp19a promoter represses transcription in vitro. Together, these results indicate that hypomethylation of the cyp19a promoter is required for normal ovarian development. Further, since estrogens are essential for ovarian differentiation, our results suggest that sb cyp19a promoter hypermethylation contributes to cyp19a transcription silencing during male differentiation, preventing the transformation of an undifferentiated gonad into an ovary. More importantly, however, we show that in a fish species where sex determination depends on the interaction between genotype and environment, exposure to abnormally high temperatures during the TSP is able to induce hypermethylation of the sb cyp19a promoter of females past a certain threshold, approaching the values characteristic of males (Figure 9). The result is that genotypic females—or, more properly in a mixed genic and environmental sex determination system, fish in which the sum of factors promoting female development is stronger than the sum of factors promoting male development— differentiate into phenotypic males, altering population sex ratios, as observed in many fish species exposed to high temperatures [4], [6]. As shown by Ospina-Álvarez and Piferrer [4], and in contrast to reptiles [1], fish studied so far have only one sex ratio response pattern to temperature: male-biased sex ratios with increased temperatures. This makes still more appealing the methylation hypothesis because the underlying mechanism could apply to all fish species with temperature influences on sex ratios. Importantly, it has been suggested that in species with GSD such as mammals, where sex determination depends on the inheritance of the sex-determining gene SRY, sex is a threshold dichotomy mimicking a single gene effect [52]. Our results indicate that such a threshold dichotomy can also apply to a completely distinct scenario: gonadal cyp19a promoter methylation levels of males vs. females, implying one shared feature of the two major sex determining mechanisms of vertebrates, GSD and TSD. Thus, temperature may modulate sex ratios through changes in the proportion of animals whose cyp19a promoter methylation level falls above or below a certain threshold. The present results demonstrate that in the European sea bass an epigenetic mechanism can affect an essential biological process, with consequences in resulting population sex ratios. Although more studies are certainly needed to confirm this, it is tempting to suggest that the epigenetic scheme described herein may be an essential component of the long-sought-after mechanism connecting environmental temperature and sex ratios in species with TSD.
Freshly fertilized European sea bass (Dicentrarchus labrax L.) eggs were collected from the Institute of Aquaculture (Castelló, Spain) and from a private farm (Base Viva, St. Pere Pescador, Spain) and transported to our experimental aquarium facilities at the Institute of Marine Sciences in Barcelona. Egg incubation and rearing during the larval and juvenile stages were performed according to standard sea bass rearing practices [53], except for the temperature treatments (see next section), and are explained in detail elsewhere [20]. Once animals reached mid-metamorphosis (standard length; SL>18 mm), juveniles were reared in 650 l fiberglass tanks under simulated natural photoperiod and fed ad libitum with pelleted food of the appropriate size.
At 330 dpf, i.e., long past the thermal regimes, fish were sacrificed and gonadal samples were collected (n = 40 fish per treatment). From each fish (∼159 mm and ∼73 g), one gonad was processed for histological identification of sex. Gonads were fixed in 4% paraformaldehyde in PBS, embedded in paraffin, cut at 7 µm thickness and stained with haematoxylin-eosin. The other gonad was snap-frozen in liquid nitrogen and stored at −80°C until further analysis to determine methylation levels and gene expression. For sexually undifferentiated fish (family 2; 94.8±0.08 mm), 20 fish per group were collected from the LT and HT treatment groups. Due to tissue amount limitations, the right gonad was used to obtain DNA and the left gonad to obtain RNA. Sex was identified based on cyp19a mRNA levels as described in Blázquez and collaborators [41] and in the Statistical analysis section below.
The gonadal sb cyp19a [27] and β-actin promoters [54] were examined to identify CpG dinucleotides that could be differentially methylated. For the sb cyp19a promoter (Figure S2), genomic DNA was obtained in the case of sexually differentiated fish from the gonads of 8–15 fish or the brains of 3–5 fish, depending on temperature and sex. For the β-actin promoter, brains and gonads from 3–5 animals were used depending also on temperature and sex. In the case of sexually undifferentiated fish, a total of 12 animals were used for sb cyp19a promoter methylation analysis. The DNA samples from each animal were individually processed and subjected to sodium bisulphite-mediated sequencing as described by Widschwendter and collaborators (2000) [55]. The targeted portion of the promoter was amplified from the bisulphite-modified DNA with two rounds of PCR by use of nested primers specific to the bisulphite-modified sequence of this region. For the sb cyp19a promoter the primers were as follows: External Forward, ATTGGTAGTTTAATGGAGGAATTT; External Reverse, AATCCCACTACAATAACATTTAAAAAC; Nested Forward, GAGGAATTTGGGAGGAATTATAAATAT; Nested Reversed, CCAAATCTACCACTATAATATCCAAAC. The primers for the β-actin promoter were as follows: External Forward, AATTTATAATTTTGGTTGGTAGTAA; External Reverse,CAAAATCTTACCTTAAAAATATATCTAC; Nested Forward, TATAATTTTGGTTGGTAGTAATTGG; Nested Reverse, CATTCACAAACCTCAACACTAACC. A hot start polymerase (Qiagen) was used in both external and nested PCR. For the sb cyp19a promoter, external PCR consisted in 5 min at 94°C; 5 cycles of 1 min at 94°C, 2 min at 55°C, 3 min at 72°C; 25 cycles of 30 s at 94°C, 2 min at 50°C, 1 min and 30 s at 72°C and a final extension of 7 min at 72°C. Subsequently, nested PCR consisted in 5 min at 94°C; 30 cycles of 30 s at 94°C, 30 s at 56°C, 30 s at 72°C and a final extension of 7 min at 72°C. For the β-actin promoter, external PCR consisted in 5 min at 94°C; 5 cycles of 1 min at 94°C, 2 min at 54°C, 3 min at 68°C; 25 cycles of 30 s at 94°C, 2 min at 50°C, 1 min and 30 s at 68°C and a final extension of 7 min at 72°C. Subsequently, nested PCR consisted in 3 min at 94°C; 30 cycles of 30 s at 94°C, 30 s at 53°C, 30 s at 68°C and a final extension of 7 min at 72°C. PCR products were separated by gel electrophoresis and gel bands were purified by Purelink Quick Gel Extraction (Invitrogen). Gel purified bands were cloned into the pCR4-TOPO vector and transformed into E. coli Topo10 chemically competent cells (Invitrogen) in the case of the sb cyp19a promoter, or into pGEM-T Easy vector (Promega) and transformed into E. coli JM109 competent cells (Invitrogen) in the case of the β-actin promoter. Then 7–10 individual clones per each fish were sequenced each in both directions and used to evaluate the seven (cyp19a) or 25 (β-actin) CpG dinucleotide positions present in the promoter region. Average methylation levels per position and fish were computed. In summary, in this study a total of 76 different animals were used. DNA methylation levels were determined, always on an individual fish basis (no pools were used), only in the gonads in some fish, whereas in others it was determined both in the gonads and also in the brain, and, as stated above, 7–10 clones per fish were sequenced in both directions. The total amount of sequenced clones in this study was ∼1100.
The efficiency of the bisulfite conversion step was evaluated by using the Bisulfite sequencing Data Presentation and Compilation (BDPC) online software (available at: http://biochem.jacobs-university.de/BDPC/index.php) [56]. The conversion rates of C, which are not in the context of a CpG, were determined in different number of clones and in all tested tissues. For the sb cyp19a promoter, the mean percentage of converted Cs was 97.98±0.9% (calculated in a subset of 35 PCR reactions) and for β-actin promoter, the mean percentage was 97.68±0.41% (calculated from a subset of 71 PCR reactions).
Since ten clones per fish were used to determine sb cyp19a promoter methylation, the maximum number of different methylation patterns (or epialleles) per fish that could be observed was 10, i.e., each clone has a different methylation pattern. Thus, for each fish from each one of the four groups the actual observed number of different cyp19a promoter methylation patterns was determined in order to check for possible PCR bias.
Total RNA was isolated from ovaries of 8 females from the HT group and 14 females from the LT group from sexually differentiated one-year-old sea bass, and 16 gonads from the HT and 16 from the LT group in the case of sexually undifferentiated or differentiating animals. Gonad tissue was homogenized with 0.5 ml of trizol and total RNA was extracted with chlorophorm, precipitated with isopropanol and washed with 75% ethanol. Pellets were suspended in 25 µl DEPC-water and stored at −80°C. One microgram of total RNA was reverse transcribed into cDNA using Superscript II (Invitrogen) and 250 ng of random hexamer primers (pdN6) following the manufacturer's instructions. Real-time PCR reactions were carried out to determine sb cyp19a gene expression. The endogenous reference gene used was 18S (validated in a previous study by our laboratory [57]). The real time RT-PCR reaction was carried out with the SYBR Green chemistry (Power SYBR Green PCR Master Mix; Applied Biosystems). The primers were: ovaro RT-F1: AGACAGCAGCCCAGGAGTTG and ovaro RT-R1: TGCAGTGAAGTTGATGTCCAGTT. PCR reactions contained 1X SYBR green master mix (Applied Biosystems), 10 pmol of each primer and 1 µl of the RT reaction. Samples were run in duplicate in optically clear 384-well plates. Cycling parameters were: 50°C for 2 min, 95°C for 10 min, followed by 40 cycles of 95°C for 15 s and 60°C for 1 min. Finally, a temperature-determining dissociation step was performed at 95°C for 15 s, 60°C for 15 s and 95°C for 15 s at the end of the amplification phase. Real-time RT-PCR data were collected by SDS 2.3 and RQ Manager 1.2 software and relative quantity (RQ) values were estimated for each reaction replicate. Specifically, the female with the lowest level of aromatase expression (i.e., highest ΔCt) was assigned as the calibration sample to calculate ΔΔCt and RQ values.
Data reported as proportions (sex ratios and methylation levels) were always arcsin square root transformed before any statistical analysis. Likewise, all RQ expression data were log-transformed to ensure normality.
A two-way ANOVA was used to investigate differences in sb cyp19a or β-actin promoter DNA methylation levels (dependent variable) considering sex and temperature as the two independent factors. Post hoc multiple comparisons were carried out with Tukey's multiple range test with the Statgraphics v16 or SPSS v19 software.
One-year-old fish were histologically sexed. Younger, sexually-undifferentiated fish were sexed based on a previous study from our laboratory that demonstrated that mRNA levels of cyp19a can be used as an early marker of phenotypic sex in the European sea bass [41]. In this case, a two-step cluster analysis (SPSS) was used to classify individuals as presumptive females and males based on the gonadal cyp19a mRNA levels (RQ). Afterwards, a two-way ANOVA was used to investigate differences between temperature and phenotypic sex as explained above. In addition, a two-tailed Student's t-test was used to check for differences in RQ between the following pairs: presumptive males and females at LT; presumptive males and females at HT; presumptive females at LT and HT; and presumptive males at LT and HT. Two-tailed Student's t-test was also used to analyze differential cyp19a expression levels among one-year old females of each temperature treatment and also to detect differences between pGL3-cyp19a methylated and unmethylated promoter vectors in the transfection assay.
To check for differences in methylation levels in specific CpG positions of the sb cyp19a promoter, a hierarchical population analysis was carried out. First, sequences were trimmed to seven nucleotides in length, one corresponding to each CpG analyzed, with two possible variants for each nucleotide: C if methylated and T if unmethylated. Then, all sequences from the same individual were considered as one population, with size equivalent to the number of sequences analyzed for each individual [59]. The four classes (i.e., treatments, LT and HT males, and LT and HT females) were considered as groups of populations. The hierarchical analysis of the molecular variance, AMOVA [59], was used to test for possible genetic differentiation among classes. When less than 5% of the 10000 pseudo-replicates presented higher genetic variance than one estimated by chance, then the genetic structure was considered not significant. AMOVA also calculated the correlation measure fixation index of population differentiation (Fst). In all cases differences were considered statistically different when P<0.05.
|
10.1371/journal.pcbi.1004638 | Somato-dendritic Synaptic Plasticity and Error-backpropagation in Active Dendrites | In the last decade dendrites of cortical neurons have been shown to nonlinearly combine synaptic inputs by evoking local dendritic spikes. It has been suggested that these nonlinearities raise the computational power of a single neuron, making it comparable to a 2-layer network of point neurons. But how these nonlinearities can be incorporated into the synaptic plasticity to optimally support learning remains unclear. We present a theoretically derived synaptic plasticity rule for supervised and reinforcement learning that depends on the timing of the presynaptic, the dendritic and the postsynaptic spikes. For supervised learning, the rule can be seen as a biological version of the classical error-backpropagation algorithm applied to the dendritic case. When modulated by a delayed reward signal, the same plasticity is shown to maximize the expected reward in reinforcement learning for various coding scenarios. Our framework makes specific experimental predictions and highlights the unique advantage of active dendrites for implementing powerful synaptic plasticity rules that have access to downstream information via backpropagation of action potentials.
| Error-backpropagation is a successful algorithm for supervised learning in neural networks. Whether and how this technical algorithm is implemented in cortical structures, however, remains elusive. Here we show that this algorithm may be implemented within a single neuron equipped with nonlinear dendritic processing. An error expressed as mismatch between somatic firing and membrane potential may be backpropagated to the active dendritic branches where it modulates synaptic plasticity. This changes the classical view that learning in the brain is realized by rewiring simple processing units as formalized by the neural network theory. Instead, these processing units can themselves learn to implement much more complex input-output functions as previously thought. While the original algorithm only considered firing rates, the biological implementation enables learning for both a firing rate and a spike-timing code. Moreover, when modulated by a reward signal, the synaptic plasticity rule maximizes the expected reward in a reinforcement learning framework.
| One of the fascinating and still enigmatic aspects of cortical organization is the widespread dendritic arborization of neurons. These dendrites have been shown to generate dendritic spikes [1–3] that support local dendritic processing [4–7], but the nature of this computation remains elusive. An interesting view is that the dendritic nonlinearities endow the neuron with the structure of a 2-layer neural network of point neurons, in particular if the dendrites show themselves step-like dendritic spikes, but also if the dendritic nonlinearities remain continuous [8–11]. Here we show that the dendritic morphology actually offers a substantial additional benefit over the 2-layer network. This is because it allows for the implementation of powerful learning algorithms that rely on the backpropagation of the somatic information along the dendrite that, in a network of point neurons, would not be possible in this form.
Error-backpropagation has become the classical algorithm for adapting the connection strengths in artificial neural networks [12, 13]. In this algorithm, an error at an output unit is assessed by comparing the self-generated activity with a target activity. Plasticity in hidden units is driven by this error that propagates backwards along the connections of the network. Synapses, however, transmit information just in one direction, making it difficult to implement error-backpropagation in biological neuronal circuitries. But this is different for dendritic trees. In the 2-layer structure of a dendritic tree information at the output site may be physically backpropagated across the intermediate computational layer to the synapses targeting the tree.
While the suggested dendritic error-backpropagation is a plasticity rule for supervised learning, it is also suitable for reinforcement learning. Instead of imposing the somatic spiking to learn pre-assigned target spike timings, the somatic spikes can be generated by the dendritic inputs alone, while learning is driven by a delayed reward signal. The synapse itself can be agnostic about the coding and the learning scenario; it learns by continuously adapting synaptic strength according to molecular mechanisms that are identical in the different scenarios.
Various experimental work revealed that synaptic plasticity depends on the precise timing between pre- and postsynaptic action potentials [14, 15] and the postsynaptic voltage [16]. It has further been shown that the specific form of this spike-timing-dependent plasticity (STDP) may vary with the synaptic location on the dendritic tree [17–19], and that synaptic plasticity in general is modulated by dendritic spikes [20–22]. Yet, no coherent view on the impact of dendritic nonlinearities on plasticity has emerged. Correspondingly, beside an early attempt to assign a fitness score to dendritic synapses [23] and the suggestion of a Hebbian-type plasticity rule for synapses on active dendrites [24], no computational framework for synaptic plasticity with regenerative dendritic events exists that would guide its experimental exploration. In our previous study, we derived a reward-maximizing plasticity rule that incorporates dendritic spikes, but no online implementation was presented [25]. Here, starting from biophysical properties of NMDA conductances [26], we consider an integrated somato-dendritic spiking model that captures the main biological ingredients of dendritic spikes and that is simple enough to derive an online plasticity rule for different coding schemes in the context of both supervised and reinforcement learning.
We model a multi-compartment neuron with several active dendritic branches, each directly linked to a somatic compartment (Fig 1A1). The subthreshold dendritic voltage in branch d is the weighted sum of normalized postsynaptic potentials (PSPs) triggered by the presynaptic spikes in the afferents projecting to that branch, u d d ( t ) = ∑ i w d i PSP i ( t ). Here, wdi represents the synaptic strength of the synapse from afferent i onto branch d that scales the PSP amplitude. The dendritic branches can generate temporally extended NMDA-spikes of a fixed amplitude, similar to experimental observations in vitro [1, 3, 5] and in vivo [7]. In our model an NMDA-spike is represented by a square voltage pulse of amplitude a and duration Δ = 50 ms (Fig 1A2–1A3). It is stochastically elicited with a rate that is an increasing function of the local subthreshold membrane potential u d d and, implicitly, of the local glutamate level. In fact, in an in vivo scenario the joint voltage and glutamate condition for triggering an NMDA-spike effectively reduces to a single condition on the local voltage alone. This is because the depolarization required to activate the NMDA receptors is only reached when enough glutamate was released, making the glutamate condition automatically satisfied at high enough voltages (see S1 Text).
The subthreshold dendritic voltage u d d ( t ) and the dendritic spike train NMDAd(t) in branch d propagate with some attenuation factor α to the soma where they add up with inputs from other branches to form the somatic voltage u s = ∑ d α ( u d d + NMDA d ) - κ. This voltage is also modulated by a spike reset kernel κ(t) incorporating the transient hyperpolarisation caused by each somatic spike (Fig 1A4). For supervised learning, the somatic spikes S(t) are imposed by an external input, whereas in reinforcement learning they are stochastically triggered with an instantaneous rate ρs(t) that is an increasing function of the somatic potential us (Online Methods).
We first consider a supervised learning scenario where somatic spikes S are enforced by one modality (e.g. a visual stimulus) while the synaptic inputs to the dendritic branches are caused by another modality (e.g. representing an auditory stimulus [27]). The strengths of the synapses on the dendrites, wdi, are adapted in order to reproduce the somatic spike train S(t) from just the dendritic input alone, without direct somatic drive. This can be achieved by ongoing synaptic weight changes, w ˙ d i, that together maximize the likelihood of observing S in response to this dendritic input. According to the two types of contributions to the somatic voltage, the sub- and supra-threshold dendritic voltages, u d d and NMDAd, the synaptic weight change can also be decomposed into a sub- and suprathreshold contribution, w ˙ d i = w ˙ d i ss + w ˙ d i sds, that take into account the subthreshold somato-synaptic (ss) and the suprathreshold somato-dendro-synaptic (sds) drive. We also refer to w ˙ d i as somato-dendritic synaptic plasticity (sdSP).
The somato-synaptic contribution is proportional to the postsynaptic error term (S − ρs) times the local postsynaptic potential PSPi induced by synapse i on that dendritic branch,
w ˙ d i ss ∝ ( S - ρ s ) · PSP i . (1)
This corresponds to the gradient learning rule that was previously derived for a single compartment neuron [28] and that was shown to be consistent with the experimentally observed STDP (see e.g. [29]). The error term in the rule ensures that if the rate ρs is too small for generating S, the weight is increased, and if the rate is too high, the weight is decreased, eventually leading in average to 〈S〉 = ρs.
The main sdSP-effect stems from the somato-dendro-synaptic contribution w ˙ d i sds. The instantaneous synaptic weight change at time t is induced by the dendritic activity Dend in branch d during the interval Δ prior to t. Any NMDA-spike elicited in this interval will affect the somatic voltage at time t, and the likelihood of a dendritic spike is itself influenced by the local synaptic potentials PSPi arriving in this interval and a few milliseconds before (Fig 1A2–1A5). Overall, we obtain an expression of the form
w ˙ d i sds ∝ ( S - ρ ∖ d s ) · Den d ∗ PSP i , (2)
where Dend ∗ PSPi captures the impact of synapse i on the triggering of an NMDA-spike in the preceding interval Δ, and ρ ∖ d s represents the instantaneous somatic firing rate in the absence of a dendritic spike in branch d (see Online Methods). A positive error term ( S - ρ ∖ d s ) tells the synapses on branch d how worth it is to increase their weights in order to trigger a local NMDA-spike; a negative error term suggests to rather decrease the weights since even without NMDA-spike from that branch the somatic firing rate, in average, is too high. When only dendritic nonlinearities without spiking are present, the rule Eq (2) simplifies to a pure 3-factor rule composed of a somatic difference factor, a dendritic factor, and a presynaptic factor that can be applied to dendrites showing supra- or sublinear dendritic summations (see Eq (9) in Online Methods).
The learning rule of Eq (2) can be interpreted as error-backpropagation for spiking neurons where a somatic error signal is propagated back to the dendrites that represent the nonlinear hidden units. These hidden units further modulate the error signal depending on their impact on the output unit. Classical error-backpropagation would also adapt the weights from the hidden units to the output unit. This would correspond to adapting the impact of NMDA spikes on the somatic voltage and could be modeled as dendritic branch strength plasticity [24, 30]. For conceptual clarity we discard from this extension, but the gradient calculations could also be applied to infer an optimal learning rule for these branch strengths.
The overall synaptic modification, Δwdi, induced by sdSP is obtained by integrating the instantaneous changes w ˙ d i over the stimulus duration, Δ w d i = ∫ 0 T w ˙ d i ( t ) d t. Using the decomposition w ˙ d i = w ˙ d i ss + w ˙ d i sds we may also write Δ w d i = Δ w d i ss + Δ w d i sds and have a closer look to the somato-dendro-synaptic contribution Δ w d i sds (Fig 1A2–1A5). We fixed a presynaptic spike at time t i pre = 0 and plotted Δ w d i sds as a function of the somatic spike time ts for the case of a NMDA-spike at td = 5 ms and 15 ms. The dendritic spike immediately after a presynaptic spike considerably extends the classical time window for causal ‘pre-post’ potentiation to a ‘pre-dend-post’ potentiation. In fact, a presynaptic spike that was taking part in triggering a NMDA-spike may indirectly contribute also to a postsynaptic spike more than 50 ms later. In turn, synaptic depression is induced in an a-causal configuration where the somatic spike comes either before the presynaptic spike or after the NMDA-spike has already decayed.
Endowed with sdSP a neuron is able to learn precise output spike-timings as shown in Fig 1B and 1C; blue) where 3 somatic spike times were imposed during the learning. The dendritic input consisted of 100 frozen presynaptic Poisson spike trains with frequency 6Hz and duration T = 500 ms. The dendritic tree had 20 branches, each being targeted by a random subset of the 100 afferents with a connection probability of 0.5. After repeated pattern presentations with somatic output clamped to the target spikes, the neuron learned to generate the target output from the synaptic input alone with a precision of a few milliseconds. The high spike-time precision is lost when synapses are modified only by the somato-synaptic contribution w ˙ d i ss (Fig 1B and 1C; pink). Because this plasticity contribution is blind to dendritic activity, small synaptic weight changes may cause undesired appearance or disappearance of NMDA-spikes. In this case dendritic spikes arise as unpredicted knock-on effects of synaptic plasticity. Note that the somato-synaptic contribution alone, being identical to the gradient rule [28], would be able to learn the precise spiking (as would also the rules in [29, 31]) if the neuron were note endowed with the dendritic spiking mechanism and instead would only show linear summation with passive voltage propagation.
We next considered a reinforcement learning scenario where synaptic modifications are modulated by a binary feedback signal R = ±1 that is applied at the end of the stimulus presentation and that assesses the appropriateness of the somatic firing pattern. While this feedback is itself an external quantity, it is assumed to induce an internal signal, e.g. in the form of a neuromodulator, that globally modulates the previously induced synaptic changes. To control the balance between reward and punishment, the internal feedback modulates the past plasticity induction by a factor (R − R∘) with a constant reward bias R∘.
When deriving a plasticity rule that maximizes the expected reward we again obtain the same sdSP (see Eqs (1) and (2)), but now integrated across the stimulus interval and then modulated by the feedback signal,
Δ w d i ∝ ( R - R ∘ ) ∫ 0 T ( w ˙ d i ss + w ˙ d i sds ) d t , (3)
see S1C Text. We refer to this rule as reward-modulated somato-dendritic synaptic plasticity (R-sdSP). Due to the term w ˙ d i sds it is effectively a 4-factor rule of the form ‘Δw = Rwrd⋅som⋅dend⋅pre’. The intuition is that the intrinsic neuronal stochasticities generate fluctuations in the somatic spiking that deviate from the prediction made by the dendritic input and cause an ‘error’ expressed in the somatic factor ( S - ρ ∖ d s ) of the rule. These fluctuations will be reinforced or suppressed by the feedback signal. As before, the synaptic modification will be strengthened if a presynaptic spike contributed to a dendritic NMDA-spike that in turn affects the somatic voltage.
We tested R-sdSP for various coding schemes. First, we considered a standard binary classification of frozen Poisson spike patterns by a postsynaptic spike- / no-spike code (Fig 2A). Each input pattern is defined as above (6Hz in 100 afferents for 500 ms) and belongs to one of two classes. For one class the soma is required to fire at least one spike while for the other class it should be silent. After repeated presentations followed by a reward signal, R-sdSP perfectly learned the correct classification of 4 random patterns. In contrast, reward-modulated STDP (R-STDP, [32]) implemented in its best performing version (see Online Methods and [33]) did not (Fig 2B). To achieve an appropriate alignment of dendritic spikes (Fig 2C and 2D), any successful learning rule needs to take account of the causal chain linking presynaptic spikes to dendritic and somatic spikes, the latter deciding upon reward or punishment. R-sdSP derived from maximizing the expected reward captures this causal relationship, but R-STDP does not, neither with a 10 ms (Fig 2B) nor with a 50 ms learning window (S2 Fig), and hence fails. Interestingly, R-STDP improves when the NMDA-spike generation is suppressed (Fig 2B, dashed). This shows that the increased flexibility in neuronal information processing provided by dendritic nonlinearities will in fact impede learning when a rule is used that does not take the nonlinearities into account.
R-sdSP is still able to correctly learn the classification even when the spike timings were noisy with a jitter up to 100 ms, or when the somatic voltage modulating the synaptic plasticity (via ρs and ρ ∖ d s) was low-pass filtered to mimic the dilution of information back-propagating to the synaptic site (S2 Fig).
Incidentally, the same task from Fig 2 can also be solved in a supervised scenario e.g. with the tempotron where, beside telling a neuron whether it should spike or not spike in response to a stimulus, the neuron is supposed to have access to the time of the voltage maximum within the stimulus interval [0, T], see e.g. [34, 35]. Although with these additional assumptions learning in principle becomes faster, the rules will again suffer from the ignorance about NMDA spikes and the possible acausality between a presynaptic spike and an immediately following somatic spike.
To apply the dendritic learning to a biological example we consider the direction selectivity of pyramidal neurons that was found to be mediated by nonlinear dendritic processing in vitro [5] and in vivo [7]. To mimick directional inputs moving in the stimulus space from right to left and left to right, we randomly enumerated the synapses across the whole dendritic tree and stochastically activated these synapses once in increasing and once in decreasing order (Fig 3A). After the stimulation, a positive reward signal was applied to the synapses when at least one somatic spike was elicited during the left-to-right patterns, or no somatic spike was elicited during a right-to-left pattern. A negative reward signal was applied in the other cases. R-sdSP, but not R-STDP, could learn such direction selectivity (Fig 3B). Individual dendritic branches may become selective to the synaptic activation order and learn to generate NMDA-spikes that, after summation in the soma, eventually trigger somatic action potentials (Fig 3C and 3D). Hence, the neuron learned to employ the dendritic nonlinearities to achieve direction selectivity, even though solving the task does not require them.
A classical task that exceeds the representational power of a point neuron is the XOR (exclusive-or) problem that is equivalent to the linearly non-separating feature binding problem [24]. In this task, the neuron has to respond exclusively to two disjoint pairs of features (e.g. to black & circle and to blue & diamond), but not to the cross combinations of these features (black & diamond and blue & circle). The presence and absence of a feature was encoded in a high and low Poisson firing rate, respectively, of a subpopulation of afferents projecting to our classifying neuron (Fig 3E). R-sdSP on the active dendrites could learn the correct responses, although due to the intrinsic stochasticity failures occurred in some cases. Classical R-STDP failed also in solving the feature binding problem problem on the dendritic tree, whether applied to pre-dend or to pre-som spike pairings (Fig 3F).
Besides learning a spike / no-spike code or a firing rate code, R-sdSP can also learn a spike timing code, i.e. to fire only at specific times. In a first task showing this, the neuron had to learn to spike at a target time ttarg = 250 ms in response to a frozen Poisson spike pattern. Deviations from this time were punished at the stimulus ending, using a graded feedback signal that increases with the magnitude of the deviation (Online Methods). During repeated pattern presentations, while applying R-sdSP and the delayed punishing signal, the postsynaptic spiking becomes concentrated in a narrow time window around the target spike (Fig 4A–4C). To understand the role of the active dendrites we separated the time course of the somatic voltage into the contribution from the subthreshold dendritic potentials and the NMDA-spikes (Fig 4D and 4E). After successful learning, the averaged NMDA-spikes form a broad ridge around ttarg on top of which the subthreshold dendritic voltages act as ‘scorers’. Before and immediately after ttarg the subthreshold voltage is hyperpolarized to prevent somatic spikes from coming too early or too late. Notice that in an individual run the summed NMDA-spikes can form plateaus that are much shorter than the NMDA-spike duration of 50 ms. In the example shown, this arises because just 5 ms after the initiation of an NMDA-spike in one branch another NMDA-spike ends in a second branch, cutting the somatic plateau short to 5 ms (Fig 4D). By virtue of the backpropagated somatic activity, R-sdSP learns to coordinate the timing of the NMDA-spikes in the different branches, creating a narrow window for a somatic spike around the target time.
Learning a spike-timing code is also possible if the rewarding / punishing signal is binary and is potentially delayed by several stimulus durations. We conceived a spatial navigation task where 7 positions on a circle are encoded each by a frozen 500 ms spike pattern in 100 afferents projecting to the dendritic branches of the model neuron as before. The task is to jump to position 0 when being in one of the other 6 positions and, after reaching position 0, staying there (Fig 5A). Actions consisted in either no jump or jumps of 1, 2 or 3 steps clock or counter clockwise. No jump is encoded by no somatic spikes, and a jump of n steps in the clock or counterclockwise direction is encoded by the first somatic spike arising in the n’th time bin to the left or right from the center (Fig 5A, inset). A positive reward signal R = 1 is delivered when the agent, being in a non-target position, directly jumps to the target, or when it is at the target position and stays there; else R = −1. After an initial average of 20 random actions needed to reach the target position, the R-sdSP modulated agents learned to eventually reach the target with a single action and stay there (Fig 5B and 5C). While initially the first somatic spike times of our model neuron were uniformly distributed across the 500 ms stimulus interval, the neuron eventually learned to respond in the appropriate time bin of 83 ms duration that encoded the correct jumps (Fig 5D). During learning, the dendritic branches develop a shared selectivity for the patterns and until the first NMDA-spikes become properly aligned in the correct time bin (Fig 5E).
We derived a synaptic plasticity rule for synapses on active dendrites that minimizes errors in the supervised and maximizes reward in the reinforcement learning scenario. More precisely, the rule follows the gradient of (a lower bound of) the log-likelihood of reproducing a given spike train for supervised learning, and the gradient of the expected reward for reinforcement learning. The rule specifies the optimal timing between the presynaptic, dendritic and somatic spikes, including the time course of the postsynaptic voltages. We showed that neurons can only exploit the increased representational power of active dendrites when synaptic plasticity is modulated by both the somatic and the dendritic spiking. The suggested somato-dendritic spike-dependent synaptic plasticity (sdSP) learns to correctly respond to synaptic input patterns coding by either frozen spikes times or firing rates, while classical STDP fails. It is remarkable that the same plasticity induction that supports the learning of precise spiking in the supervised learning scenario also maximizes the expected reward when modulated by an internal and possibly delayed reward signal, irrespective of whether the postsynaptic code is based on spike times or firing rates.
The neuron model for which we derived the gradient rules considered dendritic spikes as saturating square-shaped depolarizations triggered by the crossing of a dendritic voltage threshold. We showed that the dual voltage-glutamate criterion for NMDA spikes reduces in the presence of balanced excitation and inhibition to a pure voltage criterion. This is because the glutamate condition is always satisfied when reaching the voltage threshold. This leads to a dendritic spike scenario that also includes dendritic sodium [36] or calcium spikes [37] differing in their voltage threshold, duration and amplitude. In the supervised learning scenario, the general plasticity rule we derived consists of a somatic error term that measures the difference between the actual spiking and the instantaneous spike rate, a dendritic rate- and spike-term, and a presynaptic spike term. Potentiation is triggered if the presynaptic spike is followed by a postsynaptic spike within roughly 10 ms, and this time window is stretched to roughly 50 ms if between the pre- and postsynaptic spike an additional dendritic spike occurs. Plasticity, be it potentiation or depression, can also be boosted by a mere nonlinear dendritic depolarizations without dendritic spikes, linking the rule also to computational models considering nonlinear but continuous dendritic processing [8–11, 24]. In the reinforcement learning scenario, the same plasticity rule is modulated by a global reward signal.
As learning is driven by a somatic error term, the synapses must be able to readout this error by disentangling the backpropagating spike and the somatic voltage (or at least a low-pass filtered version of it, see S2C–S2F Fig). Synapses must also read out the local dendritic spike and potential, and the synapse-specific postsynaptic potential (PSP). The PSP may be inferred from the concentration of the local glutamate released by the presynaptic bouton. The somatic and dendritic spikes may be determined from their characteristic voltage upstrokes and sustained depolarizations, and the NMDA spike can be further detected by a rapid increase in the local calcium concentration. Finally, the (subthreshold) somatic and dendritic depolarization may be distinguished by co-sensing local ion concentrations. In fact, the synaptically induced dendritic depolarization goes together with an increased local sodium concentration while the backpropagating somatic voltage does not cause such a ion influx. We assume that synapses developed a molecular machinery to extract these quantities and infer approximate estimates for the terms occuring in our plasticity rules.
Our computational framework for active dendrites contributes to the debate whether plasticity on dendritic branches should depend on the dendritic rather than the somatic spike [38], or whether it subserves synaptic clustering [23, 39] or a homeostatic adaptation [40]. In fact, when seen in the light of learning, synaptic plasticity is predicted to depend on all the postsynaptic quantities. Based on the model of dendritic NMDA receptor conductances in an in vivo stimulation scenario, the learning rule yields a guideline for experimental testings. For instance, it is in line with the observed synaptic depression induced by a synaptically generated dendritic spike alone ([41], but see [42]), or with the extended time window for plasticity induction involving NMDA-spikes [21]. It predicts that an NMDA-spike within roughly 50 ms after an excitatory synaptic input always enhances the synaptic modification. While the sign of the synaptic modification is determined by the presence or absence of a somatic spike following the synaptic input, an additional synaptically evoked NMDA-spike will only enhance it, never revert this sign.
Dendritic structures that have been suggested to form a 2-layer network [8] offer the additional advantage of easily backpropagating the information of the output to the synaptic sites 2 layers upstream. From a computational perspective it is interesting to note that, one the one hand, 2-layer networks represent an universal function approximator [43] while, on the other hand, networks with more than two layers are difficult to be trained [13]. For the dendritic implementation this suggests to limit the internal nonlinearities to a single layer of active dendritic branches. Because stacking dendritic nonlinearities across multiple layers would cause additional cross-talk, the restriction to a single dendritic nonlinearity may just be nature’s solution to the trade-off between achieving more representational power and paying the associated signaling costs required for efficient learning.
|
10.1371/journal.pntd.0001416 | Ciguatera Fish Poisoning in the Pacific Islands (1998 to 2008) | Ciguatera is a type of fish poisoning that occurs throughout the tropics, particularly in vulnerable island communities such as the developing Pacific Island Countries and Territories (PICTs). After consuming ciguatoxin-contaminated fish, people report a range of acute neurologic, gastrointestinal, and cardiac symptoms, with some experiencing chronic neurologic symptoms lasting weeks to months. Unfortunately, the true extent of illness and its impact on human communities and ecosystem health are still poorly understood.
A questionnaire was emailed to the Health and Fisheries Authorities of the PICTs to quantify the extent of ciguatera. The data were analyzed using t-test, incidence rate ratios, ranked correlation, and regression analysis.
There were 39,677 reported cases from 17 PICTs, with a mean annual incidence of 194 cases per 100,000 people across the region from 1998–2008 compared to the reported annual incidence of 104/100,000 from 1973–1983. There has been a 60% increase in the annual incidence of ciguatera between the two time periods based on PICTs that reported for both time periods. Taking into account under-reporting, in the last 35 years an estimated 500,000 Pacific islanders might have suffered from ciguatera.
This level of incidence exceeds prior ciguatera estimates locally and globally, and raises the status of ciguatera to an acute and chronic illness with major public health significance. To address this significant public health problem, which is expected to increase in parallel with environmental change, well-funded multidisciplinary research teams are needed to translate research advances into practical management solutions.
| Ciguatera fish poisoning occurs throughout the tropics. After consuming contaminated coral reef fish, people report a range of acute neurologic, gastrointestinal, and cardiac symptoms, with some experiencing chronic neurologic symptoms lasting weeks to months. Ciguatera is largely caused by toxins from benthic microalgae of the genera Gambierdiscus that are bioaccumulated in reef fish through the marine food chain. Unfortunately, the true extent of illness and its impact on human communities and ecosystems are still not well understood. Using data gathered from Health and Fisheries Authorities of the Pacific Island Countries and Territories we identified a 60% increase in the annual incidence of ciguatera from 1988–2008 to 1973–1983 and estimate over 500,000 Pacific islanders might have suffered from ciguatera in their lifetime. The incidence of ciguatera is expected to continue to rise in conjunction with continued reef degradation and global warming, with greatest impact likely to be experienced in the developing PICTs and potentially the archipelagoes of southeast Asia. Despite this threat, little funding is available for research that might lead to better management of the problem either locally, regionally or globally.
| The developing Pacific Island Countries and Territories (PICTs) are under increasing threat from both acute and chronic diseases ranging from HIV/AIDS to obesity. In addition, people residing in PICTs are highly vulnerable to environmental impacts from the sea level rise and extreme weather events associated with global warming. Ciguatera is a prevalent tropical and subtropical disease that has been an under-appreciated cause of acute and chronic disease in island communities and might be increasing in incidence due to increasing vulnerabilities (i.e. poverty, global warming, eutrophication) in these populations [1]–[3].
Ciguatera is caused by the consumption of coral reef fish contaminated by ciguatoxin and related toxins from dinoflagellates (microalgae) and cyanobacteria [1], [2]. The ciguatoxin bioaccumulates up the food web, either directly from incidental uptake by herbivorous fish or indirectly by carnivorous fish [1]. After the consumption of coral reef fish contaminated with ciguatoxin, people experience potentially severe acute neurologic, gastrointestinal and cardiac symptoms as well as, in some cases, chronic neurologic symptoms lasting weeks to months [1], [2]. Ciguatera occurs globally, in coastal tropical waters, and is particularly prevalent across the PICTs. Cases of ciguatera have also been reported in temperate regions of the world due to travel and coral reef fish export. Ciguatera poisoning is often under-diagnosed and under-reported, with only 2 to 10% of cases reported to health authorities (Friedman et al 2008). Estimates of the incidence of ciguatera in Oceania have ranged from 0.5/10,000/year in Hawaii to 5,850/10,000/year in French Polynesia [2]. Fish is the staple protein source in many PICT communities, with many islands in the region suffering ongoing outbreaks of ciguatera leading to potentially significant impacts on large portions of the population of small island communities when toxic fish are consumed [3].
Dinoflagellates of the genera Gambierdiscus, that grow epiphytically on macro- and turf-algae on coral reefs, produce the ciguatoxins predominantly responsible for the disease known as ciguatera. Coral reef damage, or when algal growth is not controlled by herbivorous fish, provide increased potential habitat for Gambierdiscus growth that might increase the risk of ciguatera [4]. Despite extensive research, we know little about the ecology and the environmental factors that cause the blooms of the ciguatera caustive organisms, nor do we understand the role (if any) of other dinoflagellate genera including Ostreopsis (palytoxin producers) and Prorocentrum (okadaic acid and dinophysistoxins producers) or marine cyanobacteria [5]. Presently, ciguatoxin can only be detected in fish and Gambierdiscus in specialized labs, and diagnosis in humans is based almost exclusively on symptoms associated with the recent consumption of a potential ciguateric fish; factors that hamper its effective management and highlight important research needs [1], [2].
A number of factors have been associated with ciguatera cases and the presence of ciguatoxic dinoflagellates. Military activities causing coral reef damage in the Pacific, including from World War II, and nuclear test explosion programs, have been linked with outbreaks and changing incidence of ciguatera in some locations [6]. The prevalence of ciguatera in the South Pacific increases dramatically where average sea surface temperatures are at least 28 to 29°C [7]. Elevated sea surface temperatures associated with global warming are believed to already be exacerbating the extent and the range of ciguatera [8], . Reportedly, ciguatera occurrences are most prevalent in the warmest regions of the Caribbean, and all indigenous ciguatera cases have occurred where annual average temperatures are >25°C [4]. Nutrient enrichment and warming sea surface temperatures have been shown to stimulate Gambierdiscus growth which results in higher cell densities [10]. Also, benthic dinoflagellate species, including those of the genera Gambierdiscus, might have extended biogeographical ranges, induced by human activity. For example, benthic dinoflagellates are likely to be able to colonize previously unoccupied locations through transport in ship ballast [11]. Certain species of Gambierdiscus has now been found to be highly ciguatoxic compared to the other species [12], and blooms of these species are likely to contribute most to ciguatera risk.
Given changes in global climate patterns, increased degradation of coastal marine environments through coastal development and land run-off, and increased exploitation of coastal marine resources, the incidence of ciguatera cases is predicted to continue to increase in the future [4], [13]. Therefore, we hypothesized that ciguatera incidence is an increasing human health and ecological concern across the PICTs. To test this hypothesis, we report on changes in ciguatera incidence across the Pacific, and the social consequences of changing ciguatera incidence by comparing two 11 year periods of data: 1973–1983 vs 1998–2008.
The Secretariat of the Pacific Community, the Institut Louis Malarde (Tahiti), the Institut Louis Pasteur (New Caledonia), and Institute for Research and Development (IRD) organized a Ciguatera workshop held in Noumea, October 2008. At this workshop, many island nation delegates declared a need for the ciguatera concern to be better addressed. To start to understand the current extent and nature of the ciguatera problem, we distributed a questionnaire to all PICTs (Supporting information S1). Ciguatera records used in this study are housed in each PICTs government health institution (Ministries and Departments of Health and Public Health).
To obtain the ciguatera records for the period of 1998 to 2008, we first contacted the Secretariat of the Pacific Community (SPC), Division of Fisheries, Aquaculture and Marine Ecosystems (FAME) to obtain the list of institutions responsible for maintaining ciguatera records within PICTs. We considered these repositories were comaparable to the data collection repositories used in the Lewis et al. study [14]. The questionnaire (Supporting Information S1) was sent by email to the institutions in October 2009. Updates on the returning of the questionnaires were sent to PICTs on four occasions over the period, and questionnaires were returned from the PICTs up until April 2010.
The questionnaire was developed by the co-authors in collaboration with the PICTs. The questionnaire included questions and definitions from prior ciguatera studies to provide consistency of data gathering and allow comparison across studies. The 3 key sections of the questionnaire collected information on: 1) Temporal incidence of ciguatera; 2) Environmental disturbance, to examine if coral reef condition and occurrence of coral bleaching and cyclones might influence ciguatera incidence (these data were considered purely speculative on the part of the respondent); and 3) Social consequences of ciguatera including changing diet and associated medical conditions, proactive and reactive management of ciguatera, and the desire for external assistance in response to ciguatera.
Nearly all PICTs responded (17 or 85%), with half fully completing the ciguatera questionnaires. Whilst we contacted the health authorities for the ciguatera data (which were returned by the health authorities in all cases), other questions were left incomplete as they were not directly about the ciguatera health issue or were sent to the other government authorities to be fully completed.
The reported cases for the recent 11 year period showed high levels of inter-year variability within and between PICTs. In Fiji, Kiribati and French Polynesia, more cases occurred at the start of the period. Annual reported cases peaked around the middle of the period at Cook Islands, Marshall Islands, Tokelau, Mariana's, and Hawaii. Reported cases in Vanuatu peaked towards the later part of the 11 year period and since 2005, Fiji experienced an increase in the number of ciguatera cases. Finally, Palau, Hawaii, Guam, Samoa, Wallis and Futuna, and Nauru all had relatively consistent incidence rates of under 5/100,000 (Table 1). Additional data relating to ciguatera incidence within PICT archipelagos are presented in Table 2.
Within the 35 year period (1973–2008), including the study by Lewis [14] and this study, there was a clear overall increase in CPF incidence; however, the results show inter-PICT incidence variability between the two time periods (Figure 1). Cook Islands, Vanuatu, Fiji, Tokelau, Marshall Islands, Niue, Tonga, and Palau all have increased ciguatera incidence (Table 3). Others have shown a decrease in ciguatera incidence in comparison to the other PICTs (such as Tuvalu and New Caledonia). From 1973–1983, only four nations demonstrated a ciguatera incidence over 300/100,000; now seven nations have an incidence over this value. Fiji now outranks French Polynesia as the nation with the highest number of ciguatera cases. Previously only four nations had over 2,000 ciguatera cases; recently six nations reported ciguatera cases over this value. Fiji, French Polynesia, Vanuatu, Kiribati, Cook Islands, and Tokelau all demonstrated an increase in the number of ciguatera cases; New Caledonia, Tuvalu and Guam showed a decrease in the number of cases.
Statistical analysis of temporal change in ciguatera incidence showed varied results. There was no statistically significant difference between 1973–1983 mean incidence and 1998–2008 mean incidence across PICTs (p = 0.949), comparing all PICTs (except the State of Hawaii which was not presented by Lewis [14]) using paired t-test (Table 4). However, there was a highly significant difference in mean incidence, between the two time periods when comparing mean incidence across years (p = 0.002), using independent-sample t-test. Linear regression analysis of annual mean incidence from 1973 to 2008 was also statistically significant (p = 0.005) despite high inter-year variability. The rate ratio (1998–2008 mean annual incidence/1973–1983 mean annual incidence) was 1.60; therefore, there was a 60% increase in the mean annual incidence from the earlier period to the more recent period. Hawaii, North Marianas, Marshall Islands, and Palau were omitted for the independent sample t-test, regression analysis and the rate ratio due to data limitations (Table 1).
Of the 18 PICTs in this study, 11 reported on all (i.e. coral bleaching, cyclone incidence and perceived reef condition), whilst New Caledonia reported only on coral bleaching (Table 5). All three environmental disturbance types were positively related to ciguatera incidence. However, there was no statistically significant correlation between mean annual ciguatera incidence and occurrence of bleaching (p = 0.20), cyclone incidence (p = 0.17) or perceived coral reef condition (p = 0.57).
Responses to questions relating to the social consequences of ciguatera demonstrated that the incidence of ciguatera might be having a negative impact on PICT communities. Seven PICTs reported changes in diet as a result of ciguatera, whilst six PICTs reported that there was no change in diet as a result of ciguatera. Also, seven PICTs reported secondary medical problems (such as diabetes due to dietary changes) as a result of ciguatera. Five PICTs reported both a change in diet and secondary medical problems as a result of ciguatera. Seven PICTs reported taking reactive management measures (such as closure of fishing areas) to manage ciguatera outbreaks, whilst four PICTs reported taking no reactionary measures. Four PICTs reported that preventative management (such as catchment management) was occurring, whilst four PICTs reported that there was no preventative management. Eight PICTs reported that additional support would improve the management of ciguatera, whilst four reported that it would not.
There was a positive and marginally significant relationship between changing diet and per capita incidence of ciguatera (p = 0.06), and secondary medical problems and per capita incidence of ciguatera (p = 0.08) (Table 6). Neither reactive nor proactive management was correlated with per capita incidence of ciguatera. However, perceived improvement in management as a result of increased support was positively correlated with per capita incidence of ciguatera (p = 0.013). There was no significant difference (p≤0.05) in per capita incidence of ciguatera between nations that did and did not respond to questions on the social consequences of ciguatera.
This study provides four important findings. First, as hypothesized, ciguatera incidence has increased significantly in the Pacific since the 1970s, but there is significant variability in incidence within PICTs since this time. Second, predicting causes of outbreaks and consequent elevated levels of ciguatera is difficult at the scale of this study, highlighting the need for further local-scale research and management action. Third, as reported earlier [3], ciguatera incidence continues to have significant negative effects on PICT societies, including dietary changes and associated medical problems (such as diabetes). Fourth, there has been inadequate response to date, yet there is acknowledgement from a number of PICTs that assistance would aid in the management of ciguatera. Such assistance could provide appropriate support and unified action might lead to solutions to a disease that could be considered an important cause of both acute and chronic illness in the Pacific.
Based on the results of this study compared to historical analyses, the overall incidence of ciguatera per 100,000 people appears to have increased significantly in the Pacific comparing 1973–1983 (mean104 cases/100,000 [14]) with 1998–2008 (mean194/100,000). There has been a 60% increase in the annual incidence of ciguatera between the two time periods based on PICTs that reported for both time periods (Figure 1). Two nations which exemplify the potential degree of change in incidence of ciguatera are the Cook Islands, where the incidence rose from 2/100,000 to 1,554/100,000 between the two time periods; and Tuvalu, where the incidence decreased from 462/100,000 people to 83/100,000 people. Furthermore, while it might appear that ciguatera incidence rates have subsequently fallen, they are still higher than the levels reported earlier by Lewis [14]. The non significant result from the paired t-test comparing within PICT ciguatera incidence for the two time periods suggests that there is significant variability of ciguatera incidence within PICTs through time. Therefore temporal change of incidence is difficult to predict at the PICT scale. However, the independent sample t-test and regression analysis revealed a regional increase in ciguatera incidence, highlighting the need for regional action.
Using the conservative estimate that the official reported ciguatera represents 20% of actual incidence [14], then the actual average overall incidence for the region would be 970/100,000 for 1998 to 2008. Others have estimated that only 5–10% of ciguatera cases are actually reported [2]. Across the region, using the reported mean values of actual cases for the three periods for which we have data (1,762 for 1973–1983; 2,844 for 1989–1992 (South Pacific Epidemiological and Health Information Services data); and 3,607 for 1998–2008 (current study)) and using a conservative under reporting rate of 80%, we estimate that since 1973 approximately 500,000 PICT inhabitants have had ciguatera.
It is possible that there might be a reporting bias in the data because of increased research and interest in ciguatera compared to the 1973–1983 time period. However, our data demonstrate high variability of ciguatera reporting from 1998–2008 across the PICTs. It is beyond the scope of this study to ascertain the effect of immigration and translocation of people to and from some of these PICTs, with different dietary habits than the local inhabitants, on ciguatera incidence. Given the variability in the change of incidence across the region demonstrated in this study, it is clear that the overall ciguatera trend cannot be extrapolated from data for a single PICT.
We elicited a relatively poor response rate from questions relating to coral bleaching, cyclones, and degraded reef conditions. Such environmental disturbance generally occurs at finer scales, so it might be appropriate to perform a more detailed field surveys in collaboration with environmental, fisheries, and meteorology agencies in the future to better understand such effects. However, despite the methodological limitations, we have shown that there is a trend for ciguatera incidence to be higher where bleaching, cyclones, and poor reef condition have been reported.
Stronger relationships were identified between ciguatera incidence and social impacts of ciguatera outbreaks. We found a marginally significant positive relationship between changing diet and the incidence of ciguatera, and associated medical problems and incidence of ciguatera. Such problems increase financial and social burdens on PICTs. Addressing the underlying causes of ciguatera outbreaks will reduce this burden, enabling PICT authorities to redistribute their limited resources to other priorities. Management action and prevention do not correlate with ciguatera incidence highlighting the lack of a unified and systematic approach for addressing ciguatera in the region. A clear desire for assistance exists within the PICTs that have high ciguatera incidence, suggesting that PICTs would be highly receptive to an external body aiding in enabling unified and systematic action. In addition to exploring new and better apporaches to detection and treatment, research is needed into the causes of ciguatera outbreaks, including environmental and anthropogenic parameters, to explain the hypotheses raised by this study.
It is possible that the unusual collaboration of the majority of the PICTs in this project might have contributed to the observed increased reporting of ciguatera (as well as other unidentified infrastructure changes), and thus a possible reporting bias for the more recent 1998–2008 data when compared with the 1973–83. However, these data represent a decade of reporting during a period of competing public health interests and lack of surveillance resources for ciguatera in the PICTs. As with all ciguatera studies where the case definition does not include active confirmation of ciguatoxin in the fish consumed, there is the possibility of the misclassification of reported cases; however, this situation has not changed from the 1970s [2].
It is also beyond the scope of this study to speculate on the causes of the high spatial and temporal variability of ciguatera [4]. This study, instead, aimed to demonstrate that ciguatera is still of major and possibly growing concern in the region. Addressing ciguatera will require significant investment in research and continuing education campaigns. If the suspected disturbances (including coral bleaching, cyclones, shipwrecks, and port facilities) are major causes of ciguatera outbreaks, then it is likely that the general temporal pattern of increased outbreaks will continue in the region, and be a far more expensive concern in the future, if our understanding of, and response to, ciguatera is not extensively improved.
Despite over 50 years of ciguatera research in the Pacific, no comprehensive region-wide action has occurred to better manage ciguatera. Based on this study, an estimated 500,000 persons might have contracted ciguatera in the last 35 years, corresponding to a lifetime prevalence of 25%. It is remarkable that ciguatera has largely been ignored by the PICT national governments, with only two nations having an ongoing monitoring program and only one nation having a small unit devoted to researching ciguatera (Toxic Micro-algae Unit of the Institut Louis Malarde, French Polynesia). Given the rapidly changing physical environment (including global warming, extreme weather, and coral reef degradation) as well as the dependence of local populations upon fish for their physical and cultural survival, research into improved disease treatment and toxin detection, and a better understanding of the environmental factors contributing to ciguatera, are required to help reduce the likely growing adverse impacts of ciguatera.
|
10.1371/journal.ppat.1004964 | NK-, NKT- and CD8-Derived IFNγ Drives Myeloid Cell Activation and Erythrophagocytosis, Resulting in Trypanosomosis-Associated Acute Anemia | African trypanosomes are the causative agents of Human African Trypanosomosis (HAT/Sleeping Sickness) and Animal African Trypanosomosis (AAT/Nagana). A common hallmark of African trypanosome infections is inflammation. In murine trypanosomosis, the onset of inflammation occurs rapidly after infection and is manifested by an influx of myeloid cells in both liver and spleen, accompanied by a burst of serum pro-inflammatory cytokines. Within 48 hours after reaching peak parasitemia, acute anemia develops and the percentage of red blood cells drops by 50%. Using a newly developed in vivo erythrophagocytosis assay, we recently demonstrated that activated cells of the myeloid phagocytic system display enhanced erythrophagocytosis causing acute anemia. Here, we aimed to elucidate the mechanism and immune pathway behind this phenomenon in a murine model for trypanosomosis. Results indicate that IFNγ plays a crucial role in the recruitment and activation of erythrophagocytic myeloid cells, as mice lacking the IFNγ receptor were partially protected against trypanosomosis-associated inflammation and acute anemia. NK and NKT cells were the earliest source of IFNγ during T. b. brucei infection. Later in infection, CD8+ and to a lesser extent CD4+ T cells become the main IFNγ producers. Cell depletion and transfer experiments indicated that during infection the absence of NK, NKT and CD8+ T cells, but not CD4+ T cells, resulted in a reduced anemic phenotype similar to trypanosome infected IFNγR-/- mice. Collectively, this study shows that NK, NKT and CD8+ T cell-derived IFNγ is a critical mediator in trypanosomosis-associated pathology, driving enhanced erythrophagocytosis by myeloid phagocytic cells and the induction of acute inflammation-associated anemia.
| African trypanosomes are the causative agents of Human and Animal African Trypanosomosis, impairing economic development and causing death throughout the African continent. Anemia and inflammation are hallmark features of virtually every type of trypanosome infection. During experimental murine trypanosomosis, early inflammation causes enhanced red blood cell phagocytosis by cells of the myeloid phagocytic system, leading to severe anemia within 48 hours past peak parasitemia. Here, we identify the pro-inflammatory cytokine IFNγ as the main driver of the early inflammatory reaction and enhanced red blood cell phagocytosis. This IFNγ is derived consecutively by NK, NKT and CD8+ T cells, hence these cells all play a crucial role in the induction of inflammation and anemia.
| African trypanosomes cause a wide range of disease phenotypes, but a common hallmark of the infection is inflammation. Early during the course of infection, myeloid cells get activated by released parasite components such as soluble variant surface glycoproteins (sVSG) and DNA [1–7]. This gives rise to a type 1 cytokine storm which is critical for resistance [6,8–11], but is also associated with pathology development [12–16]. Indeed, coinciding with the acute inflammatory reaction, acute anemia develops, as witnessed by a 50% reduction in circulating red blood cells (RBC) within two days following peak parasitemia. After a short recovery phase, a subsequent gradually increasing loss of RBCs occurs during the chronic infection stage [13,17]. Anemia development is independent of antibodies [18] and the height of the parasitemia peak [17], and the acute nature of this phenomenon implies a consumptive etiology. Using a newly developed in vivo erythrophagocytosis assay, we have recently shown that acute anemia during Trypanosome infection is caused by enhanced RBC phagocytosis by activated cells of the myeloid phagocytic system, in combination with a decrease in RBC membrane stability [19]. More specifically, during the acute phase of T. b. brucei infection, activated liver neutrophils and monocytic cells (comprising monocytes and monocyte-derived macrophages) as well as activated spleen resident macrophages display enhanced erythrophagocytosis. This, in combination with the decreased RBC membrane stability, leads to disproportionate amount of RBC phagocytosis and hence acute anemia [19]. It is suggested that cells of the myeloid phagocytic system are ‘over’-activated by the type 1 induced inflammation early in infection, however the exact mechanism and pathway by which this occurs is unknown.
Previous studies on African trypanosome infections have established that IFNγ is required to prime macrophages in order to become fully activated and induce an efficient type 1 response [2,3,6,20]. This indicates that IFNγ production occurs very early in infection, even before macrophage activation. Although no direct evidence was provided, others have implied CD8 T cells [21–24] and VSG-specific CD4 T cells [9] to be potential sources of IFNγ during African trypanosome infections. In addition, it was recently shown in murine Toxoplasma gondii infection that IFNγ can act directly on macrophages to provoke RBC uptake [25].
In this study we aimed to elucidate the mechanism(s) and immune pathway(s) responsible for the induction of acute anemia during African trypanosome infection.
Here, using the clonal laboratory-adapted Trypanosoma brucei brucei (T. b. brucei) strain, we show that mice lacking the IFNγ receptor suffer less from infection-associated inflammation and acute anemia. Moreover, we show for the first time that during experimental trypanosome infections NK and NKT cells are the earliest IFNγ producers, followed by CD8 and CD4 T cells, and that IFNγ plays a crucial role in the recruitment and activation of erythrophagocytic myeloid cells. In addition, the results indicate that the absence of NK, NKT and CD8 T cells, but not CD4 T cells, during the early stage of infection results in a reduced anemic phenotype similar to IFNγR-/- mice.
Collectively, this study shows that NK-, NKT- and CD8-derived IFNγ is crucial for enhanced erythrophagocytosis by myeloid phagocytic cells and consequently for the induction of acute inflammation-associated anemia.
6–8 week old female C57BL/6 mice were purchased from Janvier. C57BL/6, IFNγ-/- and IFNγR-/- mice were obtained through Dr. B. Ryffel (CNRS, Orleans, France). The interferon-gamma reporter with endogenous polyA transcript (GREAT) mice were purchased from The Jackson Laboratory. These mice were housed in individual ventilated cages at the Vrije Universiteit Brussel.
C57BL/6 CD4-/-, CD8-/- and C57BL/6 nu/nu mice were a kind gift from Dr. H. Mossmann (MaxPlanck Institute, Freiburg, Germany). These mice were housed in individual ventilated cages and maintained in SPF barrier facilities at the University of Cape Town.
All experiments complied with the ECPVA guidelines and were approved by the ETHICAL COMMITTEE for ANIMAL EXPERIMENTS (ECAE) at the Vrije Universiteit Brussel (protocol #14-220-23 and #12-220-2) and the University of Cape Town, South Africa # 97/001 and 005/041.
Mice were infected by intraperitoneal (i.p.) injection of 5000 pleomorphic Trypanosoma brucei brucei AnTat1.1E parasites, which were a kind gift from N. Van Meirvenne (Institute for Tropical Medicine, Belgium). RBC counts were determined via a hematocytometer at two to four day intervals on 2,5μl blood sample collected from the tail vein and diluted 1/200 in PBS. Anemia was expressed as the percentage of reduction in RBC counts compared to non-infected animals.
For depletion of CD8 T cells, mice received the first i.p. injection of 500μg anti-CD8 Ly2 rat-anti-mouse monoclonal antibody 24 hours prior to infection. Subsequently, mice received a dose of 100 μg 2 day intervals post infection. NK and NKT cells were depleted with the anti-NK1.1 PK136 rat-anti-mouse monoclonal antibody. 250μg was given four and one day prior to infection. A dose of 300μg was given at 2–3 day intervals post infection. Depletion efficiency of CD8 T cells and NK(T) cells from both spleen and liver was assessed by flow cytometry.
For neutralization of IFNγ, wild type mice were treated with 500μg neutralizing IFNγ antibody (clone XMG1.2, Bioceros) at two-day intervals, starting at day 1 post infection. Control mice were treated with corresponding volumes of PBS.
Blood was harvested from CO2 euthanized mice by cardiac puncture and centrifuged at 10000rpm for 10 min. Serum was harvested and stored at -20°C.
Spleen and liver were harvested from CO2 euthanized mice. For myeloid cell analysis livers were perfused with cold PBS prior to harvesting. Consecutively, these livers were minced in 10 ml digestive medium (0.05% collagenase type A in Hanks' Balanced Salt Solution (HBSS); Invitrogen) and incubated at 37°C for 30 min. The digested tissue was then homogenized (GentleMacs) and filtered (40 μm pore filter). For analysis of lymphocytes, livers were homogenized and filtered (40 μm) and restricted to a 33% Percol in PBS gradient (1800 rpm, 12 min, room temperature). Spleen cells were obtained by homogenizing the organs in 10 ml RPMI medium containing 5% fetal calf serum (FCS) and filtered (40 μm pore filter). Next, liver and spleen cell suspension were centrifuged (1400 rpm, 7 min, 4°C) and the pellet treated with RBC lysis buffer (0.15 M NH4Cl, 1.0 mM KHCO3, 0.1 mM Na2-EDTA). Subsequently, the cells were resuspended in CM–medium (RPMI medium, 5%FCS, 1% L-glutamine, 1% Penicillin-Streptomycin) or RPMI 5% FCS or plain RPMI for cell culture flow cytometry or cell isolation respectively. For cell culture 3 105 cells were put in flat bottom 96-well plates and incubated at 37°C and 5% CO2. Cell culture supernatant was harvested after 24 or 48 hours and stored at -20°C.
CD8-/- mice were reconstituted with splenocytes from naïve C57BL/6 donor mice. CD8 T cells were purified via negative selection using the EasySep Mouse CD8+ T Cell Isolation kit according to the manufacturers protocol (StemCell Technologies). Obtained cell suspensions were between 80 and 90% pure. CD8 T cells were carboxyfluorescein succinimidyl ester (CFSE) labeled allowing retracement of transferred cells in acceptor mice (S1). Briefly, CD8 T cells were put at a concentration of 107 cells per ml and incubated with 5μM CFSE for 15 min at 37°C 5% CO2. Subsequently, labeled cells were incubated for 15 minutes with 10 ml PBS 1%BSA at 37°C, 5% CO2 and washed twice with the same medium at 2000 rpm, 7 minutes. Between 5x106 and 1x107 CD8 T cells were injected i.v. into acceptor mice 24 hours prior to infection and four days post infection.
C57BL/6 nu/nu mice were reconstituted with splenocytes from naïve C57BL/6 donor mice. CD8+ and CD4+ T-cell purification was performed using the antibody cocktail and density gradient method (Stem Cell Technologies) according to the manufacturers’ protocol. Obtained cell suspensions were 95% pure. 3 107 cells were injected i.p. into acceptor mice 24 hours prior to infection. CD4+ T-cell reconstituted mice were given an additional injection of 500 μg anti-CD8 mAb, 2 hours after cell reconstitution.
Blood was harvested from CO2 euthanized mice by cardiac puncture using 50μl 1000 U/ml heparin. RBCs were counted and 109 RBCs were washed twice with 15 ml PBS, 2000 rpm, 7 minutes. Next, RBCs were labeled with 2 μl pHrodo Red succinimidyl ester (pHrodo Red, Life Technologies) in a final volume of 1 ml PBS for 60 minutes at 37°C 5% CO2. Subsequently, labeled RBCs were incubated for 15 minutes with 10 ml RPMI/5% FCS at 37°C and washed twice with the same medium, 2000 rpm, 7 minutes. Labeled RBCs were resuspended in RPMI. As negative control, and for the determination of the background signal, RBCs were treated in the same manner without addition of the pHrodo dye. 7–8 week old female C57BL/6 non-infected or T. brucei infected (day 6 p.i.) mice were injected intravenously (i.v.) with 109 pHrodo labeled or unlabeled RBCs in 200 μl RPMI. After 18 hours, mice were CO2 euthanized and spleen and livers were isolated and processed into single cell suspension as described above. Next, the cells were analyzed via flow cytometry as described further.
Spleen cells were isolated as described and resuspended in ME—medium (RPMI medium, 5% FCS, 1% L-glutamine and non essential amino acids, 1% Penicillin-Streptomycin) at a concentration of 4 105 cells per 200 μl. Red blood cells were isolated and labeled as described and 2 107 labeled or unlabeled RBCs were put in co-culture with 4 105 cells in polypropylene tubes (BD Biosciences). Co-cultures were incubated overnight at 37°C and 5% CO2 with or without IFNγ stimulation (100U/ml). After overnight culture, cells were submitted to flow-cytometrical analysis.
Cells were washed with FACS medium (5% FCS in RPMI) and non-specific binding sites were blocked by incubating 20 minutes at 4°C with an Fc-blocking antibody (anti-CD16/32, clone 2.4G2). Next, cell suspensions were stained with fluorescent conjugated antibodies for 30 minutes at 4°C. Fluorescent antibodies: CD11b PE-Cy7 clone M1/70, F4/80 FITC clone C1:A3-A, Ly6C APC clone AL-21, Ly6G PerCP-Cy5.5 clone 1A8, CD45 APC-Cy7 clone 30-F11, CD4 BV421 clone GK1.5, CD8 BV510 clone 53–67, NK11 PE clone PK136 (BD Biosciences), CD64 Pe clone X54-5/7.1. (BioLegend), CCR2 Pe clone 475301, MerTK Pe clone 108928 (R&D systems), Ly6B clone 7/4 (AbD Serotec)., TCRb APC clone H57-597, CD49b Pecy7 clone DX5, NKp46 PE clone 29A1.4 (eBioscience). Following washing with FACS buffer they were analyzed on a FACS Canto II flow cytometer (BD Biosciences) and data was processed using FlowJo software (Tree Star Inc.).
Concentrations of IL15/IL15R, IL12p70 and TNFα (R&D Systems) as well as IFNγ (Pharmingen) in serum and cell supernatant were determined by ELISA according to the manufacturers’ protocol.
Statistical analysis was performed using Student-test and GraphPad Prism software (GraphPad 6, San Diego, CA). Values are expressed as mean ± standard deviation (SD) unless stated otherwise. Values of p≤ 0.05 are considered to be statistically significant.
Trypanosome infections are characterized by multiple parasitemia waves and a survival of approximately 30–35 days [8,12]. The peak of parasitemia occurs at day 5–6 post infection, followed by acute anemia development [17]. Previous research on African trypanosome infections has established an important role for IFNγ during the onset of infection. Indeed, IFNγ is crucial for macrophage activation and optimal initiation of the type 1 immune response associated with resistance to infection [6,11]. Coinciding with the peak of parasitemia and induction of anemia (day 6), a burst in serum pro-inflammatory cytokines is observed [14,17]. To investigate the role of IFNγ during T. b. brucei infection-associated pathology, IFNγR-/- mice were infected and anemia was monitored. Infected IFNγR-/- mice suffered much less from acute anemia compared to infected C57BL/6 mice (Fig 1A). Coinciding in IFNγR-/- mice, reduced amounts of pro-inflammatory cytokines IL-15 and TNFα were observed at the time of peak parasitemia (day 6) (Fig 1B).
To investigate the contribution of IFNγ in the alteration of the myeloid cell composition a detailed investigation (gating strategy S1A and S1B Fig) of liver and spleen cell composition was performed. The liver of IFNγR-/- mice showed a reduced influx of neutrophils (defined as CD11b+, Ly6G+) and monocyte-derived macrophages (defined as CD11b+, Ly6C+, MHCII+) at day 4 post infection (Fig 1C and S1C Fig). No changes were observed in monocytes (defined as CD11b+, Ly6C+, MHCII-) and resident macrophages (defined as CD11b+, Ly6C-, MHCII+). Therefore, the composition of liver myeloid cells of infected IFNγR-/- mice closely resembled that of naïve C57BL/6 mice (Fig 1D). Of note, there was no difference in myeloid cell composition between naïve C57BL/6 and IFNγR-/- mice. Similarly, the myeloid cell composition in the spleen of infected IFNγR-/- mice showed a reduced influx of monocyte-derived macrophages, without differences in percentages of resident macrophages, when compared to infected C57BL/6 mice (Fig 1E and 1F, S1D Fig). In contrast to the situation in the liver, neutrophil influx was similar to wild type mice, and the proportion of monocytes within the CD45+ population increased in IFNγR-/- mice, when compared to wild type mice (Fig 1E and 1F, S1D Fig). Treatment of infected C57BL/6 mice with neutralizing IFNγ antibody resulted in maintenance of RBC during the acute stage of the infection, confirming the result described above in IFNγR-/- mice (Fig 1G).
To determine the phagocytozing capacity of these distinct myeloid cell subsets in IFNγR-/- mice, a newly developed in vivo pHrodo-based erythrophagocytosis assay was used [19]. In this assay, RBCs are labeled with the acid-sensitive dye pHrodo ex vivo prior to i.v. injection in recipient mice. Following lysosomal uptake of labeled RBCs, phagocytozing cells become fluorescent (Fig 2A). The change in fluorescent intensity is then monitored between cells that have taken up unlabeled RBC and labeled RBC and expressed as delta median fluorescent intensity (ΔMFI) (Fig 2B). Liver neutrophils, monocytic cells, as well as spleen resident macrophages showed an increase in erythrophagocytozing capacity upon T. b. brucei infection in C57BL/6 ([19] and Fig 2C). In striking contrast, the erythrophagocytozing capacity of neutrophils drastically dropped in IFNγR-/- mice (Fig 2C and 2E). The erythrophagocytozing capacity of monocytic cells in IFNγR-/- mice was equal to that of wild type mice (Fig 2C). However, as IFNγR-/- mice display a remarkably reduced influx of monocyte-derived macrophages in the liver (Fig 1C and S1C Fig), the contribution of these cells to anemia development in IFNγR-/- mice could be minor. In contrast to wild type mice, liver resident macrophages of IFNγR-/- mice showed significant erythrophagocytozing potential (Fig 2C and 2E). Again, this resembles the situation of naïve mice, in which liver resident macrophages are the main cells involved in RBC uptake. In the spleen of infected wild type mice, neutrophils and resident macrophages showed the highest erythrophagocytozing potential (Fig 2D and 2E), whereas in IFNγR-/- mice the erythrophagocytozing capacity of these cell populations dropped significantly (Fig 2D and 2E). Spleen monocytes, which were present in higher amounts in IFNγR-/- mice compared to wild type mice (Fig 1D and 1E, S1D Fig), also showed a reduced erythrophagocytozing potential (Fig 2D). However, the erythrophagocytozing potential of monocyte-derived macrophages of IFNγR-/- mice was enhanced compared to that of wild type mice (Fig 2D and 2E). Yet, taking into account the low abundance of the monocyte-derived macrophages of IFNγR-/- mice compared to wild type mice, their contribution to acute anemia induction may be minor.
Using the in vitro pHrodo-erythrophagocytosis assay [19] we investigated the direct effect of IFNγ signaling on myeloid cells. In this setup, cells from uninfected, IFNγ-/- mice were incubated with labeled RBC and unlabeled RBC (background) in the presence or absence of IFNγ. Cellular composition and erythrophagocytosis potential was analyzed. As shown in Fig 2F, addition of IFNγ to cells from IFNγ-/- mice led to a shift in monocyte and monocyte-derived macrophage percentage. As the percentage of monocytes decreased, the percentage of monocyte-derived macrophages increased (Fig 2F). This suggests that IFNγ signaling could directly induce the differentiation of monocytes into monocyte-derived macrophages. Other surface markers should however be investigated to determine if these cells are indeed monocyte-derived macrophages. Fig 2G shows that upon incubation of naïve cells with RBC, IFNγ stimulation directly induces an up-regulation of the erythrophagocytic potential of neutrophils and monocyte-derived macrophages. This result suggests that neutrophils and monocyte-derived macrophages can directly alter their erythrophagocytic potential upon IFNγ signaling. For neutrophils it seems that IFNγ is a crucial inducer of erythrophagocytosis, as the erythrophagocytic potential is completely absent in neutrophils from IFNγ-/- mice (Fig 2G) and IFNγR-/- mice. This effect has been described before [26]. In contrast, monocyte-derived macrophages are still able to phagocytoze RBC in the absence of IFNγ, suggesting that IFNγ is not crucial for erythrophagocytosis by monocyte-derived macrophages.
Recently, we showed that during T. b. brucei infection the lipid composition of circulating RBC is altered, which coincided with an increase in susceptibility to lysis [19]. This in turn can contribute to enhanced RBC uptake and acute anemia. To investigate the role of IFNγ in the alteration of RBC membrane fragility, we performed a hemolysis experiment using resistance to osmolarity as a read-out. As indicated in Fig 3A, no difference in RBC osmotic fragility of naïve wild type and IFNγR-/- mice was observed. Upon T. b. brucei infection the same increase in osmotic fragility was observed for RBC of both infected wild type and IFNγR-/- mice (Fig 3B).
In conclusion, IFNγ appears to be indispensable for induction of the classical type 1 inflammation, leading to myeloid cell activation and recruitment, and resulting in enhanced erythrophagocytosis and acute anemia during early T. b. brucei infection. In addition, infection-associated alteration of RBC membranes is independent of IFNγ and its contribution to acute inflammation-associated anemia development seems to be minor.
It is generally accepted that IFNγ plays a key role in early stages of infection and CD8 and CD4 T cells have been indirectly suggested to be the sources of IFNγ during T. b. brucei infection [9,21,22,24]. To investigate the cellular source of IFNγ, IFNγ reporter (GREAT) mice were infected with T. b. brucei and liver and spleen were analyzed for IFNγ production (gating strategy S2A).
As previously shown [14,17], systemic IFNγ was present in quantifiable amounts at day 3 post infection (Fig 4A). This corresponded to approximately 2.6 x 106 IFNγ producing spleen cells and 0.9 x 106 IFNγ producing cells in the liver (Fig 4B). At the time of anemia induction at day 6 post infection, both spleen and liver had equal amounts of IFNγ producing cells (approximately 3.7 x 106) (Fig 4B). IFNγ production in GREAT mice was confirmed on the protein level in serum and spleen cell culture (S2B). Further detailed investigation into the cellular source of IFNγ showed that at day 3 post infection, NK cells were the dominant IFNγ producing population in the spleen, while in the liver both NK and NKT cells were the principal IFNγ producing cell populations (Fig 4C). By day 6 post infection, a shift occurred and CD4 and CD8 T cells were the majority of IFNγ producing cells in the spleen (Fig 4C). In the liver, the population of IFNγ producing NK and NKT cells also contracted, but to a lesser extent as in the spleen, and CD8 T cells become the dominant IFNγ producing cell population (Fig 4C). Upon investigation of the lymphocyte population dynamics in the spleen, a reduction in the amount of NK and NKT cells was observed by day 6 while an almost two-fold increase in CD8 T cells and CD4 T cells was observed (Fig 4D). In the liver an enormous expansion of NK cells occurred by day 3 post infection, while the NKT cell population quickly contracted (Fig 4D). CD8 T cells and CD4 T cells expanded approximately five-fold by day 6 post infection.
As this is the first time that NK and NKT cells have been implicated in T. b. brucei infection, we investigated possible activating mechanisms. For both NK and NKT cells, mechanisms of selective and non-selective activation have been described [27–29]. Investigation of serum cytokines indicated an early appearance of IL-12 and IL-15 (Fig 5A and 5B). While IL-15 generally drives NK cell proliferation and IFNγ production [30], IL-12 has been implicated in non-selective NKT cell activation [29]. Recent research into NK cell activation pathways have identified Stem cell antigen 1 (Sca-1) as a novel marker of non-selective NK cell activation. Investigation of surface markers of IFNγ producing NK cells showed that Sca-1, in addition to CD107a, is up regulated upon T. b. brucei infection (Fig 5C), indicating that they might be non-specifically activated.
Taken together, this data shows that both liver and spleen are important sources of IFNγ during T. b. brucei infection, and that NK and NKT cells are the earliest activated cells and IFNγ producers. Only by day 6 post infection, do CD8 and CD4 T cells become the dominant IFNγ sources.
As IFNγR-/- mice did not suffer as much from the acute anemia observed in wild type C57BL/6 mice, we examined the contribution of each IFNγ producing cell subset to acute anemia induction and systemic IFNγ levels. As results from the previous section indicated that NK and NKT cells were the earliest IFNγ producing cells during T. b. b. infection, we infected C57BL/6 mice depleted of NK1.1+ cells (S3A). In the absence of both NK and NKT cells, T. b. brucei infected anti-NK1.1 treated mice suffered less from acute anemia compared to control mice (Fig 6A). This reduced anemic phenotype coincided with reduced levels of IFNγ in serum and spleen cell cultures (Fig 6F and 6G). By day 6 post infection, CD8 and CD4 T cells seemed to become the dominant IFNγ producing cells. Infection of nu/nu mice, lacking both CD4 and CD8 T cells, resulted in a diminished anemia phenotype compared to C57BL/6 mice (Fig 6B), and coincided with reduced levels of IFNγ in serum and spleen cell culture (Fig 6F and 6G). To specify whether CD8 and CD4 T cells are equally important for acute anemia induction, anti-CD8-treated mice, CD8-/- mice and CD4-/- mice were infected with T. b. brucei. Both CD8-/- mice and anti-CD8 treated mice showed reduced anemia compared to wild type C57BL/6 mice (Fig 6C and 6D), coinciding with reduced levels of IFNγ in serum and spleen (Fig 6F and 6G). In contrast, CD4-/- mice presented a similar anemic phenotype as wild type mice (Fig 6E), and serum and spleen cell culture IFNγ levels were similar between CD4-/- and C57BL/6 mice (Fig 6F and 6G).
To confirm the role of IFNγ-producing CD8 T cells in the induction of acute anemia, isolated C57BL/6 CD8 T cells were adoptively transferred in CD8-/- mice and C57BL/6 nu/nu mice prior to infection (Annex S3 Fig). Adoptive transfer of CD8 T cells in these knock out mice resulted in the induction of anemia in these mice (Fig 7A and 7B). As a negative control, C57BL/6 nu/nu mice were reconstituted with CD4 T cells (Fig 7C), which did not reverse the reduced anemic phenotype.
In conclusion, together, these results demonstrated a crucial role for NK1.1+ and CD8 T cells, but not CD4 T cells, in the induction of acute anemia during T. b. brucei infection.
Recently it was shown that enhanced erythrophagocytosis by activated liver neutrophils and monocytic cells, as well as spleen resident macrophages is responsible for the induction of acute anemia during T. b. brucei infection [19]. Here, we elucidate the mechanism behind this phenomenon. We show that upon T. b. brucei infection, NK, NKT and CD8 T cells rapidly produce IFNγ, which recruits neutrophils and monocyte-derived macrophages to liver and spleen, activating them to phagocytoze RBCs and consequently induce acute anemia.
Trypanosome infections are known to induce inflammation and inflammation-associated pathology [6,8–10,12–16]. In murine T. b. brucei infection this is characterized by an early type 1 cytokine storm and the occurrence of acute anemia [14,17]. Previous research has established an important role for IFNγ in macrophage priming and consequent type 1 cytokine production [2,3,6,20]. We show that liver and spleen NK and NKT cells are the early sources of IFNγ in T. b. brucei infection. While it is commonly known that these innate lymphoid cells are permanently in a ‘pre-primed state’ which allows them to respond rapidly to multiple infections [27,31–34], this is the first evidence that these cells play a role in the regulation of T. b. brucei infection-associated inflammation. Upon murine T. b brucei infection, NK cells can get activated in a non-selective way [28]. However, for example during MCMV infection, NK cells have been reported to react in an antigen-specific way [27,28]. During infection, NKT cells could get activated by parasite-derived glycolipid antigens such as the glycosylphosphatidylinositol (GPI) anchor of VSG [35], however recent evidence showed that innate stimuli such as IL-12 and toll like receptors (TLRs) are a major mean of NKT cell activation [29]. In addition, studies of CMV infection have shown that NKT cells activate and enhance NK cell responses [32]. The exact mechanism of NK and NKT cell activation during Trypanosome infection, and to what extent they play a role in host protection, is however the subject of a different study. By day six post infection CD4 and CD8 T cells get activated and accumulate in liver in spleen. This coincides with a shift in IFNγ producing cells. In the liver CD8 T cells seem to take over IFNγ production while in the spleen CD8 and CD4 T cells both become the principal IFNγ producing cells. Antigen-specific T cell activation during T. b. brucei infection has been extensively described [9,10,36], and a T cell-dependent antibody response is crucial for control of the first parasitemia peak [37,38]. Non-specific activation of CD8 T cells has also been reported during T. b. brucei infection via a trypanosome-derived molecule called TLTF, which supposedly acts directly on CD8 T cells to induce IFNγ production [22]. IFNγ production during the early stage of T. b. brucei infection is essential for recruitment of myeloid phagocytic cells to liver and spleen. Indeed, in the absence of IFNγ the myeloid cell composition of liver and spleen closely resembles that of a naïve C57BL/6 mouse. IFNγ has been previously implicated in recruitment of TNF- and iNOS-producing Tip-DCs to liver during T. b. brucei infection [39]. The phenotype of these TIP-DCs closely resembles that of the monocyte-derived macrophages described here. In addition, recruitment of both cell types is CCR2-dependent (S3C Fig), indicating that these are most likely the same cells. However, due to the expression of F4/80 and MerTK by these cells, we favor the terminology monocyte-derived macrophages. Using the newly developed pHrodo assay, we showed that IFNγ is not only needed for the recruitment of monocytes and neutrophils to spleen and liver, but is also necessary to activate these cells as well as the resident macrophages of the spleen. Indeed, in the absence of IFNγ, these cells displayed a reduced phagocytozing potential. It must be mentioned that not only the liver myeloid cell composition of infected IFNγR-/- mice resembles that of naïve mice, but also the erythrophagocytozing potential of each cell subset. Indeed, similar to naïve mice, resident macrophages or Kupffer cells are the only cells that display erythrophagocytozing potential [19]. In the spleen of IFNγR-/- mice the monocyte-derived macrophages are the major cells that display erythrophagocytosis. However, given the small size of this cell population, the contribution to acute anemia induction could be minor. Using the in vitro approach to monitor erythrophagocytosis we show that IFNγ can directly induce an enhanced erythrophagocytic potential. This activating potential of IFNγ is common, e.g. in Toxoplasma gondii infection IFNγ has been shown to act directly on macrophages to induced enhanced phagocytosis of RBC [25]. Of note, the results presented here do not prove that the direct effect of IFNγ on the myeloid cells is the only manner of enhancing erythrophagocytosis. Other parameters such as low-grade inflammation in the absence of IFNγ could also play a role.
In contrast to CD4 T cell depletion, depletion of both NK and NKT cells or CD8 T cells conferred protection against anemia. This coincided with reduced local and systemic IFNγ levels, confirming that these cells are the major IFNγ producers during early infection. The reduced anemic phenotype upon NK1.1 depletion could indicate that these cells are necessary for CD8 T cell activation in a non-specific way. Alternatively, it could be that a certain threshold level of IFNγ needs to be reached and that therefore concomitant IFNγ production by NK NKT and CD8 T cells is needed.
Previously we showed that during T. b. brucei infection an alteration of RBC membrane occurs, which coincided with an enhanced fragility and erythrophagocytosis by myeloid phagocytic cells of both naïve and infected animals [19]. Here we showed that RBCs from infected IFNγR-/- mice are equally fragile as RBCs from infected wild type mice, indicating that this process occurs independently of IFNγ. The altered RBC fragility in IFNγR-/- mice could still prime them for more rapid phagocytosis.
The data presented in this paper show that in the absence of IFNγ, mice are protected from infection-associated acute anemia. It is interesting to mention that IFNγ-/- mice die within 20 days of infection, in contrast to wild type mice, which die around day 35 post infection [21], which could indicate that the protection against acute anemia is of no clinical significance. In contrast, the enhanced RBC clearance early in infection could even be a protective mechanism as it could be a mean of the host to diminish iron availability, hereby ‘starving’ the parasite and impeding its growth. This could be an explanation for the higher parasitemia peak in IFNγR-/- mice compared to wild type mice [21], however it was also reported that there is no correlation between parasite load and anemia induction [17], arguing against correlations between anemia and parasitemia.
In conclusion, this work describes the mechanism behind the induction of acute anemia during T. b. brucei infection. IFNγ derived from NK, NKT and CD8 T cells is crucial for the recruitment and activation of myeloid phagocytic cells in liver and spleen and consequently for the induction of acute anemia. Whether this mechanism can be extrapolated to other trypanosome infections inducing acute anemia (such as T. congolense and T. evansi) remains to be investigated.
|
10.1371/journal.pcbi.1006753 | Experimentally-constrained biophysical models of tonic and burst firing modes in thalamocortical neurons | Somatosensory thalamocortical (TC) neurons from the ventrobasal (VB) thalamus are central components in the flow of sensory information between the periphery and the cerebral cortex, and participate in the dynamic regulation of thalamocortical states including wakefulness and sleep. This property is reflected at the cellular level by the ability to generate action potentials in two distinct firing modes, called tonic firing and low-threshold bursting. Although the general properties of TC neurons are known, we still lack a detailed characterization of their morphological and electrical properties in the VB thalamus. The aim of this study was to build biophysically-detailed models of VB TC neurons explicitly constrained with experimental data from rats. We recorded the electrical activity of VB neurons (N = 49) and reconstructed morphologies in 3D (N = 50) by applying standardized protocols. After identifying distinct electrical types, we used a multi-objective optimization to fit single neuron electrical models (e-models), which yielded multiple solutions consistent with the experimental data. The models were tested for generalization using electrical stimuli and neuron morphologies not used during fitting. A local sensitivity analysis revealed that the e-models are robust to small parameter changes and that all the parameters were constrained by one or more features. The e-models, when tested in combination with different morphologies, showed that the electrical behavior is substantially preserved when changing dendritic structure and that the e-models were not overfit to a specific morphology. The models and their analysis show that automatic parameter search can be applied to capture complex firing behavior, such as co-existence of tonic firing and low-threshold bursting over a wide range of parameter sets and in combination with different neuron morphologies.
| Thalamocortical neurons are one of the main components of the thalamocortical system, which is implicated in key functions including sensory transmission and the transition between brain states. These functions are reflected at the cellular level by the ability to generate action potentials in two distinct modes, called burst and tonic firing. Biophysically-detailed computational modeling of these cells can provide a tool to understand the role of these neurons within thalamocortical circuitry. We started by collecting single cell experimental data by applying standardized experimental procedures in brain slices of the rat. Prior work has demonstrated that biological constraints can be integrated using multi-objective optimization to build biologically realistic models of neurons. Here, we employed similar techniques, but extended them to capture the multiple firing modes of thalamic neurons. We compared the model results with additional experimental data, test their generalization and quantitatively reject those that deviated significantly from the experimental variability. These models can be readily integrated in a data-driven pipeline to reconstruct and simulate circuit activity in the thalamocortical system.
| Thalamocortical (TC) neurons are one of the main components of the thalamus and have been extensively studied in vitro and in computo, especially in first order thalamic nuclei in different species [1]. One of these nuclei, namely the ventral posterolateral nucleus (VPL), relays somatosensory, proprioceptive, and nociceptive information from the whole body to the somatosensory (non-barrel) cortex [2]. The VPL is located close to ventral posteromedial nucleus (VPM), which transmits information from the face to the barrel cortex. The VPL and VPM nuclei constitute the ventrobasal (VB) complex of the thalamus [3].
Despite its key role in sensory functions, a systematic characterization of the cellular properties of the VB complex is still missing. The morphologies of VPL neurons in adult rats were described in early anatomical studies but were limited to two-dimensional drawings of Golgi-impregnated cells [4]. The general electrical properties of TC neurons maintained in vitro are known and similar in different thalamic nuclei and species with respect to the generation of two distinct firing modes, called tonic firing and low-threshold bursting [5–8]. However, a systematic description on the electrical types in the VB thalamus in the rodents is still missing.
Collecting morphological and electrophysiological data, by following standardized experimental procedures, is essential for the definition of cells types and it is the first step to constrain computational models of single neurons [9,10]. Although models of TC neurons have been published previously, they were typically aimed at studying specific firing properties and their parameters were hand tuned to achieve the desired result [11–15].
The purpose of our study is to systematically define the morphological and electrical types by collecting in vitro experimental data and to constrain biophysically detailed models of VB TC neurons of the juvenile rat. To the best of our knowledge, automatic parameter search has not been applied, thus far, to capture complex firing behavior in thalamic neurons, in particular low-threshold bursting and tonic firing. We defined the electrical and morphological types of TC neurons through in vitro patch-clamp recordings and 3D morphological reconstructions. We then extended an existing method [16] to account for their distinctive firing properties. These electrical models (e-models) were constrained by the electrical features extracted from experimental data [9,10,17,18]. Other experimental data were used to assess the generalization of the models to different stimuli and morphologies. We further performed a sensitivity analysis by varying each parameter at a time by a small amount and recording the resulting electrical features. This analysis provided an assessment of the robustness of the models and a verification that the selected features provided sufficient constraints for the parameters.
We characterized TC neurons in slices of the rat VB thalamus, by combining whole-cell patch-clamp recordings, biocytin filling and 3D Neurolucida (MicroBrightField) reconstruction, along with anatomical localization in a reference atlas [19] (Fig 1).
Visual inspection of 50 reconstructed morphologies (24 from the VPL, 26 from the VPM nuclei) revealed variability in the number of principal dendritic trunks and their orientation, in agreement with previous anatomical studies [4]. The maximum radial extent of the dendrites ranged between 120 and 200 μm and they started to branch between 20 and 50 μm from the soma (S1 Fig). We then analyzed the morphologies with two methods in order to quantitively classify different morphological types. We used algebraic topology to extract the persistent homology of each morphology and to visualize the persistence barcode [20] (Fig 2A, see Methods). Each horizontal bar in the persistence barcode represents the start and end point of each dendritic component in terms of its radial distance from the soma. The barcodes of all the morphologies followed a semi-continuous distribution of decreasing length. To quantify the differences between the barcodes, we computed the pairwise distances of the persistence images (see Methods and S1 Fig). We found that they were in general small (<0.4, values expected to vary between 0 and 1). These findings indicate that the morphologies cannot be grouped in different classes based on the topology of their dendrites. Furthermore, we performed Sholl Analysis [21] to compare the complexity of the dendritic trees (Fig 2B). We observed that all the morphologies had dense dendritic branches, with a maximum number of 50–100 intersections between 50–80 μm from the soma. When comparing the Sholl profiles for each pair of neurons we could not find any statistically significant difference (S1C Fig). Considering the results of topological and Sholl analyses, we grouped all the morphologies in one morphological type (m-type) called thalamocortical (TC) m-type.
We used an adaptive stimulation protocol, called e-code, consisting of a battery of current stimuli (see Methods for details), where the stimulation amplitude was adapted to the excitability of different neurons. This standardized protocol has previously been used to build biophysically-accurate models of cortical electrical types (e-types) [16]. However, TC neurons from different thalamic nuclei and species fire action potentials in two distinct firing modes, namely tonic firing, when stimulated from a relatively depolarized membrane potential or low-threshold bursting, from a hyperpolarized membrane potential [5–8]. We thus extended the e-code to include two different holding currents. All the neurons recorded in this study displayed tonic and burst firing, when stimulated with the appropriate holding current (Fig 1B). Moreover, we were able to classify different e-types by considering the voltage traces recorded in tonic mode in response to step current injections (Fig 1B). The majority of the cells (59.3%) showed a non-adapting tonic discharge (continuous non-adapting low-threshold bursting, cNAD_ltb e-type) while others (40.7%) had higher adaptation rates (continuous adapting low-threshold bursting, cAD_ltb e-type), as reflected by the adaptation index (Fig 1C). We followed the Petilla convention [22] for naming the tonic firing discharge (cNAD or cAD), extending it to include “_ltb” for the low-threshold bursting property. In some rare examples, we noticed acceleration in the firing rate with decreasing inter-spike intervals (ISIs) towards the end of the stimulus. Similar adapting and accelerating responses have already been described in the VB thalamus of the cat [7]. We also observed stereotypical burst firing responses within the same cell, with variation of the number of spikes per burst in different cells, but the burst firing responses alone were insufficient to classify distinct e-types.
Multi-compartmental models comes with the need of tuning a large number of parameters [23], therefore we constrained the models as much as possible with experimental data. We first combined the morphology and the ionic currents models in the different morphological compartments (soma, dendrites and axon). Given that the reconstruction of the axon was limited, we replaced it with a stub representing the initial segment [16]. We used previously published ionic current models and selected those that best matched properties measured in rat TC neurons (see Methods). The kinetics parameters were not part of the free parameters of the models.
The distribution of the different ionic currents and their conductances in the dendrites of TC neurons is largely unknown. The current amplitudes of the fast sodium, persistent and transient (A-type) potassium currents were measured, but only up to 40–50 μm from the soma [24]. Indirect measures of burst properties [15] or Ca2+ imaging studies [25] suggest that the low-threshold calcium (T-type) channels are uniformly distributed in the somatodendritic compartments. We thus assumed different peak conductance in the soma, dendrites and axon for all the ionic currents, except for ICaT, which had the same conductance value in the soma and dendrites. We then extracted the mean and standard deviation (STD) of different electrical features in order to capture the variability of firing responses from different cells of the same e-type [9,10,17] (Fig 3). We observed that some features extracted from tonic firing responses had distinct distributions between the cAD_ltb and cNAD_ltb e-types (Fig 3A).
For optimizing the models’ parameters, we chose features that quantified passive (input resistance, resting membrane potential), burst and tonic firing properties (number of spikes, inverse of inter-spike intervals, latency to first spike, adaptation index), action potentials shape (amplitude, half-width, depth of the fast after-hyperpolarization). We aimed at finding the minimal set of features that captured the most important properties in the two firing modes. This set was a trade-off between comprehensively describing the experimental data (i.e. extracting all possible features), which can lead to over-fitting and loss of generalizability, and a too small set that would miss some important characteristics. For the tonic firing responses, we used three stimulation amplitudes (150%, 200%, 250% of firing threshold) which have been shown to reproduce the complete input-output function of the neurons [16,17]. Responses to two hyperpolarizing steps of different amplitudes (−40% and −140% threshold) constrained the input resistance (conductance of the leak current) and the conductance of currents activated in hyperpolarization, for example the h-current, Ih (sag_amplitude feature). We included baseline voltage values to the optimization objectives to ensure that the models were in the right firing regime and spike count to penalize models that were firing in response to the holding currents. To constrain the low-threshold burst we used features (such as number of spikes) which are influenced by specific ionic currents, for example the low-threshold calcium current, ICaT.
The average value and STD of each feature were used to calculate the feature errors (Fig 4C). Each error measured how much the features of the models deviated from the experimental mean, in units of the experimental STD. We used a multiobjective optimization approach (MOO), where each error was considered in parallel. To rank the resulting models after optimization, we considered model A better than model B if the maximum error of all the features of A was smaller than the maximum error of all the features of B.
By applying this MOO procedure, we generated multiple models with distinct parameter combinations for each of the twenty-two free parameters (Fig 4B). The free parameters were allowed to vary between the upper and lower bounds shown in Fig 5B. The models reproduced well the key firing dynamics observed in the experimental recordings. They showed a low-threshold burst when stimulated from a hyperpolarized membrane potential, crowned by a variable number of sodium spikes (Fig 4B). In the tonic firing regime, they reproduced adapting and non-adapting firing discharges as observed in the two e-types. These results indicate that the ion channels included in the models were sufficient to reproduce the experimental firing properties and that different e-types in TC neurons could be generated by different ion channel densities.
We found that different sets of parameter values reproduced the target firing behavior (Fig 5B). We further analyzed models that had all the feature errors below 3 STD. Models’ voltage responses reflected the characteristic firing properties of TC neurons (S3 Fig), indicating that the selected set of features and ion channels were sufficient to capture the two firing modes, in both the adapting and non-adapting e-types. The voltage traces from different models showed small differences in spike amplitude, firing frequency, and depth of the after-hyperpolarization, as reflected by the variability of features values (Fig 5C), arising from differences in ion channel densities between models.
Spike-shape related features (e.g. AP. amplitude) in the different models covered the space of the experimental variability, while for some features (e.g. input resistance, Rinput), all models tended to cluster on one of the tails of the experimental distribution. Rinput relates to the neuron passive properties and depends both on the number of channels open at rest (inverse of the leak conductance in the model) and the size of the cell. Given that all the models for a given e-type were constrained on a single morphology, this result is not surprising. Other features, such as sag amplitude were less variable in the models compared to experiments. We hypothesized that this depended on the variable stimulation amplitudes applied to different experimental cells, while all the models were stimulated with the same current amplitudes.
Some other features were systematically above or below the experimental values in both e-types. We suggest that this depend on the exact dynamics of some specific ion channels. For example, the amplitudes of the first and second spikes in the burst tended to be similar or above and below the experimental values, respectively. This can depend on the specific activation/inactivation properties of some ionic currents, for example the transient sodium current (INaT) and delayed potassium current (IKd). During the rising phase of the low-threshold spike, INaT in the model is readily activated and generated a first spike with higher amplitude, but is not repolarized enough by the activation of IKd. At higher potentials, reached towards the peak of the low-threshold spike, the availability of INaT and other depolarizing currents seemed reduced and generated a spike with smaller amplitude. Sensitivity analysis confirmed that INaT and IKd had an influence on the amplitude of the first and second spike in the burst. Furthermore, these two currents operate together with currents that generate the burst, such as the low-threshold calcium current (ICaT) and the Ih in shaping the amplitude of the second spike in the burst. Interestingly, the models also tended to have lower instantaneous frequency of the first two spikes in the burst (Inv. 1st ISI) and this feature had similar sensitivity (but of opposite signs) to the amplitude of the second spike in the burst.
Another possible explanation is the lack of some ionic currents in the model, for example some specific subtype of potassium channels that promote higher firing rates (Kv3.1 and Kv3.3). While neurons of the thalamic reticular nucleus are known to express this channel subunit [26], the expression in TC neurons has not been confirmed yet. The dynamics of IKd could also explain why the after-hyperpolarization (AHP depth) tended to be smaller in the models compared to the experimental values. AHP depth is also influenced by other ionic currents, such as high-threshold calcium current (ICaL), calcium-activated potassium current (ISK) and the intracellular calcium dynamics. The number of action potentials (Num. of APs) in different conditions (No stim, Ihold) ensured that the models did not spike in the absence of a stimulus or in response to the holding current. For this reason, all the experimental and model feature values in Fig 5C are equal to 0.
We examined the diversity of the parameter values with respect to the initial parameter range (Fig 5B). Most of the optimized parameter values spanned intervals larger than one order of magnitude. On the other hand, some parameter values were restricted to one order of magnitude, for example the permeability of the low-threshold calcium current PCaT. This result is in agreement with experiments showing a minimum value of ICaT is critical to generate burst activity and this critical value is reached only at a certain postnatal age [27]. The value of PCaT was constrained by features measuring burst activity (such as number of spikes, frequency, etc.).
We used different stimuli for model fitting (current steps) and for generalization assessment (current ramps and noise). We simulated the experimental ramp currents, by stimulating the models with the appropriate holding currents for the two firing modes and a linearly increasing current. We first compared visually the model responses with the experimental recordings (Fig 6A). In burst mode, the models reproduced the different behaviors observed experimentally: absence of a burst, small low-threshold spike, burst (S4A Fig). Moreover, the latency of burst generation substantially overlapped with the experimental one. However, a small fraction of models (1.2%) generate repetitive burst that we have never observed in the experimental recordings (S4B Fig). These models were quantitatively rejected by considering the number of spikes and the inter-spike intervals. In tonic mode, the latency to first spike, the voltage threshold, the shape of the subsequent action potentials and the increase in firing frequency were comparable with the experimental recordings (Fig 6A). In addition, we quantified the generalization error to ramp stimuli (Fig 6C), by considering the latency to first spike, firing frequency increase over time (tonic mode) or number of spikes (burst mode).
Although conductance-based models can be fit by using step and ramp currents, these stimuli are different from synaptic inputs, which can be simulated by injecting noisy currents. To test the response to such network-like input, we used a noisy current varying accordingly to an Ornstein-Uhlenbeck (OU) process [28] to compare models’ responses with the experimental data. Each experimentally recorded cell was stimulated with the same OU input, scaled by a factor w. Experimentally, w was calculated by evaluating the responses to previous stimuli. We developed a similar approach to generate the noise stimuli in silico (see Methods). The noise current was injected on top of the holding currents used during the optimization. We found that the models reproduced well the subthreshold potential, spike times and the distribution of single spikes and bursts (Fig 6B). Moreover, we quantitatively evaluated the generalization to the noise stimulus by extracting features (e.g. number of spikes) and comparing them with the experimental mean.
We computed generalization errors for each model, which were calculated similarly to the optimization errors (Fig 6C). We considered a model acceptable after generalization if it had all generalization errors <3 STD and we found that the majority of the models (>90%) passed the generalization test (Fig 6D).
We assessed the robustness of the models to small changes in their parameter values. To that end, we varied each parameter at a time by a small amount (± 2.5% of the optimized value) and computed the values of the features. A sensitivity value of 2 between parameter p and feature y means that a 3% change in p caused a 6% change in f. We ranked the parameters from the most to the least influential and the features from the most sensitive to the least sensitive.
Some features resulted to be more sensitive to parameter changes, both in term of magnitude of the sensitivity and number of parameters (e.g. adaptation index, inverse of inter-spike intervals, ISIs, AHP depth). Most of these features describe the model firing pattern, which depend more on the interplay between the different ionic currents than on the specific activation/inactivation dynamics. Conversely, spike shape-related features were less sensitive to parameter changes (e.g. AP half-width, AP amp.) because they depend more on specific ionic current dynamics (e.g. IKd, IL, INaT,). Some features were very weakly influenced by small parameter changes, e.g. baseline voltage, which depend more on the holding current amplitude, than on the model parameters (Fig 7A).
The conductance of the leak current gleak emerged as the most influential parameter (Fig 7A). An increase in gleak caused a decrease in firing frequency (inverse of ISIs) in both the tonic and burst firing modes. These results are easy to interpret when considering Ohm’s law: increasing gleak means decreasing the input resistance of the model, so that for the same input current the voltage response becomes smaller. The second most influential parameter was the conductance of the persistent sodium current gNaP in the dendrites, which increased the tonic firing rate as expected from a depolarizing current. Interestingly, gNaP had an effect on the late phase of the low-threshold burst (inverse last ISI—burst), suggesting that the low-threshold burst is initiated by the activation of IT and modulated by INaP. An increase in the permeability of the low-threshold calcium current PCaT, known to be one of the main currents underlying low threshold bursting, enhanced burst firing responses (it increased the inverse of ISIs, i.e. the firing frequency) and had effects on some of the tonic features. Increasing the somatic permeability of the high threshold calcium current PCaL decreased the tonic firing rate, despite being a depolarizing current. Increasing PCaL means higher Ca2+ influx and higher activation of the Ca2+-activated potassium current (ISK). The parameter gSK had indeed a similar effect on the features and thus clustered together with parameters regulating the intracellular calcium dynamics γCa and τCa (Fig 7B). Sag amplitude, that is known to depend on the activity of Ih, was mainly influenced by change in gleak, PCaT and gh. In summary, each parameter influenced at least one feature. These results indicate that the model ability to generate tonic and burst firing is robust to small changes in parameter values and that all the parameters were constrained during the optimization by one or more features.
We then analyzed which features depended similarly on parameter changes, as they may add superfluous degrees of freedom during parameters search. Fig 7B shows the same sensitivities as in Fig 7A, clustered by their similarities (see Methods). Features clustered together if they were sensitive to similar parameter combinations and parameters clustered based on their similar influence on the features. Not surprisingly, the same tonic features measured at different level of current stimulation clustered together (e.g. AP amplitude and half-width, AHP depth, latency of the first ISI) and tonic firing features belonged to a cluster that was different from burst features. Some features measured in tonic mode (such as AP half-width and AP amp.) clustered together because they depended mainly on the dynamics of INaT and IKd: increasing the conductance of INaT increased the amplitude of the APs and decreased its duration. This was also true for the amplitude of the 1st AP in the burst. Features measured in burst mode had similar sensitivities because they depend on currents that are active at relatively hyperpolarized potential (such IH and ICaT).
We optimized the parameters for the adapting and non-adapting e-models in combination with two different experimental morphologies selected at random and then tested them with the other 48 morphologies. Considering that morphologies could not be classified in different m-types based on topological analysis of their dendrites and that TC neurons have been shown to be electrically compact [15], we expected the electrical behavior to be conserved when changing morphology. Nonetheless, different neurons vary in their input resistance Rinput and rheobase current Ithr due to variation in the surface area. Variation in Rinput and Ithr made the current amplitude applied during the optimization inadequate to generate the appropriate voltage trajectories. We thus devised an algorithm to search for the holding current to obtain the target holding voltage (for example −64 mV or −84 mV for tonic and burst firing, respectively) and Ithr from the desired holding voltage. The different e-model/morphology combinations (me-combinations) were evaluated by computing the same feature errors calculated during optimization (Fig 8A). For each morphology, we selected the e-model that generated the smallest maximum error. We chose the value of 3 STD as the threshold to define which me-combinations were acceptable [29], yielding 48 acceptable me-combinations out of the 48 tested (Fig 8A). All me-combinations reproduced burst and tonic firing (Fig 8B).
Given that the generalization of the electrical models to the other 48 morphologies worked well, we can conclude that the morphological properties of the modeled neurons are very similar, at least for properties that have an impact on the electrical models (e.g. surface area, diameters of the compartments).
Our objective was to apply and extend an existing data-driven pipeline to identify the cell types and build models of VB thalamocortical neurons that reproduced the multiple firing modes that have been experimentally observed. We successfully modelled these novel firing types, by including additional stimulation protocols and features to constrain the low-threshold burst.
Our morphological and electrical data were used to define the properties of VB TC neurons in the rat. We found two electrical types (e-types) of TC neurons, but no objectively different morphological types (m-types) were revealed either using Sholl analysis [21] or topological analysis of dendritic branching [20]. We cannot exclude that refinements to these methods will reveal different m-types similar to the ones described in the visual thalamus of the mouse [30]. We also showed that automatic parameter search can be applied to build biophysically and morphologically detailed models. This method was already applied to model canonical firing behavior in cortical [9,10,16,17], hippocampal [31], cerebellar granule neurons [32] and corticospinal neurons [33]. To the best of our knowledge, such an automatic parameter search has not been used previously to capture different firing modes and complex firing behavior such as low-threshold bursting in thalamic neurons. Standardized electrophysiological protocols allowed us to identify for the first time in juvenile rat adapting and non-adapting e-types of TC VB neurons that were previously observed in other species [7]. This finding suggests that the intrinsic properties of TC neurons contribute to adaptation, a key phenomenon for filtering out irrelevant stimuli, before sensory information reaches the neocortex. Further experiments are needed to elucidate the relative contribution of intrinsic mechanisms and network properties to adaptation in somatosensory systems. We named the two main e-types continuous non-adapting low-threshold bursting (cNAD_ltb) and continuous adapting low-threshold bursting (cAD_ltb) by following and extending existing conventions [16,22,31].
In this study, we improved upon previous morphologically and biophysically detailed models of tonic and burst firing in TC neurons [12,13,15] by explicitly constraining the parameters with experimental data, without hand-tuning of their values. Unlike previous models, we chose a multi-objective optimization for a methodological and a scientific reason: it is more time-efficient, reproducible, and it approximates the variability in ionic channel expression of biological neurons [31,34–36], as shown by the family of acceptable solutions we found. However, experiments aimed at quantifying ion channel conductances are essential to assess if these solutions fall between biological ranges. Furthermore, we tested the generalization capability of the models and found that more than 90% of the models were comparable with the experimental data.
Nonetheless, we noticed some inaccuracies when comparing the voltage traces with the experimental data when assessing the generalization of the models. For instance, some models tended to generate small transient oscillations in response to ramp stimuli in burst mode. This result is not surprising, considering that the exact kinetics for all the ionic currents are not available and that there are known limitations in models of ionic channels derived from the literature or from other models [37,38]. In particular, modifications to the kinetics of the low-threshold calcium current was shown to explain the propensity to generate oscillatory bursts in TC neurons of other nuclei and species [39]. More generally, we included ion channels that were used in previous models and that were validated with experimental data whenever possible. We undertook an extensive literature review to use channel kinetics derived from recordings in rat TC neurons from the ventrobasal (VB) thalamus or other first-order thalamic nuclei, whenever the data was available (see Methods). Moreover, we cannot exclude that some ionic currents were missing from our models and that they could have improved their fitness.
TC neurons have been shown to be electrically compact [15] and could, in principle, be modeled as a single compartment. However, active mechanisms need to be located in the dendrites in order to ensure synaptic integration and amplification [40]. Information regarding specific conductances or firing properties in the dendrites of TC neurons is limited. For this reason, dendritic parameters in our models may be underconstrained. However, the sensitivity analysis revealed that dendritic parameters did not appear to be the least constrained because they influenced different tonic and burst-related features.
We included in the model fitting and validation pipeline a sensitivity analysis, which is often neglected in computational neuroscience [41]. Although we cannot use our simple univariate approach to explore multidimensional parameter correlations and principles of co-regulation of ion channels expression, it is useful to find better constraints for parameters optimization. The selection of the features is indeed a step that still requires care and experience by modelers. Furthermore, this type of sensitivity analysis allows to identify parameters that can be traded-off during the optimization and that can be removed in order to reduce the dimensionality of the problem. In our study, parameters related to the calcium dynamics were shown to influence the features in a very similar fashion. This type of analysis is of particular importance in future work aimed at using the full diversity of ion channels that can be inferred from gene expression data. Gene expression data could also provide additional constraints on the choice of ion channels and indicate the ones that are missing in our models. More in detail, we propose that sensitivity analysis should be a fundamental tool in selecting which conductances are successfully optimized by the available experimental constraints. The example we showed is a local approach, applied to a specific solution to the optimization problem, which showed that our models are robust to small parameter changes. This analysis can be extended to study how the sensitivities vary in the neighborhood of different solutions.
In conclusion, we systematically studied the morphological and electrical properties of VB TC neurons and used these experimental data to constrain single neuron models, test their generalization capability and assess their robustness. Further work will validate these models in response to synaptic activity, in order to include them in a large-scale model of thalamocortical microcircuitry [16].
Experimental data were collected in conformity with the Swiss Welfare Act and the Swiss National Institutional Guidelines on Animal Experimentation for the ethical use of animals. The Swiss Cantonal Veterinary Office approved the project following an ethical review by the State Committee for Animal Experimentation.
All the experiments were conducted on coronal or horizontal brain slices (300 μm thick- ness) from the right hemisphere of male and female juvenile (P14-18) Wistar Han rats. The region of interest was identified using the Paxinos and Watson rat brain atlas [19]. After decapitation, brains were quickly dissected and sliced (HR2 vibratome, Sigmann Elektronik, Germany) in ice-cold standard ACSF (in mM: NaCl 125.0, KCl 2.50, MgCl2 1.00, NaH2PO4 1.25, CaCl2 2.00, D-(+)-Glucose 50.00, NaHCO3 50.00; pH 7.40, aerated with 95% O2 / 5% CO2). Recordings of thalamocortical neurons in the VB complex were performed at 34°C in standard ACSF with an Axon Instruments Axopatch 200B Amplifier (Molecular Devices, USA) using 5–7 MΩ borosilicate pipettes, containing (in mM): K+-gluconate 110.00, KCl 10.00, ATP-Mg2+ 4.00, Na2-phosphocreatine 10.00, GTP-Na+ 0.30, HEPES 10.00, biocytin 13.00; pH adjusted to 7.20 with KOH, osmolarity 270–300 mOsm. Cells were visualized using infrared differential interference contrast video microscopy (VX55 camera, Till Photonics, Germany and BX51WI microscope, Olympus, Japan).
Membrane potentials were sampled at 10 kHz using an ITC-18 digitizing board (InstruTECH, USA) controlled by custom-written software operating within IGOR Pro (Wavemetrics, USA). Voltage signals were low-pass filtered (Bessel, 10 kHz) and corrected after acquisition for the liquid junction potential (LJP) of −14 mV. Only cells with a series resistance <25 MΩ were used.
After reaching the whole-cell configuration, a battery of current stimuli was injected into the cells and repeated 2–4 times (e-code). During the entire protocol, we defined offset currents in order to keep the cell at −50 mV (tonic firing) or −70 mV (burst firing) before LJP correction and applied them during the entire protocol. The step and ramp currents were injected with a delay of 250 ms in the experiment. In the models, the stimuli were injected with a delay of 800 ms, to allow for the decay of transients due to initialization. Each stimulus was normalized to the rheobase current of each cell, calculated on-line as the current that elicited one spike (stimulus TestAmp, duration 1350 ms). The stimuli used in the experiments, for fitting and testing the models were:
Neurons that were completely stained and those with high contrast were reconstructed in 3D and corrected for shrinkage as previously described [16]. Reconstruction used the Neurolucida system (MicroBrightField). The location of the stained cells was defined by overlaying the stained slice and applying manually an affine transformation to the Paxinos and Watson’s rat atlas [19].
Electrical features were extracted using the Electrophys Feature Extraction Library (eFEL) [42]. We calculated the adaptation index (AI) from recordings in tonic mode (Step 200% threshold) and classified TC VB neurons into adapting (AI> = 0.029) and non-adapting (AI<0.029) electrical types. AI was calculated using the eFEL feature adaptation_index2 and corresponded to the average of the difference between two consecutive inter-spike intervals (ISI) normalized by their sum. The cut-off value was calculated after fitting a Gaussian mixture model to the bimodal data, using available routines for R [43,44]. In order to group data from different cells and generate population features, we normalized all the stimuli by the rheobase current Ithr of each cell. To calculate Ithr, we used IDRest and IDThresh and selected the minimal amplitude that evoked a single spike. Along with the voltage features, we extracted mean holding and threshold current values for all the experimental stimuli. Description of the features and the details on their calculation are available on-line [42]. Current stimuli applied during the optimization and generalization were directly obtained from the experimental values or automatically calculated by following the experimental procedures (e.g. noise stimulus).
Reconstructed morphologies were analyzed to objectively identify different morphological types. The Sholl profiles of each pair of cells was statistically tested by using k-samples Anderson-Darling statistics. This test was preferred to the most common Kolmogorov-Smirnov test, because it does not assume that the samples are drawn from a continuous distribution. The different Sholl profiles are indeed an analysis of the intersections with discrete spheres.
To compare the topological description of each morphology we transformed the persistence barcodes into persistence images and calculated their distances as in [20]. Briefly, we converted the persistence barcode, which encodes the start and end radial distances of a branch in the neuronal tree, into a persistence diagram. In the persistence diagram, each bar of the barcode is converted into a point in a 2D space, where the X and Y coordinates are the start and end radial distances of each bar. The persistence diagram was then converted into a persistence image by applying a Gaussian kernel. We used the library NeuroM [45] to perform Sholl and morphometrics analyses.
We used Hodgkin-Huxley types of ionic current models, starting from kinetics equations already available in the neuroscientific literature. Along with kinetics of the ionic currents, we stored information on the experimental conditions, such as temperature and LJP, by using the software NeuroCurator [46]. Whenever the data was available, we compared simulated voltage-clamp experiments to experimental data from juvenile rats. Ionic currents Ii were defined as functions of the membrane potential v, its maximal conductance density gi and the constant value of the reversal potential Ei:
Ii=gimixhiy(v−Ei)
mion and hion represent activation and inactivation probability (varying between 0 and 1), with integer exponents x and y. Each probability varied according to:
n′(v)=(n∞(v)−n)/τn(v)
where n∞(v) is a function of voltage that represents the steady-state activation/inactivation function (normally fitted with a Boltzmann curve) and τn(v) is a voltage-dependent time constant. Exceptions to this formalism are ionic currents that do not inactivate (y = 0) and ionic currents with (in)activation processes mediated by two or more time constants. Calcium currents (ICaT and ICaL) were modeled according to the Goldman-Hodgkin-Katz constant field equation and had permeability values instead of conductance [47].
NEURON 7.5 software was used for simulation [56]. We used NEURON variable time step method for all simulations. For the sake of spatial discretization, each section was divided into segments of 40 μm length. The following global parameters were set during optimization and generalisation: initial simulation voltage (−79 mV), simulation temperature (34°C), specific membrane capacitance (1 μF/cm2), specific intracellular resistivity 100 Ωcm for all the sections, equilibrium potentials for sodium and potassium were 50 mV and −90 mV, respectively.
BluePyOpt [18] with Indicator Based Evolutionary Algorithm (IBEA) were used to fit the models to the experimental data. Each optimization run was repeated with three different random seeds and evaluated 100 individuals for 100 generations. The evaluation of these 300 individuals for 100 generations was parallelized using the iPython ipyparallel package and took between 21 and 52 h on 48 CPU cores (Intel Xeon 2.60 GHz) on a computing cluster. Each optimization run typically resulted in tens or hundreds of unique acceptable solutions, defined as models having all feature errors below 3 STD from the experimental mean.
We performed a sensitivity analysis of an optimization solution by varying one parameter value (pm) at a time and calculating the electrical features from the voltage traces (y+ and y-). We defined the sensitivity as the ratio between the normalized feature change and the parameter change, which for smooth functions approximates a partial derivative [57,58]. The features changes were normalized by the optimized feature value. For small changes of parameter values, we assumed that the features depend linearly on its parameters. We could thus linearize the relationship between the features and the parameters around an optimized parameter set and calculate the derivatives. The derivatives were calculated with a central difference scheme [57].
We collected the derivatives (sensitivities) in the N X M Jacobian matrix, with N representing the number of features and M the number of parameters.
To rank parameters and features we computed their relative importance by calculating their norms (the square root of the summed squared values) from the Jacobian columns and rows, respectively. To cluster parameters based on similar influences on the features and to cluster features that were similarly dependent on the parameters, we used angles between columns (or rows) to compute distances D between parameters (or features):
D=1−|cosθ|
Features where thus considered similar if they depended in a similar manner on the parameters, independent of sign or magnitude.
|
10.1371/journal.ppat.1005140 | Experimental Malaria in Pregnancy Induces Neurocognitive Injury in Uninfected Offspring via a C5a-C5a Receptor Dependent Pathway | The in utero environment profoundly impacts childhood neurodevelopment and behaviour. A substantial proportion of pregnancies in Africa are at risk of malaria in pregnancy (MIP) however the impact of in utero exposure to MIP on fetal neurodevelopment is unknown. Complement activation, in particular C5a, may contribute to neuropathology and adverse outcomes during MIP. We used an experimental model of MIP and standardized neurocognitive testing, MRI, micro-CT and HPLC analysis of neurotransmitter levels, to test the hypothesis that in utero exposure to malaria alters neurodevelopment through a C5a-C5aR dependent pathway. We show that malaria-exposed offspring have persistent neurocognitive deficits in memory and affective-like behaviour compared to unexposed controls. These deficits were associated with reduced regional brain levels of major biogenic amines and BDNF that were rescued by disruption of C5a-C5aR signaling using genetic and functional approaches. Our results demonstrate that experimental MIP induces neurocognitive deficits in offspring and suggest novel targets for intervention.
| A growing body of evidence has established the importance of the in utero environment on neurodevelopment and long-term cognitive and behavioral outcomes. These data suggest factors that disrupt the tightly regulated in utero environment can modify normal neurodevelopmental processes. Approximately 125 million pregnancies worldwide are at risk of malaria infection every year. However the impact of in utero exposure to MIP on fetal neurodevelopment is unknown. Here we use a mouse model of malaria in pregnancy to examine the impact of maternal malaria exposure on neurocognitive outcomes in offspring. We observed impaired learning and memory and depressive-like behavior in malaria-exposed offspring that were neither congenitally infected nor low birth weight. These neurocognitive impairments were associated with decreased tissue levels of neurotransmitters in regions of the brain linked to the observed deficits. Disruption of maternal C5a complement receptor signaling restored the levels of neurotransmitters and rescued the associated cognitive phenotype observed in malaria-exposed offspring. This study provides the first evidence implicating a causal link between pre-natal exposure to malaria, complement signaling and subsequent neurocognitive impairment in offspring.
| Each year, an estimated 125 million pregnancies worldwide are at risk of malaria infection [1]. Plasmodium falciparum infections during pregnancy are more frequent, and associated with higher parasite burdens and worse clinical outcomes than those of non-pregnant individuals [2,3]. MIP has profound maternal and fetal health consequences including increased risk of maternal anemia, preterm birth, stillbirth, fetal growth restriction (FGR) and low birth weight infants (LBW), resulting in an estimated 200,000 infant deaths annually [4]. MIP is characterized by the accumulation of parasitized erythrocytes (PEs) and monocytes/macrophages in the placenta [2,3]. While it is believed that this localized placental immune response contributes to adverse birth outcomes, the precise mechanism by which parasite and monocyte accumulation in the placenta results in poor pregnancy outcomes remains unknown. Recent evidence supports a role for altered angiogenesis and resulting placental vascular insufficiency [5,6].
The complement system is an essential component of the innate immune response to microbial pathogens [7–9]. Excessive complement activation, notably generation of the anaphylatoxin C5a, has been implicated in mediating deleterious host responses and poor clinical outcomes to infections [8,10]. Malaria infection is known to induce activation of the complement system through multiple pathways, and recent studies support a mechanistic role for C5a in the pathophysiology of severe malaria and malaria in pregnancy [10–14]. Complement activation has also been proposed as a common pathway mediating adverse pregnancy outcomes in the absence of infection [15,16]. Excessive C5a generation was implicated as a mediator of placental injury in murine models of spontaneous miscarriage and FGR [17]. Moreover, human studies have associated complement split products (e.g. C3a, C5a) with pregnancy complications [18,19].
Recent evidence has also identified an essential role for the complement system in both normal and abnormal neurodevelopmental processes [20–22]. Complement proteins and their receptors are widely expressed within the central nervous system and play a major role in regulating normal synaptic development and function [23].
Alterations in the in utero environment as a result of maternal infection may have profound and long-term implications for the developing fetus. Recent studies indicate that immunological stress at the maternal-fetal interface can alter later-life brain development and behaviour [24,25]. Despite the potential public health implications, little is known about the impact of in utero exposure to MIP on fetal and infant neurological development. Based on the above evidence implicating C5a in both neurodevelopment and MIP-associated adverse birth outcomes, we tested the hypothesis that in utero exposure to experimental MIP (EMIP) alters offspring neurodevelopment and that disruption of maternal C5a receptor (C5aR) signaling would rescue EMIP-induced neurocognitive injury in exposed offspring.
LBW, as a result of preterm birth or FGR, is known to be associated with impaired neurocognitive development [26,27]. Since MIP may cause LBW, these infants would be expected to experience an increased risk of neurocognitive impairment; however the majority of fetuses exposed to malaria in utero do not develop LBW. Therefore, in order to avoid LBW as a confounder and isolate the effects of malaria exposure alone on offspring neurodevelopment, we reduced the inoculum given to dams in a validated model of EMIP [28] from 106 to 105 PEs. This inoculum was associated with the presence of parasitized erythrocytes in the placenta and localized inflammation in the placenta (S1 Fig and S2 Fig). However the 105 inoculum was associated with lower maternal peripheral parasitemia (Fig 1A) and less marked placental pathology than that previously reported with a dose of 106 PEs [28]. This modification eliminated the LBW phenotype in this model and resulted in equivalent birth weights (from 1 to 20 weeks of age) in control pups compared to offspring exposed in utero to EMIP (Figs 1b, 5b and S3 Fig, S4 Fig and S5 Fig). No significant differences were observed in the length of gestation or litter size in this lower inoculum EMIP model (S1 Table). Placentas from malaria-infected litters (wild-type and C5ar-/-) showed placental inflammation as indicated by increased expression of tumor necrosis factor (TNF), interferon gamma (IFNϒ), intracellular adhesion molecule-1 (ICAM-1) and monocyte chemotactic protein 1 (MCP-1, CCL2) (S2 Fig, p < 0.05). Wild-type mice showed increased expression of ICAM and reduced expression of MCP in comparison with C5ar-/- mice in placentas from both uninfected and malaria-infected litters (S2 Fig, p < 0.05). Absence of congenital infection was confirmed by blood smears and PCR of fetal blood.
To investigate the impact of in utero EMIP-exposure on neurocognitive performance, we compared EMIP-exposed pups to unexposed controls using a battery of standardized neurocognitive tests [29–31]. Exposed offspring showed impaired novel object recognition (NOR) in the NOR test of non-spatial learning and memory, and increased immobility in the tail suspension test (TST), a test of depressive-like behavior. Performance in the NOR test was impaired in EMIP-exposed offspring compared with unexposed offspring (P = 0.0004; Fig 1C). Differences observed between groups could not be attributed to other behavioral factors including differences in time of initial exploration of objects or motor behavior during testing (Fig 1D, S3 Fig). Immobility in the TST was increased in EMIP-exposed offspring compared with unexposed offspring (P = 0.004; Fig 1E). The behavioral deficits persisted to adulthood in EMIP-exposed offspring. Exposed mice tested at 20 weeks of age showed impaired performance in the NOR test (P = 0.001; Fig 1F) and increased immobility in the TST (P = 0.0002; Fig 1H).
We performed MRI to determine if the observed neurocognitive phenotype in EMIP-exposed mice was associated with changes in regional brain volumes. Prior to imaging, all mice were tested in the NOR test to confirm their behavioral phenotype. In utero exposed offspring showed impaired performance in the NOR test compared with unexposed offspring (P = 0.0009; S3 Fig). Volumetric analysis of brain volume in 63 distinct regions revealed no differences between EMIP-exposed and control mice (S2 Table).
A significant correlation across all WT mice (exposed and unexposed) was observed between total entorhinal cortical volume (volume of left and right cortices together) and performance in the NOR test (Spearman’s rho, 0.4912, P = 0.0044; Fig 1I). The mouse brain atlas [32] used to define the entorhinal cortex is depicted in Fig 1J. These data confirm a role for the entorhinal cortex in performance in the NOR test as suggested by previous reports [33,34]; however no difference was observed between malaria-exposed and unexposed animals.
Previous studies have shown that malaria in pregnancy is associated with altered placental vascular development [5]. We hypothesized that fetal cerebral vasculature may also be modified in malaria-exposed offspring, and that altered cerebrovascular development may contribute to the observed neurocognitive phenotype. Using a novel imaging approach in fetal mice, we performed micro-CT scans of fetal cerebral vasculature at G18. To our knowledge, this is the first time micro-CT has been used to visualize fetal cerebral vasculature. Using this technique, we identified all major cerebral vessels in fetuses and determined that there were no qualitative differences in major vessel architecture (Fig 2A–2D, S3 Table). In order to assess the impact of malaria–exposure on small vessel development, we further examined fetal cerebral vasculature with automated vessel tracking of the 3D images [35]. Vessel tracking analysis revealed a significant increase in the total number of vessel segments associated with in utero malaria-exposure (Fig 3A, P < 0.05). Malaria-exposure did not result in significant changes to total vessel length (Fig 3B).
Examination by MRI or micro-CT may not be sufficiently sensitive to detect subtle neurological features, such as changes in neuronal connectivity, capable of altering neurocognitive outcomes. Therefore, we next investigated levels of biogenic amine transmitters (dopamine, norepinephrine and serotonin) in four regions of interest (frontal cortex, temporoparietal cortex, striatum and hippocampus) based on their previously established involvement in the behavioral phenotypes we observed [36–38]. All tissue was harvested from animals that had been tested behaviorally to confirm their phenotype (Fig 1C–1E). Wild-type malaria-exposed offspring showed decreased tissue levels of dopamine (P < 0.01; Fig 4A) and serotonin (P < 0.005; Fig 4B) in the frontal cortex, norephinephrine in the temporoparietal cortex (P < 0.05; Fig 4C) and serotonin in the striatum (P < 0.05; Fig 4D) compared with wild-type unexposed offspring. Tissue levels of the catecholamine metabolite homovanillic acid were reduced in the frontal cortex and hippocampus of wild-type exposed mice (P < 0.05; S4 Table). Tissue levels of these analytes in each of the regions tested are reported in S4 Table. Maternal peripheral parasitemia (ranging from 14–31% on the day of delivery; Fig 1A) was not associated with differences in the observed levels of major biogenic amines, MRI or micro-CT imaging or neurocognitive outcomes.
Based on evidence linking C5a to both neuropathology and the pathophysiology of malaria [5,11,13,14], we examined the impact of genetic disruption of the C5a-C5aR signaling on neurocognitive outcomes in EMIP-exposed offspring. The deficits in NOR performance observed in WT EMIP-exposed offspring were completely rescued in C5aR deficient (C5ar-/-) EMIP-exposed offspring (P < 0.001) (one-way ANOVA and post-test, P < 0.004; Fig 5C). Again, no differences in time of initial exploration or motor behaviour were observed (Fig 5D, S4 Fig). Similarly, although immobility was increased in EMIP-exposed WT offspring in the TST, these features of affective-like behaviour were absent in EMIP-exposed offspring where C5aR signaling was disrupted (one-way ANOVA and post-test, P < 0.005; Fig 5E). Rescue of the neurocognitive deficits observed in EMIP-exposed C5ar-/- offspring persisted to adulthood (Fig 5F–5H). When tested at 20 weeks of age, EMIP-exposed WT mice showed impaired performance in the NOR test compared to EMIP-exposed C5ar-/- offspring (P < 0.001) and unexposed WT and C5ar-/- controls (one-way ANOVA and post-test, P < 0.002; Fig 5F). Performance of exposed C5ar-/- offspring was similar to unexposed controls (Fig 5F). Adult EMIP-exposed WT offspring, similar to malaria-exposed young mice, showed increased immobility in the tail suspension test compared with unexposed WT offspring (P < 0.001), and this effect was rescued in C5ar-/- offspring (one-way ANOVA and post-test, P < 0.0001; Fig 5H).
To provide a separate line of evidence that disruption of C5aR signaling rescues neurocognitive deficits in exposed offspring, we examined the impact of functional blockade of C5a in malaria-infected wild-type dams using C5a antisera [39]. Treatment of dams with anti-C5a antibody rescued the performance of EMIP-exposed offspring in the NOR test and TST. Offspring of dams treated with C5a antisera showed no significant difference in performance compared with unexposed offspring (P > 0.05; Fig 5I). However, performance in the NOR test was impaired in EMIP-exposed offspring and exposed offspring of dams treated with control sera (one-way ANOVA, P = 0.012; Fig 5I). EMIP-exposed offspring and exposed offspring of dams treated with control sera showed increased immobility in the TST compared with unexposed offspring (P < 0.01) and EMIP-exposed offspring of dams treated with C5a antisera (P < 0.001) (one-way ANOVA and post-test, P < 0.0001; Fig 5K). We performed additional testing on this cohort of animals to examine the impact of the saliency of the stimuli on cognitive performance. No significant difference in freezing behavior (a read out of contextual fear conditioning-based learning) was observed between groups on Day 2 or Day 3 of contextual fear-conditioning (one-way ANOVA and post-test, P > 0.05; S5 Fig).
C5a has been shown to be directly neurotoxic in vitro [22] and blockade of C5aR signaling in experimental models of MIP is associated with increased placental vascular development [5]. Therefore, to begin to examine putative mechanisms by which disruption of C5aR signaling may prevent neurocognitive injury, we performed MRI and micro-CT imaging of fetal cerebral vasculature in unexposed and malaria-exposed C5ar-/- offspring. We observed no volumetric changes as determined by MRI (S2 Table) as a result of EMIP-exposure in C5ar-/- offspring. Although micro-CT scans of fetal cerebral vasculature (Fig 2E–2H) at G18 revealed a significant increase in total vessel segments in malaria-exposed wild-type offspring (Fig 3A), disruption of C5a-C5aR signaling did not significantly reverse these changes. Therefore, neither changes in brain volumes as determined by MRI, nor microvascular development as assessed by micro-CT, provided an explanation for the cognitive impairments observed and their rescue by C5a-C5aR blockade.
We next extended our analysis to examine the impact of EMIP-exposure on monoamine transmitter levels in adult WT and C5ar-/- mice. In contrast to EMIP-exposed WT mice, regional brain levels of biogenic amines were not significantly decreased in EMIP-exposed C5ar-/- offspring (P>0.05, Students t-test, S5 Table). We normalized the levels of transmitters of EMIP-exposed WT and C5ar-/- offspring to the mean of their respective unexposed controls (Fig 6A–6F). Exposed C5ar-/- offspring displayed significantly higher levels of serotonin in the frontal cortex (P = 0.0028; Fig 6B), norepinephrine in the temporoparietal cortex (P = 0.012; Fig 6C) and serotonin in the striatum (P = 0.009; Fig 6D), compared to EMIP-exposed WT mice. Given the established role of BDNF in regulating brain monoamine levels [40,41], we determined whether decreased fetal brain BNDF levels were associated with the observed decrease in biogenic amines and whether disruption of C5aR signaling would rescue these levels. We observed decreased BDNF transcript levels in EMIP-exposed WT offspring (Fig 6G); whereas BDNF levels were restored in EMIP-exposed C5ar-/- offspring (one-way ANOVA and post-test, P <0.001, Fig 6G).
This study provides the first evidence implicating a causal link between pre-natal exposure to malaria, C5a-C5aR signaling and subsequent neurocognitive impairment in offspring. Our findings indicate that in utero exposure to maternal malaria infection can alter the cognitive and neurological development of offspring. We observed impaired learning and memory and depressive-like behavior that persisted to adulthood in EMIP-exposed offspring that were neither congenitally infected nor LBW. These neurocognitive impairments were associated with decreased tissue levels of major biogenic amines in cortical and subcortical regions of the brain. Genetic or functional disruption of maternal C5aR signaling restored the levels of BDNF and cerebral biogenic amines and rescued the associated cognitive phenotype observed in EMIP-exposed offspring.
Immunological stress at the maternal-fetal interface is associated with an increased risk of neurodevelopmental disorders in offspring [25,42–44]. MIP is characterized by the accumulation of parasitized erythrocytes and monocytes/macrophages in the intervillous space, creating a localized immune response in the placenta [6]. It is well established that components of the innate immune system, including complement factors, play diverse roles in angiogenesis, inflammation, neurogenesis and neurodevelopment [45,46]. Increased peripheral and placental levels of C5a are observed in women with MIP and are associated with adverse pregnancy outcomes [5,11]. C5a is a potent initiator of pro-inflammatory as well as anti-angiogenic pathways [9,10]. Together these observations are consistent with several potential mechanisms of impaired neurodevelopment including enhanced neuro-inflammation, altered neurovascular development, or dysregulation of complement-mediated neurodevelopmental processes [47].
Complement components are synthesized in the CNS by microglia, astrocytes and neurons and may be overexpressed in response to injury or inflammation [46,48]. Neurons, unlike peripheral cell types, do not express high levels of complement regulatory proteins, such as CD59, CD46, CD55 and CD35, suggesting that they may be particularly susceptible to complement-mediated injury [49]. A growing body of evidence supports an important role for the complement system in normal neurodevelopment, synapse formation and synaptic pruning [20,21,23]. Complement components tag excess synapses for elimination during pruning, facilitating the formation of mature patterns of neuronal connectivity [20]. Reduced levels of complement have been associated with decreased levels of synaptic pruning in the hippocampus and neocortex, a process critical for synaptic refinement during development [20,21,50]. These findings suggest that increased complement activation during development, as occurs in MIP, could lead to excessive synapse elimination and altered neuronal connectivity [23]. Moreover in models of lipopolysaccharide (LPS)-induced preterm birth C5a is reported to have direct neurotoxic effects on fetal cortical neurons in vivo and in vitro [22]. These neurotoxic effects were associated with C5a-induced glutamatergic excitotoxicity [22]. Collectively these data support the hypothesis that MIP-induced complement activation at the maternal-fetal interface may alter fetal neural networks and disrupt normal brain developmental processes. While increased peripheral and placental levels of C5a are observed in women with MIP [5,11], it is unclear whether fetal complement activation also occurs and whether it contributes to altered neurodevelopment. Future studies using the EMIP model could examine this question by mating heterozygote parents to generate WT, heterozygote and C5ar-/- offspring and determining the relative contribution of maternal versus fetal complement activation to neurocognitive outcome.
We propose that the cognitive deficits we observed in EMIP-exposed offspring are mediated, at least in part, by a reduction in regional brain levels of biogenic amines. Biogenic amines are reported to be central to learning and memory in the NOR test and depressive-like behavior in the TST [30,37,51]. We observed reduced serotonin and dopamine in the frontal cortex, reduced norepinephrine in the temporoparietal cortex and reduced serotonin in the striatum of EMIP-exposed WT offspring. EMIP-exposed offspring do not develop reductions of biogenic amines, in these regions, when C5aR signaling is disrupted. Norepinephrine has been associated with arousal and attention; responses to novelty that facilitate object recognition [33,51]. Cortical monoamine function, including dopamine and serotonin, has also been linked to performance in the NOR test [36,38,52,53]. Previous studies have implicated altered basal ganglia and cortical monoamine levels in TST behavior [37,54]. Specifically, pharmacological treatment with monoamine reuptake inhibitors that increase monoamine availability induce increased mobility in the TST [30,54,55]. Based on our behavioral and HPLC data we postulate that in utero exposure to malaria induces localized and subtle changes in neuronal development. We did not observe any behavioral deficits in the CFC test, a test of learning and memory using a highly salient and aversive foot-shock stimulus [56]. This suggests that prenatal exposure to EMIP does not induce a global impairment in learning and memory but alters behavioral performance in a task-specific manner.
Tight regulation of neurotrophic factors, in particular BDNF, is critical for normal neurodevelopment [40,57]. During embryogenesis BDNF regulates axonal and dendritic differentiation [40]. Immune responses to infections may alter BDNF levels, and disruption in BDNF-regulated processes can lead to alterations in brain monoamine levels and behavioral phenotypes in adulthood [41,58]. Our data, together with the above observations, support a model of pathogenesis whereby MIP-induced C5 activation impairs in utero neurodevelopment via effects on inflammation, synaptic pruning, neural network formation and regulation of BDNF, leading to reduced regional levels of monoamines and impaired cognitive performance in malaria-exposed offspring. When C5a-C5aR signaling is disrupted by genetic or functional approaches, there is reduced neurotoxicity, preserved regulation of BDNF and brain monoamine levels, and improved neurocognitive outcomes.
In addition to a role in neurodevelopment and neurodegenerative disorders, C5a is a potent initiator and amplifier of anti-angiogenic pathways and could theoretically alter neurodevelopment through angiogenic pathways as has been proposed for placental vascular development and remodeling during MIP [5,10,11]. Since C5a-C5aR blockade has been show to improve placental vascular development, it is also possible that rescue of the cognitive phenotype we observed in malaria-exposed C5ar-/- offspring is the result, at least partly, of changes in placental function.
Developmentally, neurogenesis and angiogenesis are tightly linked [59]. They utilize the same genetic and regulatory pathways and dysregulation in one system may alter developmental processes in the other. Therefore, we used a novel imaging approach to investigate whether the neurocognitive deficits in exposed-offspring were linked to altered neurovascular development as proposed [47]. Using micro-CT imaging, we observed an increased number of vessel segments, indicative of more vessel branching in malaria-exposed wild-type offspring. Whether this increase in cerebral vascular development represents a compensatory response to malaria-associated neurotoxicity will require further study. Overall this finding is consistent with previous observations that malaria also alters placental vascular development [5,10,60,61]. However, in the current study, disruption of C5a-C5aR signaling did not significantly reverse the vascular changes and did not provide a clear explanation for the cognitive phenotype observed and its rescue with C5aR blockade
Inflammatory conditions during pregnancy are associated with poor neurodevelopmental outcomes [25,44,62,63]. For example, maternal IL-6 cytokine surges have been reported to induce an increase in the forebrain neural precursor pool via activation of the embryonic neural stem cell self-renewal pathway [64]. Such inflammation-induced changes in early neurogenesis could have a significant impact on cognitive development. We have previously shown that C5a can enhance pro-inflammatory cytokine responses, including IL-6, to malaria-infected erythrocytes [11]. Our data do not exclude a role for neuro-inflammation in EMIP-associated adverse neurocognitive outcomes but rather suggest that both enhanced inflammation and altered neurodevelopment may be mediated through a shared pathway, C5a-C5aR signaling.
In summary, we show that in utero exposure to malaria infection disrupts normal cognitive and neurological development of offspring in a model of MIP and implicate activation of C5 in the pathobiology of this phenotype. In the clinical setting, MIP is commonly associated with LBW and there is a well-established link between LBW and increased risk of developmental delay [26,27]. Therefore, the cognitive deficits we observed would be expected to be incrementally increased by other MIP-associated birth complications including LBW caused by fetal growth restriction and preterm birth. Collectively our observations suggest a broader potential impact of malaria exposure in utero on neurocognitive outcomes since many malaria infections in pregnancy do not result in an obvious birth phenotype.
Factors that prevent normal neurological development of successive generations of children place enormous financial and social burdens on low resource countries. Persistent neurocognitive impairments as a result of MIP could have broad implications as pregnancies that occur in malaria endemic regions are at risk of MIP [65]. It is essential to identify preventable risk factors that contribute to developmental delay in children. Our data suggest that MIP is one such factor that can be targeted in order to improve cognitive development and school performance in malaria-endemic regions. A prospective study is underway to confirm these findings in children exposed to malaria in utero in sub-Saharan Africa (NCT01669941).
The EMIP model used in this study was based on a previously validated murine model of MIP, which replicates key pathogenic factors of MIP [28]. Female BALB/c mice (wild-type or C5ar-/-) between 6–8 weeks of age were mated with male BALB/c (wild-type or C5ar-/-) mice (8–9 weeks) were obtained from Jackson Laboratories (Bar Harbor, ME). C5ar-/- females were mated with C5ar-/- males, therefore all offspring were also C5ar-/-. Naturally mated pregnant mice were infected on G13 with 105 P. berghei ANKA-infected erythrocytes in RPMI media via injection into the lateral tail vein. A lower dose of innoculum (105 P. berghei ANKA-infected erythrocytes compared with 106) was used in this study to eliminate a low birth weight phenotype and increase the number of live births. Control pregnant females were injected on G13 with RPMI media alone. Thin blood smears were taken daily and stained with Giemsa stain (Protocol Hema3 Stain Set, Sigma, Oakville, ON) to monitor parasitemia. For pharmacological blockade experiments polyclonal rabbit antiserum raised against rat C5a or pre-immune control rabbit antiserum (Sigma G9023) was administered via tail vein injection 2 hours prior to malaria infection (0.25mL) and 72 hours post infection (G16) (0.25mL). Immediately following delivery all pups were given to surrogate (BALB/c wild-type) dams. All mice were weighed weekly beginning at one week of age. All litters were weaned at 3 weeks of age.
All experimental protocols were approved by the University Health Network Animal Care Committee (Animal Use Protocol number 1615 5/01/2014) and performed in accordance with the Canadian Council of Animal Care guidelines and current University Health Network regulations.
Placental tissue was collected from uninfected and malaria-infected females at gestational day 19 and whole placentas were immediately fixed in 20x volume of 10% formalin for 48 hours then transferred to 70% alcohol. Paraffin-embedded non-consecutive sections were stained for hemotoxylin-eosin (H&E) and examined under a light microscope (Olympus, BX41, Olympus Corporation).
Behavioral testing was conducted with male offspring beginning at 4 weeks of age and terminating at 7 weeks of age in the order the tests are presented below. In some experiments, testing was performed at 20 weeks. During testing, the experimenter alternated between testing mice from each experimental group. Offspring from a minimum of 4 different litters were used in each testing cohort. All testing was done with the experimenter blinded to the testing group.
A separate group of male offspring were behaviourally tested in the NOR test and TST at 5–6 weeks (Fig 1) and were euthanized at 8 weeks of age for MRI. All animals were weighed prior to behavioral testing and prior to perfusion. To minimize the likelihood of neurological changes resulting from behavioral tests using aversive stimuli, as in the CFC test, offspring to be perfused for MRI were only tested in the NOR and TST.
Microwave fixation of tissue was used to examine biochemical changes in biogenic amines as they relate to treatment group and performance in behavioral paradigms. All mice were tested prior to microwave fixation in the NOR test and TST. The microwave fixation procedures used here have been previously described in detail [30]. Briefly, all mice were euthanized at 8 weeks of age with a brief pulse (~0.9s) of high intensity microwave radiation (8 kW, 60Hz, 56 Amp) focused to the head and administered by a 10 KW magnetron (model TMW-4012C, Muromachi Kikai, Tokyo, Japan). Microwave fixation allows for rapid heat-inactivation of enzymes in situ and avoids confounding results due to post-mortem changes. Immediately following heat-inactivation, the heat-inactivated brains were dissected regionally on ice and stored at -80°C prior to analysis. Levels of dopamine (DA), norepinephrine (NE), serotonin (5-HT) and metabolites were assayed in perchloric acid tissue extracts with a Dionex HPLC system and electrochemical detector (DIONEX, Sunnyvale, CA, USA). Biogenic amines were selected based on extensive evidence linking these neurotransmitters with learning, memory and behavioural performance. HPLC was performed only on tissue from the temporoparietal cortex, frontal cortex, striatum, hippocampus and cerebellum based on the well-established role of these specific regions in learning, memory and motor behavior which impact performance in the NOR and TST [30,36,37,51,54]. As described previously, the chromatographic conditions included a C18 reverse-phase column (Acclaim 120, 150 x 4.0 mm2 cartridge, 5 μm particle size) at 30°C. The mobile phase consisted of sodium acetate (100 mM) tetrasodium EDTA (0.125 mM), 1-octane sufonic acid (432 mg/l) and 5.0% methanol (final pH = 3.6), delivered at a flow rate of 0.75 mL/min with a UltiMate 3000 pump. Samples (25 μl) were injected automatically with a refrigerated autosampler (UltiMate 3000 autosampler). The electrochemical detection (ESA Coulochem III 5011A analytical cell with a 5020 guard cell) was conducted at a working electrode potential of -400 mV.
Uteri were extracted from dams at gestational day 18 and anesthetized via hypothermia (immersion in ice-cold PBS). Each individual fetus is then extracted from the uterus while maintaining the vascular connection to the placenta. The embryo is briefly resuscitated via immersion in warm PBS to resume blood circulation. Embryo’s that could not be resuscitated are not perfused and were removed from the study. A catheter is then inserted into the umbilical artery and the fetus is perfused with saline (with heparin, 100units/mL) followed by radio-opaque silicone rubber contrast agent (Microfil; Flow Technology, Carver, MA). The perfusions were performed using two different lots of Microfil. Following perfusion specimens are post-fixed with 10% Formalin and imaged using micro-computed tomography (micro-CT). Specimens were scanned at 7.6 um resolution for 1 hour using a Bruker SkyScan1172 high resolution Micro-CT scanner. 996 views were acquitted via 180-degree rotation with an X-ray source at 54 kVp and 185 uA. Three-dimensional micro-CT data were reconstructed using SkyScan NRecon software. Each micro-CT image was manually masked to exclude extracerebral vessels using a cerebral vascular atlas as a guide [69]. The structure of the vasculature was identified automatically using a segmentation algorithm as described in detail previously [35]. Images that showed evidence of rupture of a major vessel or incomplete perfusion of Microfil were excluded from the analysis (9 of 38 specimens, S3 Table). Univariate ANCOVAs were conducted to compare the number of vessel segments and the total length of all vessel segments as a function of group, with dataset as a covariate (to control for the variance from using different lots of Microfil). A linear model was used to estimate the effect of dataset and the total segments and length were adjusted accordingly. Tukey contrasts were used to test differences between the adjusted means. Analysis was performed on wild-type (unexposed (n = 8) and malaria exposed (n = 7)) and C5aR knock out mice (unexposed (n = 7) and malaria exposed (n = 7)).
RNA extraction was performed on snap-frozen fetal brain tissue and placental tissue collected at G19. The EMIP model followed the same protocol outlined above. Dams were sacrificed at G19, yolk sacs were dissected from uteri, fetuses were removed and weighed, and fetal brain tissue and placentas were snap frozen and stored at -80°C until analyzed. Fetal viability was determined by assessing pedal withdrawal reflex. Non-viable fetuses (i.e., lacking the pedal withdrawal reflex) were considered aborted. Only viable fetuses and placentas from viable fetuses were used in the analysis. Tissue was homogenized in TRIzol (0.5mL/100mg tissue; Invitrogen, Burlington, ON) according to the manufacturer’s protocol and RNA was extracted. Extracted RNA (2 μg per sample) was then treated with DNase I (Ambion, Streetsville, ON) and reverse transcribed to cDNA with SuperScript III (Invitrogen, Burlington, ON) in the presence of oligo (dT) primers (Fermentas, Burlington, ON) with sequences listed below. Residual RNA was degraded with RNase H (Invitrogen, Burlington, ON). Sample cDNA was amplified in triplicate with SYBR Green master mix (Roche, Laval, QC) in the presence of 1 μM both forward and reverse primers in a Light Cycler 480 (Roche, Laval, QC). Transcript number was calculated based on Ct compared to the standard curve of mouse genomic DNA included on each plate by Light Cycler 480 software (Roche, Laval, QC), and expression in fetal brain was normalized to geometric average of the housekeeping genes GAPDH and β-actin expression levels. Expression in placental tissue was normalized to the housekeeping genes GAPDH and HPRT. A normalization factor was generated for each sample by dividing the mean sample expression by the mean expression of the housekeeping genes. The expression of each target gene was then divided by the normalization factor for that sample to adjust for experimental variation in gene expression [70]. RPTCR Primer Sequences: (5’–3’): GAPDH: TCAACAGCAACTCCCACTCTTCCA–TTGTCATTGAGAGCAATGCCAGCC, β-actin: GCGCCCATGAAAGAAGTAAAA–TTCGATGACGTGCTCAAAAG, HPRT: GGAGTCTGTTGATGTTGCCAGTA–GGGACGCAGCAACTGACATTTCTA, BDNF: GCGCCCATGAAAGAAGTAAA–TTCGATGACGTGCTCAAAAG. ICAM-1: CGGAAGGGAGCCAAGTAACTG–CGACGCCGCTCAGAAGAA, TNF: GACAGACATGTTTTCTGTCAAACG–AAAAGAGGAGGCAACAAGGTAGAG, IFNγ: TTCTGTCTCCTCAACTATTTCTCTTTG—CCCCACCCCCAGATACAAC, MCP: ACCACAGTCCATGCCATCAC—TTGAGGTGGTTGTGAAAAG
Student’s t-test, one-way ANOVA or ANCOVA (non-parametric Kruskal-Wallis, P < 0.05) was used to examine statistical significance between experimental groups where indicated. Post-tests on all groups were conducted using Dunn’s multiple comparison test or Tukey contrasts where indicated (P < 0.05).
|
10.1371/journal.pcbi.1002556 | The Impact of Phenotypic Switching on Glioblastoma Growth and Invasion | The brain tumour glioblastoma is characterised by diffuse and infiltrative growth into surrounding brain tissue. At the macroscopic level, the progression speed of a glioblastoma tumour is determined by two key factors: the cell proliferation rate and the cell migration speed. At the microscopic level, however, proliferation and migration appear to be mutually exclusive phenotypes, as indicated by recent in vivo imaging data. Here, we develop a mathematical model to analyse how the phenotypic switching between proliferative and migratory states of individual cells affects the macroscopic growth of the tumour. For this, we propose an individual-based stochastic model in which glioblastoma cells are either in a proliferative state, where they are stationary and divide, or in motile state in which they are subject to random motion. From the model we derive a continuum approximation in the form of two coupled reaction-diffusion equations, which exhibit travelling wave solutions whose speed of invasion depends on the model parameters. We propose a simple analytical method to predict progression rate from the cell-specific parameters and demonstrate that optimal glioblastoma growth depends on a non-trivial trade-off between the phenotypic switching rates. By linking cellular properties to an in vivo outcome, the model should be applicable to designing relevant cell screens for glioblastoma and cytometry-based patient prognostics.
| In this work, we develop a spatial mathematical model in order to analyse the growth behavior of the brain tumour glioblastoma. Tumours of this type have a diffuse boundary, with considerable local invasion of surrounding brain tissue, making surgery difficult. At the cellular level, the progression of a glioblastoma is known to depend on the balance between cell division (proliferation) and cell movement (migration). Based on recent evidence, our model assumes that each cell in a glioblastoma tumour resides in either of two mutually exclusive states: proliferating or migrating. From a probabilistic model of switching between these two phenotypes, we go on to derive equations that link cellular phenotypes to disease progression. The model has several possible applications. For instance, it could be used to predict the rate of disease progression in an individual patient, and to improve screening methods.
| Cancer progression is the macroscopic outcome of numerous cellular processes, such as elevated proliferation rates, defects in apoptosis regulation and abnormal angiogenesis [1]. In the development of targeted anticancer therapies, the proliferation, survival and angiogenesis phenotypes are often singled out as the most important. Recently, however, much attention has been given to cancer cell migration as a possible therapeutic target, since it underlies both the local invasive process whereby cancer cells degrade and move through the adjacent tissue, and the formation of distant metastases.
The importance of cancer cell migration is perhaps most evident in the common brain tumour glioblastoma, which is characterised by rapid and infiltrative growth into the surrounding brain tissue. In glioblastomas, neoplastic cells are often found at a long distance (several centimeters) from the main tumour mass. This diffuse growth pattern presents a difficult clinical problem, since residual ‘satellite cells’ can mediate rapid recurrence of the disease after surgery [2]. Key factors that underlie glioblastoma cell invasiveness include high migration speeds in comparison to other types of cancer (up to 100 m/h) and the fact that brain parenchyma provides a penetrable substrate for invasion [3]. Thus, inhibition of migration pathways might constitute an interesting complement to standard glioblastoma therapies that seek to inhibit cell proliferation rate or angiogenesis. Several pathways have been suggested to mediate the highly migratory phenotype of glioblastoma cells, including signaling via Focal adhesion kinase (FAK) [4], Phosphoinositide 3-kinase PI3K [5] and Signal transducer and activator of transcription 3 (STAT3) [6]. Other concepts for targeting of migration have also been proposed, including inhibition of integrins [7], perturbing the interactions between ECM components [8], and administering lithium chloride [9]. Further, potential gene targets have also been revealed using molecular profiling efforts [10].
However, the potential of migration as a therapeutic target is complicated by the strong dependency between migration and proliferation phenotypes. Early in vitro experiments by Giese et al. [11] showed that when plated on a substrate, that supports migration, the proliferation rate of glioblastoma cells is markedly reduced. Later, it was shown that cells at the tumour's invasive rim proliferate more slowly than cells in the central parts of the tumour, again suggesting that migration has a ‘cost’ in terms of reduced proliferation [2]. These and several additional observations led to the so called ‘go or grow’ hypothesis, stating that migration and proliferation are mutually exclusive phenotypes. Although a molecular explanation for this dichotomy still is missing, it has been suggested that cytoskeleton dynamics could be limiting, as it is involved in both cell division and force generation in migration [12].
In recent experiments, in vivo imaging of fluorescent glioblastoma cells enabled direct observation of phenotypic switching between the ‘go’ state (migration) and the ‘grow’ state (proliferation). More precisely it was observed that glioma cells move in a saltatory fashion, where bursts of movement are interspersed by periods of immobility, and it is during these stationary periods that the cells divide [13], [14]. Taking these observations into account allows for a more comprehensive understanding of glioblastoma progression, where tissue-level traits, such as progression rate, emerge from cell-level behaviour. Mathematical models at the resolution of individual cells enable a quantitative connection between these scales, and can hence be of great assistance.
Here, we focus on the relationship between cell-level phenotypic switching in glioblastoma, and the properties of the tumour as a whole. In particular we elucidate how the growth rate of the tumour and speed of invasion depends on the specific underlying microscopic parameters, such as phenotypic switching rates, rate of apoptosis et cetera. Please note that the we use the word ‘invasion’ to denote the process by which glioma cells spread into and displace the surrounding brain tissue, and do not refer to branched finger-like growth patterns. Although several models of glioma growth have previously been proposed (see next section), this model is the first to connect experimentally measurable cell-level traits with gross tumour volume in an analytical way. This yields hope for the future understanding of glioma biology and therapy, since it is the understanding of how drug induced changes on the cell-level scale propagate to the organ scale, that are required in order to accurately predict therapeutic outcome.
In the following we first review previous work in the field of glioblastoma modelling and then proceed by introducing our individual-based (IB) stochastic model of glioma growth. From this model we derive an approximate continuum description of the system, whose properties are compared to the IB-model. We proceed to analyse the continuum model to reveal the influence of the model parameters on the rate of spread of the tumour, and finally discuss our results in the context of other models and experimental results.
The growth of glioblastomas was first modelled by means of a continuum approach, which captures the two main features of glioma cells: proliferation and migration ([15], [16], and [17], chapter 11). In that model, the partial differential equation (PDE) that describes the time evolution of the concentration of glioma cells in space and time has the form:(1)where the migration is captured by a diffusion term with diffusion coefficient (first term) and proliferation of the glioma cells is described by a locally logistic growth function with growth rate (second term). This equation is known as the Fisher (or Kolomogorov) equation, and was first derived in order to describe the spread of an advantageous gene in a spatially extended population [18], [19]. The derivation, originally by Fisher and later refined by Kolmogorov, starts by assuming a contact distribution that describes the probability of migration between two spatial locations, and by then assuming that all moments of that distribution higher than two are negligible (known as the diffusion approximation, see for example [20]) one arrives at the above equation.
The Fisher equation has been of particular interest since it gives rise to traveling wave solution , whose shape is preserved and position in space is shifted at a speed as time progresses. Significant interest has been devoted to determining the wave speed , and it has been shown that for reasonable initial conditions (exponentially decaying [20] or of compact support [19]) the wave speed is given by [21]. The speed of propagation thus depends on both the motility, captured in the diffusion constant , and the rate of proliferation . Both and can be determined from time-course Magnetic Resonance Imaging (MRI) data from actual patients, and it has been shown that their values are of prognostic power [22].
The above modeling approach rests on the assumption that glioblastoma cells follow a random walk (which at the macroscopic scale corresponds to the diffusion of cells). Recently this assumption has been under scrutiny, and this has led to a number of explorations of non-random migration, i.e. where migration is influenced by biological processes such as cell-cell signaling, oxygen pressure, nutrient availability and phenotype switching. In one line of work, Aubert et al. [23] used an individual based (IB) model to show that attraction between glioblastoma cells is likely to influence the dynamics of tumour invasion. Deroulers et al. [24], derived the macroscopic PDE for this case, obtaining a density dependent diffusion equation ( in terms of eq. (1)), whose solution deviates significantly from the Fisher-Kolmogorov PDE (see also [25], [26]). Khain et al. [27], used IB models to characterise the role of hypoxia in glioblastoma, showing that reduced oxygen levels may down-regulate cell-cell adhesion, leading to increased motility.
The cellular behaviour implied in the ‘go-or-grow’ hypothesis (see Introduction) is also thought to affect migration and growth dynamics of glioblastomas, in a manner that is not captured by the Fisher-Kolmogorov equation. Hatzikirou et al. [28] proposed a lattice-gas cellular automaton model in which the switching between the proliferative (P) state and migratory (M) state is driven by lack of oxygen, and went on to show that in the corresponding macroscopic (Fisher) equation, there is a tradeoff between diffusion and proliferation reflecting the inability of cells to migrate and proliferate simultaneously. Similar results where obtained by Fedotov and Iomin [29] but with a different type of model known as continuous time random walk model. That model contains two distinct subpopulations (P-cells which are stationary and divide and M-cells that perform random walks), and a cell switches from one compartment to the other after a time (and respectively) which is exponentially distributed. They analytically show that the spreading rate is smaller than one would expect from the Fisher equation (1). Finally, Lewis and Schmitz have studied the general relation between organism migration and proliferation and its impact on population spread using reaction-diffusion equations [30]. They show that the system exhibits travelling wave solutions and that the wave speed depends on the rates of switching between the states.
The model that we propose draws from these previous models, but is different in some crucial ways. We consider two distinct subpopulations with a stochastic switching in between (as in Fedotov and Iomin, and Lewis and Schmitz), but instead of starting with a continuum description, we begin with an IB-model in which the cells occupy a lattice and obey size exclusion (as in Deroulers et al.), and from that derive a system of PDEs. This allows for an analytical treatment of the IB-model which establishes a connection between cell characteristics and the macroscopic behaviour of the system previously not demonstrated.
The cells are assumed to occupy a -dimensional lattice (we will consider ), containing lattice sites, where is the linear size of the lattice and each lattice site either is empty or holds a single glioma cell. This means that we disregard the effects of the surrounding brain tissue, such as the different properties of grey vs. white matter [31], and the presence of capillaries which might influence the behaviour of the cancer cells. But since the soft tissue in the brain presents little resistance to invading cancer cells and the precise nature of interaction with stromal cells is unclear, focusing on the dynamics of the glioma cells is a reasonable first approximation. Further, the process of angiogenesis, which has been modelled extensively [32], is ignored, and we hence assume the growing tumour to be well vascularised. For the sake of simplicity we do not consider any interactions between the cancer cells (adhesion or repulsion), although this could easily be included, and we also disregard other types of mechanical interactions, such as cell pushing (see Discussion).
The lack of knowledge of the intra-cellular dynamics and extra-cellular cues that lead to the phenotypic switching behaviour poses a problem, but we will circumvent it by, as a first approximation, considering the switching as a stochastic event. The behaviour of each cell is therefore modelled as a time continuous Markov process where each transition or action occurs with a certain rate, which only depends on the current and not previous states, known as the Markov property. The rates are interpreted in the standard way, i.e. if transition in a variable from state to occurs at rate then the probability of a transition occurring in the time interval is given by(2)where means that remaining terms are bounded from above by , and thus that in the limit of small the transition probability is proportional to .
Each cell is assumed to be in either of two states: proliferating or migrating, and switching between the states occurs at rates (into the P-state) and (into the M-state). A proliferating cell is stationary, passes through the cell cycle, and thus divides at a rate . The daughter cell is placed with uniform probability in one of the empty neighbouring lattice sites (using a von Neumann neighbourhood). If the cell has no empty neighbours cell division fails. A migrating cell performs a size exclusion random walk, where each jump occurs with rate . Size exclusion means that the cell can only move into lattice sites which were previously empty. If only a motile subpopulation is considered, size exclusion does not affect the macroscopic diffusive nature of a population of random walkers, but if two or more subpopulations are taken into account then, as we shall later see, diffusion becomes non-linear. Lastly, cells are assumed to die, through apoptosis, at a rate independent of the cell state. Since this type of cell death is associated with cell shrinking and rapid removal of the dead cell, a cell which goes through apoptosis is instantly removed from the lattice and leaves an empty lattice site behind.
The stochastic process is depicted schematically in figure 1. In fact the whole system comprises a continuous time Markov chain with a finite, but very large state space, containing different states, where is the linear size of the system and the 3 comes from the three distinct lattice states: empty, P-cell and M-cell.
We will consider a lattice of linear size with a spacing of , the typical size of a cancer cell. For the most part we will consider the system in dimensions, which means that we simulate a lattice, which corresponds to a 4 slice of tissue. This is of course considerably smaller than a clinically relevant glioma, but sufficient to capture the effects of the phenotypic switching on tumour growth rate. The time scale of the model is set to agree with that of the cell cycle (approximately 24 hrs [2]) which means that the proliferation rate , and that we scale all other parameters accordingly. We are mainly interested in the effect of the phenotypic switching rates and on the growth of the tumour and they will therefore be varied within a biologically reasonable range. It follows from equation (2) that the time spent in one phenotypic state is exponentially distributed with parameter and thus that the average time spent in each state is given by (cell cycles). It has been observed that the switching from a stationary to motile state (and back) does not occur faster than on the time scale of one hour [13]. This gives an upper limit on the transition rates, which since time in the model is measured in cell cycles, is given by .
The motility rate is set to . This means that a motile cell on average moves one lattice site, i.e. , in a time cell cycles, which gives a linear velocity of , that lies within experimentally determined values of [14] and [13]. The rate of apoptosis is set to the value , which is small compared to the other transition rates in the model.
Our concern is the influence of the microscopic cell-level parameters on the growth rate of the tumour as a whole, and we will therefore measure the size the tumour after a fixed time for a given set of initial conditions, as a function of the phenotypic switching rates. More specifically we will measure the tumour mass (the total number of cells), and also later, quantify the rate of spread by measuring the velocity of the tumour interface. The precise initial condition of the model has little impact on the long-term rate of spread (data not shown), but in line with the clonal origin of cancer we initialise the model with a single cell (in the P-state). All simulations of the IB-model are carried out using the commonly employed Gillespie algorithm [33].
Figure 2 illustrates the results of simulating the model in two dimensions for cell cycles when in three different ways. Panel (a) shows the result of a single simulation, where P-cells are coloured blue and M-cells are red, (b) shows the results of the model averaged across a large number of realisations and gives the occupancy probability of finding a cell at location on the lattice, and (c) shows a slice through this function . This figure gives us a general idea of the growth dynamics of the model. The tumour grows with a radial symmetry, and exhibits a solid core, while the tumour margin is diffuse and somewhat rugged. Please note that the time span considered in this simulation is smaller than the time scale of actual glioblastoma growth, which usually occurs on the time scale of months to years, but still sufficient to investigate the dynamics of the model.
In order to quantify the dependence on the phenotypic switching rates we measured the tumour mass at in the parameter range . The results are displayed in figure 3a and show a strong dependence on the two parameters. For all cells are in the proliferative state, and as expected the mass is independent of . The other extreme where gives rise to tumours with a zero mass, which occurs since the motile cells cannot multiply and eventually die off due to the small but non-zero apoptosis rate . These results are intuitive, but what is more interesting is that tumour cells with intermediate switching rates are the ones that give rise to the largest tumours. Although migratory behaviour does not directly contribute to an increase in the number of cancer cells it has the secondary effect of freeing up space which accelerates growth compared to the tumours dominated purely by proliferation (). The results suggest that for each there is a which gives a maximal tumour growth rate. These results also hold for the more biologically plausible 3-dimensional case (see figure 3b). Although the maximal tumour mass seems to occur for a smaller , and the region of parameter space giving rise to small tumours is considerably larger (upper left region), the qualitative behaviour is similar. The implications of the observation that influences tumour size in a non-monotone way will be discussed later, and we will now proceed to an analytical treatment of the problem.
In an effort to get a deeper understanding of the somewhat unintuitive relationship between tumour growth rate and phenotypic switching rates we will derive a set of two coupled PDEs which will serve as an approximate way of describing the time evolution of the occupancy probability (see figure 1b and c). For the sake of clarity we will however constrain the derivation to a one-dimensional system. In fact, radially symmetric travelling wave solutions with constant velocity do not exist for , but instead the velocity of the front depends on the local curvature. However, for large enough times the interface of the circular (spherical) solution has almost zero curvature and its dynamics is well approximated by the one dimensional solution. A further simplification in our derivation is that we assume the occupancy probabilities of neighbouring sites as independent, which in practice means that we for example allow ourselves to write: (site empty and site occupied) = (site empty)(site occupied).
The derivation is carried out in two steps: firstly, a set of coupled master equations, for the two sub-populations, are derived by considering the processes which alter the occupation probabilities at a given site, and secondly these master equations are approximated by a set of PDEs. In brief, the second step is achieved by identifying the on-lattice master equations with a set of coupled PDEs, which when discretised on the length scale of the lattice spacing, equal the master equations. The full derivation is given in Methods and results in the following system of coupled PDEs:(3)(4)Here denotes the density of proliferating cells, and that of the motile cells. In equation (3) we recognise the first term as a diffusion term, modulated by a density-dependent prefactor and the second term as a logistic growth term. The remaining terms are due to the switching between the subpopulations and to apoptosis. In the equation for the motile cells (4) there is also density-dependent diffusion, but of a different type. This is typical of a two species size exclusion process [34], and contains the second-derivative of both species. The values of the diffusion constants are and , and depend crucially on the choice of spatial scale, which for simplicity is chosen to be that of the cell size . If a coarser spatial scale is considered then the diffusions constants would have to be scaled accordingly (see Methods for details). We will now proceed to investigating the properties of this system of PDEs through both numerical solutions and analysis.
The first question one might ask about a system of equations that presumably describes tumour growth is if it exhibits tumour invasion and hence travelling wave solutions, and further how the model parameters influence the wave speed, i.e. the velocity of the invading tumour front. The results from the IB-model (figure 2 and 3) suggest that the switching rates strongly influence the tumour mass, and hence we expect them to also have an effect on the speed of invasion.
In order to investigate this, we first solved the continuum model (3)–(4) numerically (which actually corresponds to reverting to the master equations eq. (8)–(9)), for a range of parameter values, in the domain . The initial condition was set to , and , meant to represent a situation where a tumour is initiated by a small number of proliferating cells () and no migratory cells (). In fact the balance between and in the initial condition is largely irrelevant for the long-term dynamics of the model, the exceptions being the extremes and , when flow between the phenotypes is unidirectional or completely blocked. The boundary conditions of the domain were set to no-flux.
The results can be seen in figure 4 and shows the occupancy probabilities after and 50 cell cycles. From these results it is clear that the system exhibits an invading front of cancer cells, similar to what is observed for the Fisher equation. The leftmost panel (a) shows the dynamics of a tumour which only contains proliferating cells, while (b) and (c) exhibit a mix of P- and M-cells. The solutions remain stationary in a moving frame, suggesting that travelling wave solutions exist, with wave speeds 1.48, 1.88 and 1.63 respectively. These numerical results mirror what was seen in figure 2, where an intermediate gave rise to the largest tumours. Please note that the wave speed for the case is roughly what one would expect from a Fisher equation with and , since . However, this is not what occurs in the IB-model where the tumour interface moves at an average velocity of . The source of this discrepancy is the assumption of independence between sites, which applies the least in this particular case ( = 0), when there is no random motion within the cell population. Migration of the cells tends to break up the correlations that build up as the tumour is growing, and as we later shall see, the continuum approximation works better when the cells are more motile.
The observation that the numerical solutions are stationary in a moving frame suggests the existence of travelling wave solutions. In order to close in on these solutions, and get an estimate of their velocity, we will make use of the travelling wave ansatz: and with , where is the velocity of the interface. The problem of determining how depends on the model parameters is solved by applying phase-space analysis (see Methods), and boils down to a four-dimensional eigenvalue problem, namely to find the smallest such that the eigenvalues of the Jacobian all have imaginary part equal to zero. This problem is analytically intractable, but provides us with a numerically easy way of determining the velocity.
Although phase-space analysis does not yield an analytic closed-form expression for the wave speed , it still provides us with a computationally simple way of determining the velocity of the tumour margin in the model: for a given set of parameter values we start by setting and calculate the eigenvalues of the Jacobian (18) (or equivalently the roots of the corresponding characteristic polynomial ). If not all eigenvalues are real we increment slightly and reevaluate the eigenvalues. This procedure is terminated as soon as we find all eigenvalues real, and the value of for which this occurs corresponds to the wave speed for those parameter values.
In order to test the validity of the wave speed analysis we compared the wave speeds obtained in the continuum and IB models with those from the phase space analysis. For the continuum model an estimate of the wave speed was obtained by, from the initial condition (for proliferating cells), and (for migrating cells), integrating the equations (3)–(4) for 200 time steps (cell cycles). From these solutions we estimated the velocity of the front by measuring the position of a reference point , defined as the point where , as a function of time. The comparison between the speed of propagation in the numerical solution and the wave speed obtained from the phase space analysis is shown in figure 5. The agreement is fairly good and the discrepancies are probably due to error in integration and the deviation in the numerical solution from a perfect travelling wave, which from a given set of initial conditions, is only attained in the limit . However, since we are interested in biologically relevant scenarios the time frame considered is reasonable.
When it comes to the IB model, we have to take into account the stochastic nature of the model, and therefore need to estimate the average margin velocity from a large number of simulations (100 independent realisations). Each simulation was started with a single P-cell at the center of the lattice and the model was simulated for 100 time steps (cell cycles). In each time step the location of the cells was recorded and from this we calculated the occupation probability of finding any cell at location at time . The wave speed was then approximated by taking the average propagation speed of in the and direction (as in the PDE case). In comparing with the two-dimensional simulations we need to rescale the diffusion coefficient , since cell movement occurring tangential to the two-dimensional front does not contribute to its propagation. The result can be seen in figure 6, which shows that the analytical result is in good agreement with the discrete individual-based model. The disparity between the IB-model and the analytic answer is largest for small , when the dynamics are dominated by proliferation. This is to be expected since for larger the movement of the cells decorrelates the sites, and hence our assumption about site independence is closer to truth. The analytical results recapitulates the non-monotone dependence on and using this method we found that the largest tumours occur when , i.e. when the ratio between the switching rates is 1∶2.
Naturally the other model parameters also affect the rate of tumour invasion (see figure 7). Increasing the proliferation rate leads initially (for small ) to an increase in velocity according to , while for there is a cross-over to a linear dependence with , with . The motility rate also influences the wave speed in a non-linear way according to the relation , which holds for all . Finally, increasing the rate of apoptosis , as expected, decreases the wave speed, and does so in a non-linear way. Actually the dependency on looks very much like that of a second-order phase transition, where the derivative diverges at a critical point , and we have for that (see inset of figure 7c). We observed that the critical apoptosis rate , above which no travelling wave solutions exists and hence the tumour disappears, depends on the other parameters of the model, but that the critical exponent is independent of the other parameters.
Our model gives considerable insight into the dependency between five cell-level parameters (switching rates and , motility rate , proliferation rate and rate of apoptosis ) and the macroscopic dynamics of tumour growth and invasion. Focusing on the impact of the phenotypic switching rates we showed that tumour cells with a small and large (see fig. 3) give rise to small tumours (low ) while those characterised by a large and intermediate grow into large tumours (high ). To see why this is the case, consider a one-dimensional growth process in which the tumour expands in a narrow channel. If , then the tumour expands only through proliferation of the cells at the interface (since interior cells cannot divide), and the interface thus moves with velocity , equal to the proliferation rate. If then cells at the interface spend some time in the motile state, freeing up space and allowing previously blocked cells to proliferate. This process increases the interface velocity, but it is also clear that a large has a negative effect on tumour growth, since if fewer cells are in the proliferative state and can thus take advantage of the space created via cell migration. From this line of reasoning it is clear that the tumour interface velocity will depend on in a non-monotone way. Taken together, our analysis shows that and affect glioblastoma progression by altering the composition and structure of the tumour interface, and that for each the velocity attains a maximum, which occurs at .
The above reasoning, and our model, do however not take into account the effects of mechanical forces between the cells. In particular it is, in real tumours, possible for cells to push one another and hence to divide and move, although there is no free space. This process will most likely lessen the positive effect of cell migration on tumour growth, but since it has been experimentally established that few cell divisions occur in the core of the tumour due to pressure build-up and hypoxia, we believe that the conclusions of our model still hold to a large extent.
A similar trade-off between proliferation and migration has in fact been observed in the models of Hatzikirou et al. and Fedotov and Iomin [29]. Although a formal comparison with the former model is difficult, the macroscopic equations that Hatzikirou et al. derive show that the number of rest channels (comparable to the likelihood of a cell proliferating), increases the proliferation rate, but at the same time decreases the motility of the cells. In the work of Fedotov and Iomin [29] a similar trade-off is present. Using a continuous time random walk model they showed that if the waiting times in the P- and M-state are exponentially distributed (as in our model) then the margin velocity is non-monotone in the ratio and that the maximum velocity is achieved for . However it should be noted that their model does not take size exclusion into account, and hence yields an overestimate of the effects migration has on invasion.
A trade-off between proliferation and migration has also been investigated in relation to cancer stem cells and tumour progression by Enderling et al. [35]. They showed that cell migration can lead to the formation of secondary tumour loci, in a process termed self-metastasis, which might accelerate tumour growth, depending on the ratio of migratory and proliferative behaviour. In a related study it was shown that cancer stem cell migration might lead to branched tumour morphology and that it can increase the chance of tumour recurrence [36]. These modelling results together with those presented in this study highlight the importance of cell migration in tumour progression and motivate future experimental studies.
We have also demonstrated that the other parameters in the model affect the speed of invasion. Firstly, the impact of the motility rate and the proliferation rate imply that the wave speed dependence observed in the Fisher equation (1), , also holds in our system, when equating the diffusion constant with the motility rate and the proliferation rate with (at least for small and biologically realistic values of ). The Fisher equation has been shown to give an accurate macroscopic description of glioblastoma progression in vivo [22], which also lends support to our model. The correspondence to the Fisher equation is particularly interesting since it allows for a connection between cell level characteristics ( and ) and tissue-scale behaviour, and suggest a means of parametrising our model, not only using single-cell measurements, but also from tissue-level data, such as MRI-scans. From consecutive images the position and hence velocity of the invading tumour margin can be determined and compared with the results of the model.
Secondly, we observed a second-order phase transition in the velocity with respect to the rate of apoptosis . This means that there is a critical apoptosis rate above which no tumour can grow and that for we have , where is parameter-dependent, but seems to hold for all parameter values. The discontinuous behaviour of is interesting, not only from a theoretical perspective, but also because it implies that if a high enough rate of apoptosis is induced, it may not only retard tumour growth, but in fact lead to regression. However, these results should viewed with caution, since the model would need to be modified and extended in order to properly account for the dynamics of drug delivery and treatment (cf. [37]).
While data from Farin et al. [13] served as the impetus for our model, we note that a few additional experiments support our modeling assumptions. First, the general observation that glioma cells sampled from invasive fast-growing tumours are characterised by a blend of proliferative and migratory behaviour [2] supports our results, although only in a qualitative way. Second, a recent study on different glioma subclones obtained from the same patient identified a particular cell type as being particularly invasive. Subsequent analysis of proliferation of these clones (determined by Ki67-staining) showed that the most invasive subclone (giving rise to the largest tumours in vivo) had the lowest proliferation rate [38]. Although the subclones were not subject to a motility assay, these results still diminish the importance of cell proliferation in determining tumour growth rate, and future studies that measure both proliferation and migration could be even more useful in this respect.
In order to gain further experimental support for our model, we plan in future work, to measure the five cell specific parameters directly. Such measurements should be possible by applying live imaging microscopy techniques to primary glioblastoma-derived cell cultures. A first application of such measurements could be exploited to develop the model further, to predict progression for an individual patient based on cell-level phenotyping, and to develop chemical compound screens where the impact of a chemical on the model parameters are observed. This might in turn lead to a strategy to define in vivo-relevant compounds more likely to inhibit progression.
The current model is however far from these highly set goals, and there are a number of extensions that would make the model more realistic. In its current form the model does not account for cell-cell adhesion, which could be incorporated letting the motility rate be dependent on the neighbourhood of the cell [25], [26]. The preferential migration along capillaries and myelin tracts, and the tendency for glioma cells to divide at capillary branch points, is also something that could be included. A further complication is that cancer cells within a real glioblastoma are not identical with respect to their behaviour, but exhibit both genotypic and phenotypic heterogeneity, e.g. cells with a migratory phenotype tend to be located at the tumour boundary whereas dividing cells are commonly found in the main tumour mass, a fact which is not captured by the current model.
Despite this we would still expect the results of our model to hold at least with respect to the large-scale behaviour of the tumour. The real situation is also complicated by the fact that cancer cells are selected for based on their phenotype. One hypothesis which emerges from our model is that selection could drive the behaviour of the cells to the optimal balance between and , although this hypothesis would require a model that allows for population heterogeneity in order to be tested.
Adding these mechanisms would of course make the model less tractable from an analytical point of view, but this trade-off between simplicity and reality is something that all modellers must deal with.
Let us consider a one-dimensional lattice indexed by the integers. We let denote the probability of finding a P-cell at site at time , and equivalently let represent the occupation probability of M-cells. The general strategy is to formulate two coupled master equations for the occupation probabilities, which will then be approximated by a set of PDEs, amenable to a wave speed analysis that hopefully will reveal the influence of on tumour growth.
Let us first consider . Which are the processes that affect this quantity at a given site?
Summarising all these processes we can write:where the first term is a loss term due to apoptosis, while the second and third term are due to phenotypic switching. The final term is due to cell division from the neighbouring sites, and here we have made use of the independence assumption discussed above. After dividing both sides by and going to the limit we end up with the following expression:(5)
In order to simplify the expression and also draw parallels to continuum systems we define a discrete Laplace operator(6)where corresponds to the spacing of the lattice. Equation (5) can now be written as(7)
If we now turn to the motile cells, the following processes affect :
Taking all these processes into account we can write
The first three terms can be recognised as apoptosis and switching terms, while the fourth and fifth are due to movement out of and into the site. After a bit of algebra and making use of the discrete Laplacian defined in eq. (6) we get
In summary we have that the time evolution of the occupation probabilities are described by the following coupled equations:(8)(9)
Please note that despite the similarity to PDEs, that describe the changes of a quantity in continuous space and time, these equations are defined on the lattice and describe the probability of finding a cell of a specific type in a certain location. In many instances it is natural to proceed by taking the spatial continuum limit of the discrete master equation(s), but in this case, where we are considering expansion via both cell movement and pure cell division (the case ), things are a bit more delicate, because when the size of the cells tend to zero () so does the contribution of cell division to tumour expansion. In order to achieve a sensible continuum limit, a certain scaling in space and time is required, which implicitly assumes that cell motility occurs on a much faster time scale than cell division [39], something which is generally not the case in the case of glioma biology.
However in order to proceed with the analysis and make use of the toolbox of real analysis we will approximate the above equations with the following PDEs:(10)(11)
The motivation behind this choice is that the master equations (8) and (9) are the (space) discretised versions of (10) and (11). The diffusion constants are given by and , where is the spacing of the lattice, which we for simplicity measure in terms of cell size, and accordingly set . This means that we consider the dynamics on the length scale of a single cell. Please note that the unit of the diffusion constants and is , while the unit of the underlying proliferation and migration rates is . The correspondence between the master equations and PDEs is, however, not rigorous and implies that the analytic results obtained for the PDEs are not in general valid for the master equations, but, as we shall see, still reflect the behaviour of the IB-model to a large extent.
For the sake of clarity let us recapitulate the method applied to the Fisher equation (1) in order to calculate its speed of invasion. The travelling wave ansatz () turns the Fisher equation into second order ODE in the variable . By introducing the variable the ODE is turned into a two-dimensional autonomous system. The system has two fixed points and , and by calculating the eigenvalues of the Jacobian (a determinant of partial derivatives) at the two fixed points, one finds that the fixed point at (corresponding to the invaded state) is a saddle point (independent of ), while the characteristics of the one at the origin depend on . For the fixed point is a stable spiral, while for it is a stable node. The heteroclinic orbit connecting the two states goes from (1,0) through the third quadrant (, as in figure 4), and only if the origin is a stable node does it enter the fixed point without spiralling around and attaining negative values of the density of cancer cells. Negativity would be inconsistent with the non-negative solution of the equation (), and shows that the smallest possible wave speed is given by .
In our case the travelling wave ansatz transforms the system of PDEs (10)–(11) to the following set of coupled ordinary differential equations (ODE):(12)(13)where prime indicates derivative with respect to , and where we have expressed the diffusion coefficients in terms of and . Because and represent occupation probabilities we seek solutions and for all . In order to perform a phase-space analysis we need to transform the coupled ODEs to an autonomous system by introducing the variables and . This expands equation (12) and (13) into the following four-dimensional system:(14)with boundary conditions(15)where(16)and(17)
The boundary conditions reflect the fixed points of the system, which are and , and correspond to the healthy and invaded state respectively. In the limit the invaded fixed point simplifies to , in which case only the relative magnitude of the switching rates determines the equilibrium occupation probabilities (cf. the values of and at in figure 4b and c).
What will help us determine the wave speed is the characteristics of these fixed points, or more precisely the one at the origin. This method only gives a lower bound on the wave speed, but this minimal turns out to be the one attained for relevant initial conditions for the Fisher equation [21], although this requires further proof [40].
We will now apply the same kind of reasoning of non-negativity as for the Fisher equation in order to obtain a minimal wave speed for our system (3)–(4). The properties of the fixed point at the origin are determined by linear stability analysis and depend on the eigenvalues of the Jacobian , where the 's are the right hand sides of equation (14) and correspond to the independent variables and . The Jacobian evaluated at the origin is given by(18)whose eigenvalues are given by the zeros of the characteristic polynomial(19)
The roots of this equation have, for the biologically relevant parameter values and non-zero real part, , implying that the fixed point is hyperbolic and thus that its characteristics are fully determined by linear stability analysis [41]. The aim is now to find the smallest such that all roots of have imaginary part equal to zero, since only then are we guaranteed trajectories which do not oscillate around the origin, and remain positive in the variables and , which is required since these variables represent non-negative occupation probabilities. Determining the smallest such turns out to be intractable from an analytic point of view, and we will therefore resort to numerical solution of the eigenvalue problem.
|
10.1371/journal.pcbi.1000853 | Statistical Analysis of 3D Images Detects Regular Spatial Distributions of Centromeres and Chromocenters in Animal and Plant Nuclei | In eukaryotes, the interphase nucleus is organized in morphologically and/or functionally distinct nuclear “compartments”. Numerous studies highlight functional relationships between the spatial organization of the nucleus and gene regulation. This raises the question of whether nuclear organization principles exist and, if so, whether they are identical in the animal and plant kingdoms. We addressed this issue through the investigation of the three-dimensional distribution of the centromeres and chromocenters. We investigated five very diverse populations of interphase nuclei at different differentiation stages in their physiological environment, belonging to rabbit embryos at the 8-cell and blastocyst stages, differentiated rabbit mammary epithelial cells during lactation, and differentiated cells of Arabidopsis thaliana plantlets. We developed new tools based on the processing of confocal images and a new statistical approach based on G- and F- distance functions used in spatial statistics. Our original computational scheme takes into account both size and shape variability by comparing, for each nucleus, the observed distribution against a reference distribution estimated by Monte-Carlo sampling over the same nucleus. This implicit normalization allowed similar data processing and extraction of rules in the five differentiated nuclei populations of the three studied biological systems, despite differences in chromosome number, genome organization and heterochromatin content. We showed that centromeres/chromocenters form significantly more regularly spaced patterns than expected under a completely random situation, suggesting that repulsive constraints or spatial inhomogeneities underlay the spatial organization of heterochromatic compartments. The proposed technique should be useful for identifying further spatial features in a wide range of cell types.
| Several reports suggest functional relationships within the spatial organization of the nucleus, gene regulation and cell differentiation. However, it still remains difficult to extract common rules, mostly because i) most data have been gathered on limited sets of nuclear elements and in nuclei outside their normal physiological environment, and ii) few three-dimensional (3D) quantitative measures have been performed. Thus, we questioned whether common nuclear organization principles exist in the animal and plant kingdoms. For that purpose, we investigated the 3D distribution of centromeres/chromocenters in five populations of animal and plant nuclei: rabbit embryos at 8-cell and blastocyst stages, rabbit mammary gland epithelial cells and Arabidopsis thaliana plantlets. We set up adapted procedures to segment confocal images and developed a new analytical methodology based on distances between positions within the nucleus and centromeres/chromocenters. We showed that in all systems, despite large differences in chromosome number (44 in rabbit; 10 in A. thaliana) and genome size (rabbit estimated size 2.77 Gbp; A. thaliana 125 Mbp), centromeres/chromocenters form significantly more regularly spaced patterns than expected under a completely random situation. This suggests that, whatever their specific features, conserved rules govern the spatial distribution of genomes in nuclei of differentiated cells.
| In eukaryotes, the interphase nucleus is organized into distinct nuclear “compartments”, defined as macroscopic regions within the nucleus that are morphologically and/or functionally distinct from their surrounding [1]. Complex relationships between the spatial organization of these compartments and the regulation of genome function have been previously described. Furthermore, changes in nuclear architecture are among the most significant features of differentiation, development or malignant processes. Thus, these findings question whether topological landmarks and/or nuclear organization principles exist and, if so, whether these architectural principles are identical in the animal and plant kingdoms. To investigate nuclear organization principles, multidisciplinary approaches are required based on image analysis, computational biology and spatial statistics.
Spatial distributions of several compartments, which can be proteinaceous bodies or genomic domains, have been analyzed. Chromosome territories (CT), areas in which the genetic content of individual chromosomes are confined [2], [3], are usually radially distributed, with gene-rich chromosomes more centrally located than gene-poor chromosomes. Some studies report that chromosome size could also influence CT location [4]–[7]. Centromeres may be close to the nuclear periphery and those located on chromosomes bearing ribosomal genes are generally tethered to the nucleolar periphery [4]. Transcription sites, as well as early replicating foci, assumed to correspond to active chromatin, are more centrally located, whereas inactive heterochromatin tends to be at the nuclear periphery. At a finer level, active genes widely separated in cis or located on different chromosomes can colocalize to active transcription sites [8]–[10], whereas proximity to centromeric heterochromatin or to the nuclear periphery is generally associated with gene silencing [11]–[14]. Changes in the transcriptional status of genes have been frequently associated with their repositioning in the nucleus relative to their CTs, the nuclear periphery or the repressive centromeric heterochromatin [13], [15]–[20]. Furthermore, large reorganization in nuclear architecture (e.g. CTs, heterochromatic compartments, centromeres, speckles, nucleoli,..) can accompany some differentiation, development, malignant processes or natural variations [21]–[32].
However, it still remains difficult to extract common rules and establish comparisons due to various limitations. Indeed, most data have been gathered on limited sets of nuclear elements in isolated plant cell nuclei or in nuclei from immortalized animal cell lines outside their physiological environment, except for circulating blood cells. Little is known about possible differences in nuclear organization of cells within their tissue [33]. Some studies compared nuclear organization in primary cells versus cell lines, in cell lines versus tissues, and in 2D culture versus 3D cultures; these studies suggested that tissue architecture is involved in the control of nuclear organization [34]–[36]. In addition, data on nuclear organization in plant cell nuclei in situ are rare [37], [38]. Finally, few three-dimensional (3D) studies and quantitative measures have been performed to investigate spatial nuclear organization [39]–[42].
The statistics used to analyze the data were mostly based on radial patterns of nuclear elements, such as genes, chromosome territories, and centromeres. Radial positions have been measured with respect to the nuclear geometric center or the nuclear envelope [43], [44]. Spatial affinity between several elements has been investigated and spatial correlations have been assessed through central angles, for example between the radii joining homologous chromosome territory centers and the nuclear center [5], [45]. Alternative approaches based on distances between elements have been developed. Distances between a small number of elements, like two pairs of homologous alleles, were used for testing spatial attraction or repulsion [39]. Remarkably, spatial statistics tools, such as distance functions, that have been developed in ecology or epidemiology for analyzing spatial point patterns [46] have rarely been applied in nuclear organization studies. For example, (cross) nearest-neighbor distances have been used to analyze large numbers of nuclear elements, such as molecular complexes, PML bodies, or RNA Polymerase II foci [47], [48]. Alternatively, all pairwise inter-distances have been used to analyze the spatial distribution of chromocenters [49] and nucleocapsids [50].
In spatial statistics, data are usually collected through a sampling window over a single realization of a point process. This point process is generally considered as unbounded and spatially homogeneous. Such a theoretical framework makes sense in applications in which the investigated phenomenon extends far beyond the observed region. By contrast, analyses of nuclear spatial patterns are based on images of entire nuclei: the whole domain of interest is observed. Furthermore, one should not consider observed nuclear patterns as realizations of spatially homogeneous point processes.
Another difference is that replicated data are available as the analysis is carried out on a sample of nuclei. Recently, distance functions have been extended to replicated spatially heterogeneous point patterns [51], [52]. For instance, an extended F-function has been used for analyzing spatial patterns of transmissible spongiform encephalopathy lesions in brain tissue [53]. The extended F-function takes into account the expected spatial heterogeneity of the point process intensity. To estimate this intensity, the replicated patterns are first registered to locate all observed points in a common coordinate system. However, this type of preliminary registration is not possible for nuclei due to the lack of identifiable nuclear landmarks. Hence, further developments are required to make spatial statistics tools appropriate for nuclear spatial organization studies.
In this study, we develop an approach to furthering the analysis of nuclear spatial organization. Spatial distributions of nuclear compartments were quantified using the cumulative distribution functions of nearest-neighbor distances (G-function) and of distances between arbitrary points within the nucleus and their nearest compartment (F-function). The analysis of G- and F-functions was designed specifically to cope with patterns observed in non-registered and variable (both in size and shape) domains.
We applied this new approach to the investigation of the 3D distribution of centric/pericentric heterochromatin in five interphase nuclei populations belonging to the animal and plant kingdoms [54]. The centric/pericentric compartment was chosen due to its dual structural and regulatory functions. Indeed, it usually behaves as a transcription repressor and is essential for genome organization and the proper segregation of genetic information during cell division [55], [56]. This compartment often clusters and forms chromocenters [57]–[59]. We studied nuclei of cells at various differentiation stages, in three biological systems: rabbit embryos at the 8-cell and blastocyst stages, differentiated rabbit mammary epithelial cells during lactation, and differentiated cells of A. thaliana plantlets.
We found non-completely random and significantly more regularly spaced patterns than expected under complete randomness of the centric/pericentric heterochromatin compartment in the five differentiated cell populations, suggesting the existence of inter-kingdom nuclear organizational rules and possible nuclear regularities.
The most common or comparable markers of the centromeres/chromocenters were chosen in the three biological systems, rabbit embryos, rabbit mammary gland and A. thaliana. The non-histone heterochromatin protein 1 (HP1β) family plays an important role in chromatin organization and transcriptional regulation of both heterochromatin and euchromatin compartments [60], [61]. Several HP1 isoforms are usually present in higher eukaryotes with various specificities and localization [60], [62]. The human or mouse HP1β isoform is usually used as marker for pericentric heterochromatin regions. However, our preliminary experiments revealed that immunodetection of HP1β in rabbit embryo, as well as in rabbit mammary gland nuclei, did not exhibit enough contrast to delineate the pericentromeric heterochromatin blocks (Figure 1A). By contrast, immunolabeling of centromeric proteins (CENP) using sera of patients with autoimmune diseases led to dots with significant differences in contrast, which allowed the positioning of the centromeres. HP1β–labeling was retained to label the whole nucleus in embryos.
In A. thaliana, LHP1, the HP1 homolog, is mainly involved in gene regulation and does not colocalize with centromeric heterochromatin [63], and therefore could not be used to follow heterochromatic centromeres. However, well-defined chromocenters can be revealed by DAPI staining in interphase nuclei, which mostly include centromeric and pericentromeric heterochromatic regions [57], [58].
Therefore, nuclei of rabbit embryos, mammary gland and A. thaliana plantlets were labeled with CENP and HP1β, CENP and DAPI, and DAPI alone, respectively to visualize centromeres/chromocenters and the nuclear volume (Figures 1 and 2).
After acquisition and treatment of a first set of images, capture conditions needed for a proper segmentation and for the best quality measurements were set up. We paid particular attention to i) the setup of the minimal (background) and maximal intensity level, ii) the spacing of the optical planes, and iii) the procedure to limit squeezing between slides and coverslips, particularly in the case of whole embryos. The acquisition parameters defined at this stage remained unchanged for the rest of the project. The resulting collection of images and the acquisition protocols have been deposited on the ICOPAN (“A 3D Image Collection of Plant and Animal Nuclei”) website (http://amib.jouy.inra.fr/icopan).
Five populations of nuclei were analyzed. Nuclei were from rabbit embryos at the 8-cell and blastocyst stages, and from rabbit differentiated mammary epithelial cells (DMEC). DMEC nuclei were easily identified among nuclei of other mammary cell types based on their relative position within the tissue [64]. The DMEC flank the lumen of acini and are surrounded by elongated myoepithelial cells. Both cell types are buried within a stroma composed of adipocytes, fibroblasts and vascular cells. In A. thaliana plantlets, two populations of nuclei were analyzed based on their shapes: rounded or elongated nuclei (Figure 2 A and A′).
The size and shape parameters were determined for the five populations of nuclei and highlighted both a certain nuclear diversity between the various systems and homogeneity within each of them (Table 1, Figures 1 and 2). Shape analysis was detailed by determining flatness, compactness, and elongation indexes to characterize the 3D morphology of the studied nuclei. At the two rabbit embryonic developmental stages, the nuclei compactness value was rather low due to deep invaginations in the nuclear volume (Figure 1A). At the 8-cell stage, flatness was high (1.7, Table 1) and the main direction of flattening was closest to the Z-axis for almost all nuclei (28/29). The observed flattening may be due to the embryos being pressed between slides and coverslips. For blastocyst nuclei, the proportion of nuclei with the main direction of flattening close to the Z-axis was lower (25/41) and the flatness parameter for the 16 other nuclei was close to the overall average (1.36 vs 1.40). This suggested that, although experimental artifacts were partly responsible of the observed flattening, blastocyst nuclei were naturally relatively flat.
The observed DMEC nuclei were rather regular and spherical (high compactness and low flatness values). The main direction of flattening was closest to the Z-axis for most nuclei (57/79) suggesting that the low observed flattening may be partly experimental.
The three rabbit nucleus populations showed unimodal distributions of volume, compactness and elongation, as expected in homogeneous populations. In A. thaliana, the nuclear volume within the population of rounded nuclei exhibited a unimodal distribution (Figure 3A), as did compactness and elongation (data not shown). Flatness distribution was also unimodal and was concentrated in the lower flatness range (Figure 3B). The distributions of the size and shape parameters thus confirmed that, though they were not selected based on cellular type, the rounded nuclei constitute a morphologically homogeneous population. Similar homogeneous distributions were observed within the population of elongated nuclei (data not shown). The main direction of flattening was closest to the Z-axis for 76% (45/59) of the rounded nuclei and flatness was close to 1 (i.e., no or moderate flattening) for the remaining nuclei. Similar observations were made within the population of elongated nuclei, in which 79% (48/58) of nuclei presented flattening oriented along the Z-axis. Thus, nuclear flatness measurements in the five analyzed populations suggested some experimental effects and a natural flatness in rabbit blastocyst nuclei.
Various segmentation procedures were developed to adapt to the size and contrast of the objects. Centromeres in rabbit embryo and mammary gland nuclei were revealed by CENP immunolabeling. In both cases, images were denoised with median and Gaussian filters, and the background lowered with a top-hat transform by size. Some of the CENP spots appeared to be outside of the nucleus mask, because of the elongation caused by the microscope's point spread function. To avoid truncating some of the spots, the nuclear masks were enlarged with a morphological dilation. Objects smaller than 0.02 µm3 were then removed in the masked CENP image.
In rabbit embryo, centromeres could not be extracted using a fixed threshold over all nuclei because of the high level of remaining background signal. Rather, the a priori knowledge of the number of centromeres (44) was used in searching for a threshold value that would produce, at most, 44 connected objects, starting with a threshold of 1 and incrementing by 1.
With this method, mean values of 42.8 and 43.7 centromeres were counted in rabbit 8-cell and blastocyst nuclei, respectively (Table 1). To assess the quality of the segmentation, subsamples of nuclei (6 at the 8-cell stage and 8 at the blastocyst stage) were checked visually. This revealed that at the 8-cell stage, 2.7% of segmented regions turned out not to be associated to HP1β labeling and thus not to be centromeric spots (false positives) and 4.7% of centromeric spots (as assessed by their association with HP1β labeling) were missed by the segmentation (false negatives). The false positive rate was 4.9% at the blastocyst stage, whereas no false negatives were identified. Finally, the centromeres were mapped within the 3D nucleus model (Figure 1F) for subsequent spatial analyses.
In rabbit mammary gland nuclei, a threshold computed as the median of the 11 brightest regional maxima divided by 4 was applied to each image. We identified a mean of about 38 centromeres per nucleus (Table 1). To visually check the result of the segmentation process, all input images were overlaid with their segmentations (Figure 1E′). Among 2996 segmented spots, 30 (1.0%) were considered to be false positives on the basis of their size or position. About 20 centromeres (0.7%) were under the threshold that had been set for intensity or size and were therefore not detected during segmentation (false negatives). The total number of visually detected centromeres was always lower than 44. Centromeres were mapped within the 3D nucleus model (Figure 1F′) for subsequent spatial analysis.
In A. thaliana, chromocenters could not be accurately detected via intensity thresholding. We thus developed an alternative strategy based on the fact that chromocenters have spherical or ellipsoidal shapes and present a positive contrast relative to their immediate neighborhood. Using a 3D watershed transform [65], the nucleus was partitioned into regions (Figure 2B). Each region was assigned a value given by the average intensity in the neighborhood of its barycenter. To correct for possible over-segmentation of chromocenters or nucleoli, region merging was repeatedly applied until all differences between values of adjacent regions were above a predefined threshold. The contrast of non-chromocenter regions adjacent to dark regions, such as the nucleolus, was reduced using a morphological region closing (Figure 2C). The contrast of each region was then computed as the average difference between its value and those of its neighbors, weighted by their sizes to limit the influence of small regions with exceptionally high or low values.
The contrast of each region was multiplied by its compactness to obtain a shape/contrast criterion that enhances chromocenters at the expense of other regions, even if they display similar intensities (Figure 2D). Using the ImageJ software [66], a threshold was then interactively set to a value ensuring the extraction of all chromocenters (Figure 2E). All segmentations were visually checked and compared to the original images by an experienced experimenter. Identified false positives were removed using the Free-D software [67]. Finally, the chromocenter regions were mapped within the 3D nucleus model (Figure 2 F and F′) for subsequent spatial analysis and their sizes quantified by their equivalent spherical diameters.
A few false negatives, generally corresponding to small and weakly labeled chromocenters that had been smoothed out when computing the Gaussian gradient, were also identified during the visual examination of segmented images. For rounded A. thaliana nuclei, the algorithm detected 470 chromocenters and the number of false negatives was 27 (error rate of 5.4%). For the elongated nuclei, 633 chromocenters were detected and 11 false negatives were identified (error rate of 1.7%).
The number of detected chromocenters differed between rounded and elongated nuclei (Table 1). Five to 10 chromocenters (average 8.0±1.5) were detected per nucleus in rounded nuclei. Our results therefore confirmed previously published data indicating that A. thaliana diploid cells (2n = 10) contain 4 to 10 chromocenters, due to a non-random association of homologous chromocenters or the coalescence of chromocenters containing rDNA repeats [58], [68]. Six to 17 (average 10.9±2.4) chromocenters were detected in elongated nuclei. Plants contain cell types with different ploidy levels that may vary from 2C (where 1C is the haploid genome complement) to 64C [69]. Previous studies reported a positive correlation between polyploidy and nuclear volume [38], [70]. Our data thus suggested that elongated nuclei, which on average contained more than 10 chromocenters and were ∼2 times larger than rounded nuclei (Table 1), were extracted, at least for a certain proportion of them, from endoreduplicated cells and that this population of nuclei may represent nuclei from cells that have undergone further differentiation.
Following the image processing stage, chromocenters and centromeres were segmented as regions within nuclei. To analyze their spatial distribution, all regions were represented by their centers of gravity, with, in the A. thaliana case, their equivalent spherical diameters. For the sake of brevity, we refer below to chromocenters/centromeres to mean their centers of gravity, with, in the A. thaliana case, their associated diameters.
Our method encompasses four key steps that can be summarized as follows and will be detailed below:
Distance functions are standard tools in the statistical analysis of spatial point processes [46]. The nearest neighbor distance function G of a point pattern is the cumulative distribution function of the distance X between a typical point (i.e., a uniformly randomly chosen point) of the pattern and its nearest neighbor (Figure 4):Computing this function from the point pattern is straightforward. The F-function is the cumulative distribution function of the distance Y between a typical position within the nucleus and its closest point in the pattern:Thus, F(y) is the nuclear volume fraction that lies at a distance less than y from a point of the pattern. We are considering the distribution of centromeres/chromocenters as a finite process within the bounded nuclear space. To analyze populations of nuclei, we developed an original strategy, based on F- and G-functions, which does not require nucleus registration.
A stochastic scheme was adopted to compute the F-function corresponding to a nuclear point pattern. A number of independent evaluation points NE were generated uniformly at random within the nucleus. For each evaluation point, the distance to the closest point of the pattern was determined (Figure 4). The cumulative distribution function F(y) was then estimated by the proportion of evaluation points for which this distance is below y. Setting NE to 10000 was sufficient to smooth out the effect of evaluation point sampling on the F-function.
To determine whether the spatial distribution of centromeres/chromocenters obeys any organizational rule, the observed distributions were compared against a completely random distribution, conditioned on the observed numbers of centromeres/chromocenters and, in A. thaliana, on chromocenter sizes. Due to the arbitrary shape of the nucleus, the expected distance functions under CRBPP cannot be determined analytically. A Monte-Carlo approach was therefore adopted, whereby the distance functions were computed over sets of patterns simulated according to CRBPP. For each nucleus, random patterns were generated with the same number of centromeres/chromocenters as detected within the nucleus. In A. thaliana, each random point was also assigned the radius of one chromocenter. This hardcore distance defined a sphere within which no other point was allowed to fall and a minimum distance between the point and the nuclear envelope. Taking care of rabbit centromere sizes was not necessary because of their small size. The CRBPP distance functions were estimated by computing averages over a number P1 = 500 of such independent patterns.
Observed and CRBPP mean F-functions obtained for the three nuclei from Figures 1A, 1A′ and 2A are displayed in Figure 5 (A: rabbit embryo; B: rabbit mammary gland; C: A. thaliana leaflets). As illustrated by these examples, the observed F-functions were frequently located on the left side of the CRBPP ones, and presented a steeper slope. On average, the observed distance between any nuclear position and the closest centromere/chromocenter was thus lower and less variable than expected under CRBPP. This suggests a non-completely random and regular distribution of centromeres/chromocenters within the nucleus (see Figure 4). Discerning any particular trend by visually examining G-functions was much more difficult (data not shown). This may be due to the fact that i) this function is potentially less discriminant than the F-function and the fact that ii) it was estimated from a smaller pool of data (number of detected centromeres or chromocenters compared with an arbitrary number of arbitrary positions for F-function).
The next step in our analysis was to test, at the population level, the statistical significance of the differences between the observed F- and G-functions and the CRBPP theoretical F- and G- functions. Due to the arbitrary shape of the nucleus, the fluctuations under CRBPP of the distance functions around their averages are not analytically accessible and a Monte-Carlo approach was designed. To avoid under-estimation of these variations, they were estimated using a second set of randomly generated P2 = 500 patterns. For each simulated pattern, the difference between its distance function and the CRBPP theoretical function was defined as the signed difference of maximum amplitude. Taking for example the F-function, we thus have:where F0 is the F-function under CRBPP computed using the Monte-Carlo approach described above.
We also computed the difference between the observed pattern distance function and that of the CRBPP; this yielded a total of P2+1 differences. A p-value (the probability of observing, under CRBPP, a difference at least as large as that observed) could then be computed for each nucleus as the proportion of random patterns with a difference equal to or above that observed. Since it quantifies a spatial repartition, this p-value was called spatial distribution index (SDI). For example, low values of the SDI associated to the F-function indicate regularity in the patterns (evenly distributed points) while high values correspond to clustered patterns. Under the hypothesis that centromeres/chromocenters obey CRBPP, the SDI within a population is uniformly distributed between 0 and 1. Our test thus consists in comparing the observed SDI distribution with the uniform distribution. This was done using the two-sided Kolmogorov-Smirnov test (α = 5%) in the R statistical software package [72].
Within the five groups, the distributions of the SDI based on F-function were significantly different from the uniform distribution (Figure 6 and Table 2). Hence, the spatial distributions of centromeres and chromocenters are different from the completely random distribution. Besides, the histograms were concentrated in the lower range of SDI (Figure 6), meaning that the observed F-functions were generally above the CRBPP ones. This analysis thus demonstrated that the centromeres or chromocenters tend to form more regularly spaced patterns than expected under CRBPP. For G-functions, the distributions of the SDI within the five groups (data not shown) were also significantly different from the uniform distribution (Table 2). The SDI histograms were concentrated in the upper range, meaning that the observed G-functions were generally below the CRBPP ones and that the nearest centromere/chromocenter was on average farther away than expected under CRBPP. Thus, though departure from complete randomness was less pronounced, analysis by G-function was consistent with the results obtained with F-function.
Lastly, we examined whether the regularity of the spatial distribution of centromeres or chromocenters could be explained by the experimental flattening of the nuclei along the Z direction. A link between flatness and SDI was tested on the nuclei with a Z-oriented minor axis. As flatness (e.g., Figure 3B) and rank (Figure 6) distributions largely deviate from normality, a non-parametric test based on Kendall's tau rank correlation [73] was applied (α = 5%). In A. thaliana (rounded and elongated nuclei) and blastocyst nuclei, no correlation was found between flatness and the F-function-associated SDI (Table 2). In 8-cell embryo and mammary gland nuclei, correlations between flatness and F-function SDI turned out to be significant (Table 2). However, the values of tau remained rather small (about 30%). Therefore, even for those two latter cell populations, the regularity of centromere patterns cannot be considered as caused by flattening. Similarly, no relation was found between nucleus flatness and the SDI associated with G-function. Consequently, the possibility that flattening along the Z direction is responsible for the regular spatial distribution of centromeres/chromocenters was ruled out.
The Rabl configuration observed in salamander [74], fission yeast [75], drosophila embryos [76], as well as in some plants such as Allium cepia [77], wheat, rye, barley and oats [78], [79] but not in A. thaliana nor in mammals, is probably the most spectacular example of a non-random organization in interphase nuclei. In this configuration, the centromeres are clustered at one end of the nucleus and telomeres at the opposite end. Besides this extremely recognizable organization, detection of any regularity in the complex nuclear organization is difficult and requires the development of specific 3D image analysis and statistical tools. Indeed, as nuclei exhibit large morphological fluctuations both within and between cellular types [80], spatial normalization is required. However, no nuclear landmark has yet been identified as a suitable reference to perform this standardization [40]. In this study, we addressed this issue and developed original tools adapted to various biological models, allowing a common treatment of data and extraction of rules. Building on well-established distance functions from spatial statistics, we designed a new methodological framework to analyze replicated samples without explicit nucleus registration.
Spatial statistics distance functions have only rarely been applied to study nuclear organizations. Nearest neighbor distances (G-function) have been used to characterize the spatial distribution of transcription factors [47], [48], [81], centromeres [49], and other nuclear compartments [82]. The distribution of all inter-distances (quantified through the pair correlation function or the K- and L-functions) has also been considered [49], [81]. To our knowledge, the present study is the first to rely on the F-function (the cumulative distribution of the distance from arbitrary nuclear positions to the nearest centromere/chromocenter) to investigate nuclear architecture. In contrast to the F-function, the G, K and L-functions and the pair correlation function only depend on the relative positioning of the points within the pattern, irrespective of their absolute positioning within the nucleus. For instance, a translation of a clustered pattern from the nucleus center to its periphery has no effect on the G-function while it will tend to shift the F-function towards the left (larger empty nuclear region). Thus, in the finite and bounded context of nuclear organization studies, the F-function captures more information about spatial repartition and therefore represents a potentially more discriminant and sensitive tool to detect differences between spatial distributions. In accordance with this, our results showed more pronounced departure from complete randomness with F-function than with G-function, be it in individual function plots or SDI distributions. The J-function, a more recently introduced function [83], combines the F- and G-functions and has an easy interpretation for point patterns with random size. However, it needs to be further elaborated for analyses conditioned on the number of points.
The completely random patterns that were simulated here contained a number of points equal to that actually observed in the patterns and not to the number of chromosomes in the species. In A. thaliana, it is known that centromeres aggregate into a variable number of chromocenters. In the rabbit, exactly 44 centromeres have never been visually observed in interphase nuclei, be it because of aggregation, limited optical resolution, or labeling issues. Hence, the number of chromosomes only provides an upper bound on the expected number of points. For this reason, our analysis has focused on the spatial distribution of observed patterns.
The F-function was computed using a Monte-Carlo scheme for sampling the nuclear space. An alternative computational scheme to estimate the F-function consists in computing the Euclidean distance map (EDM) [84], [85], which gives the distance between each voxel of the nucleus and the closest voxel containing a centromere or chromocenter center of gravity; the F-function is then given by the normalized cumulative histogram of the distance map. In our preliminary experiments, we have implemented, tested, and compared the two approaches. We have retained the estimation of F based on randomly generated points throughout the nucleus for essentially two reasons. The first one is that this approach retains the subpixel precision at which the positions (centers of mass) of centromeres and chromocenters are computed. On the contrary, using distance maps introduces a loss of precision because these positions must be rounded to voxel (integer) coordinates. This is not critical in our case because the voxel size is small as compared to the nucleus size. However, this effect could bias the analysis when processing data extracted from images with large voxel sizes. The second reason is computational efficiency: our experience did not reveal any computational advantage of the EDM approach over the stochastic one.
Standard methods for computing empirical G- and F-functions usually include boundary corrections [86]. Since point patterns are traditionally observed within sampling windows, nearest point distances can indeed be over-estimated due to the exclusion of some points from the experiment. For the G-function, for example, the Hanisch correction [87] consists in discarding the recorded points whose nearest neighbor is farther than the boundary. Such standard estimation corrections have been applied in studies on nuclear patterns [49], [81]. In this context, however, no point (here, centromere or chromocenter) is expected outside the nucleus. Hence, edge-corrections should be all the more avoided. Indeed they may decrease statistical power by reducing the number of analyzed points and potentially bias the analysis, especially if the assumed sampling window is computed from the pattern itself [49]. To obtain unbiased estimates of G- and F-functions, we proposed an alternative computational scheme that takes into account the actual boundary of the nucleus and involves no edge-correction.
Previous spatial statistical analyses of populations of nuclei have been conducted based on distance functions either averaged without normalization [49], [81] or pooled after standardization with respect to the greatest inter-object distance, to account for nuclear size fluctuations [47]. This left the difficult issue of nuclear shape normalization unsolved. A first significant contribution of the approach described here is to take into account both size and shape variability. This was achieved by comparing, for each nucleus, the observed distribution against a reference distribution estimated by Monte-Carlo sampling within the same nucleus. This implicit normalization (each nucleus being its own reference) circumvents the unfeasibility of an explicit nucleus registration in the absence of identified nuclear reference points [40]. A second significant contribution of the present study is the introduction of a test for complete randomness at the population level. For each nucleus, the departure from the completely random spatial distribution was quantified through a spatial distribution index (SDI). This SDI could have been used to independently classify each centromere or chromocenter pattern as completely random (CRBPP) or not. Then a global conclusion concerning the population level could have been drawn by a simple proportion test. However, by such a binarization of the SDIs, one focuses mainly on extreme patterns (clustered or regular) and may fail to detect slight deviations from complete randomness. Avoiding binarization gives more sensitivity to detect spatial structure.
Overall, our methodology therefore allows for sensitive and unbiased statistical assessment of distribution differences against reference distributions. Using this approach, we showed that, in three biological systems belonging to plant and animal kingdoms and in the five types of interphase nuclei, centromeres/chromocenters form significantly more regularly spaced patterns than expected under a complete random situation. Interestingly, this feature was found in biological systems with extremely different numbers of chromosomes (44 in rabbit versus 10 in A. thaliana) and different genome sizes (rabbit estimated size 2.77 Gbp; J. Johnson, Broad Institute, personal communication [https://www.broadinstitute.org/ftp/pub/assemblies/mammals/rabbit/oryCun2/Stats.pdf]; A. thaliana last estimated size 125 Mbp [http://arabidopsis.org/portals/genAnnotation/], [88]). This suggests that conserved rules govern the spatial distribution of genomes, whatever their specific features.
In A. thaliana, 2D analyses suggested non-uniform distributions of some nuclear elements, such as telomeres, mostly localized in the vicinity of the nucleolus [89], as well as the non-random association of homologous chromocenters or chromocenters that contain homologous rDNA repeats [58]. More recent studies based on 3D imaging suggested the existence of a radial arrangement pattern in A. thaliana interphase nuclei, with centromeres localized predominantly at the nuclear periphery [43], [90]. Our results are consistent with a non-random distribution of A. thaliana heterochromatic domains, but further demonstrate their tendency to form regular patterns. The prevailing hypothesis about the spatial organization of A. thaliana interphase nuclei is that of a globally random distribution of chromosome territories under non-specific constraints and interactions [2], [90], [91]. According to this view, the regular distribution of chromocenters we have observed could merely result from the partitioning of the nucleus into distinct chromosome territories. Whether more specific mechanisms of mutual repulsion exist remains to be elucidated, but in any case, the concomitant aggregation of some centromeres into chromocenters and the regular distribution of chromocenters suggest that constraints or control mechanisms are exerted at multiple scales to determine the positioning of heterochromatic domains in A. thaliana nuclei.
A priori, a SDI distribution may differ from CRBPP due to spatial heterogeneity or due to spatial interactions (or both). The volume occupied by the nucleoli, which can be rather large in some biological systems (rabbit), are sources of heterogeneity that have not been taken into account in our analyses since simulated points could fall anywhere within the nucleus. However, the only bias that may have resulted from neglecting these excluded nuclear regions is a bias toward artifact aggregation. Preliminary calculations, performed on a subset of rabbit nuclei, showed that taking into account the nucleolus volume accentuated the deviation from the CRBPP (data not shown). This confirmed that the observed regular dispersion of centromeres/chromocenters was not due to nucleoli volume omission. Nevertheless, further developments will be necessary to integrate the nucleoli into the 3D nuclear modeling and analyses. Other spatial inhomogeneities, due for example to the size and nuclear position of the chromosome territories, may also impact on the distribution of centromeres/chromocenters, especially in species with a large range of chromosome sizes and shapes (e.g. rabbit) [92]. It would therefore be interesting to refine the model by taking into account, whenever possible, the identity of all or most of centromeres/chromocenters. In rabbit, probes allowing the detection of four centromeric DNA families are already available [93].
Previous authors have reported clustering of pericentric heterochromatin during terminal differentiation, e.g. during myogenenesis [94], in human neutrophils [28], human and mouse neurons [95], [96] and rat myoblasts [97]. In these biological models, centromeres cluster into chromocenters and comparisons between undifferentiated and differentiated cell nuclei show that the number of chromocenters decreases during the differentiation process. These results may seem contradictory with the regular patterns revealed by the F-function, especially in A. thaliana in which chromocenters are also observed. However, our analysis concerned the spatial distribution of chromocenters instead of their number. In particular, each given observed pattern of chromocenters was compared with simulated random patterns of the same size. Results showed that the observed patterns are more regularly spaced than completely random patterns of the same size. Therefore, it appears that centromeres in A. thaliana differentiated cells cluster into chromocenters that are regularly spaced within the nucleus. In differentiated rabbit cells, centromeres do not form chromocenters, but the numbers of detected centromeres in mammary gland cells indicate that a small fraction of centromeres may cluster. In the rabbit mammary gland model, centromeres and a few small clusters of centromeres are more regularly spaced than completely random patterns.
The approach and the tools we developed could now be applied to other nuclear compartments and -even more interestingly- to other differentiation stages, to determine whether this tendency to form more regularly spaced patterns than expected under complete randomness might represent a “signature” of the differentiated states. This methodology can also be generalized to investigate further the properties of intra-nuclear spatial distributions. Indeed, our strategy consists in comparing, on the basis of appropriate distance functions, observed nuclear organizations to patterns simulated from a reference distribution. Introducing constraints on the simulated point patterns, as for example interactions with the nuclear envelope or other nuclear compartments, will allow evaluating further hypotheses beyond randomness.
All experiments involving animals were carried out according to European regulations on animal welfare.
All animals were handled following ethical rules for animal welfare according to the INRA ethics statement. Rabbit embryos were obtained by natural fertilization of superovulated mature New Zealand White rabbit females, as already described [98], [99]. Superovulation was induced by two 0.25 mg, two 0.65 mg and one 0.25 mg intramuscular injections of follicle-stimulating hormone (FSH, Stimufol, Mérial, Lyon, France) given 12 hours apart. Females were mated with males 12 hours after the last FSH injection and 30 IU of human chorionic gonadotropin (hCG, Choluron, Intervet, Angers, France) was injected a few minutes after mating. In rabbit, fertilization occurred at ∼12 hours post coitum (hpc). Two-cell embryos were collected in the rabbit oviduct at 24 hpc and were further cultured in vitro in B2 medium supplemented with 2.5% fetal calf serum (Sigma) in 5% CO2 atmosphere at 38.5°C until the 8-cell (48 hpc) and blastocyst (100 hpc) stages.
Mammary glands were collected from 16-day lactating New Zealand INRA-1077 rabbits, one hour after suckling. Mammary gland tissues were cut into small pieces, fixed in 4% paraformaldehyde (PFA) in phosphate-buffered saline (PBS) for 30 min at room temperature (RT), washed three times with PBS and equilibrated in 40% sucrose before embedding in Cryomount (Histolab) and snap freezing in liquid nitrogen. Samples were conserved at −80°C. Cryosections of about 14 µm in thickness were prepared on a Reichert Jung cryo-microtome (Leica, Wetzlar, Germany), deposited on slides (SuperFrost Plus glass slides, Menzel-Gläser J1800AMNZ) and stored at −80°C until use.
A. thaliana plantlets (Col-0 accession) were grown in vitro as previously described [25]. Three-week-old plantlets were fixed for 30 min in 4% PFA in PBS buffer (PFA-PBS), under vacuum, at room temperature. The fixative was replaced, and plantlets were fixed for an additional 30 min in the same conditions. Up to 8 fixed seedlings were transferred into an Eppendorf tube and gently ground in 500 µl of extraction buffer (10 mM Tris HCl pH 7, 4 mM spermidine, 1 mM spermine, 5 mM MgCl2, 0.1% triton X-100, 5 mM β-mercaptoethanol). Nuclei suspension was filtered through a 50 µm nylon mesh. After gentle centrifugation (500× g, 3 min), the pellet was washed in 1× PBS, treated with 0.5% triton X-100 in PBS and washed in PBS. Nuclei were resuspended in 30 µl PBS.
Rabbit embryos at 8-cell and blastocyst stages were fixed overnight at 4°C in 4% PFA-PBS, permeabilized 30 min at RT with 0.5% Triton X-100, and blocked with 3% bovine serum albumin in PBS (BSA-PBS) for 1 hour [30].
Fixed mammary gland sections were incubated in 50 mM NH4Cl in PBS for 15 min and washed with PBS. They were then permeabilized with 0.5% Triton X-100 for 30 min at RT, washed again with PBS, and blocked with 2% BSA-PBS for 1 hour at RT.
Immunoprocessing was then similar for rabbit embryos and mammary gland sections. Fixed embryos and slides with fixed mammary gland sections were incubated with the primary antibodies overnight at 4°C in 2% BSA-PBS. After three washes with 0.05% or 0.1% tween-20 in PBS at RT (15 min each), incubation with the secondary antibodies was performed for 1 hour in 2% BSA-PBS at RT followed by three washes (10 min each) with 0.1% tween-20 in PBS at RT. For double immunostaining, primary antibodies and secondary antibodies were mixed together at the same final dilutions as for simple immunodetection. Rabbit embryos were then deposited on slides and mounted in VECTASHIELD medium (Vector laboratories, Burlingame, CA). Mammary gland sections were washed once in PBS, counterstained with DAPI (1 µg/ml in PBS for 5 min, at RT), washed in PBS for 5 min at RT and mounted under a coverslip with ProLong Gold antifade reagent (Invitrogen).
The suspension of A. thaliana nuclei was spotted on a slide and left to evaporate at 4°C for 20 min, before mounting in VECTASHIELD medium with 1 µg/µl of DAPI for DNA counterstaining.
HP1β was detected with a mouse monoclonal anti-HP1β antibody (clone 1 MOD 1A9, dilution 1∶250), and revealed with a lissamine–rhodamine-conjugated anti-mouse secondary antibody (Jackson ImmunoResearch, dilution 1∶150). Centromeres were detected in rabbit embryos and in rabbit mammary glands by a human autoantibody against centromeres (HCT-0100, Immunovision, dilution1∶300) followed by FITC-conjugated donkey anti-human antibody (Jackson ImmunoResearch, dilutions 1∶150 in rabbit embryos and 1∶300 in rabbit mammary glands).
Embryos were scanned with a Zeiss LSM 510 confocal laser-scanning microscope equipped with a ×63/NA 1.4 oil immersion objective. Z stack images were acquired at intervals of 0.24 µm with 488-, 543- and 633-nm wavelengths of the lasers and with an XY voxel size of 0.04 µm.
Images of mammary gland sections were captured with an optical sectioning microscope attached to an AxioObserver imaging Apotome system (Zeiss) (×63/NA 1.4 oil immersion objective). Z stack images were acquired at intervals of 0.24 µm on two channels (DAPI and FITC), with an XY voxel size of 0.1 µm.
A. thaliana nuclei images were captured on a Leica DM IRE2 SP2 AOBS spectral confocal microscope equipped with a 405 nm diode (25 mW) using a ×63 HCX PL APO objective (NA 1.2). Z stack images were acquired at intervals of 0.122 µm, with an XY voxel size of 0.05 µm.
The anisotropy of voxel sizes in XY-Z was taken into account in all subsequent image processing and spatial analysis procedures.
Images can be freely retrieved in their native formats together with detailed acquisition protocols at the ICOPAN Website (http://amib.jouy.inra.fr/icopan).
Nucleus contours were determined on the HP1β (embryos) and DAPI (mammary gland and A. thaliana) images.
Images of rabbit nuclei were denoised with a median filter and a Gaussian filter. They were subsequently segmented through two different pathways.
HP1β images (embryos), on which several nuclei are present, were analyzed with the Insight Toolkit (ITK) library. The robust automatic threshold selection method (RATS) [100] was used to compute a threshold to ‘binarize’ the HP1β images. The threshold is computed as the mean of the intensity values in the HP1β image weighted by their Gaussian gradient magnitude. To avoid the high gradient values in the nucleus caused by non-homogeneous content, the small bright and dark zones were removed with a 3D area opening and a grayscale fill hole transformation before computing the gradient. The joined masks of nuclei were separated using a watershed transform on the distance map. Truncated nuclei at the image border as well as objects smaller than 200 µm3 were removed.
A semi-automated procedure was developed to segment mammary gland nuclei from the DAPI image. DAPI signal was denoised with a median and a Gaussian filter, and manually thresholded to produce partial nuclear masks. The DAPI signal was mostly present on the border of the nuclei. As a result, thresholding this signal results in an incomplete nucleus, in which the center is not filled and the border is not continuous. The nuclear borders were thus closed with a morphological closing transform with a large round kernel, and content of the nuclei was filled with a binary hole filling transform. The masks of the different nuclei were then separated by a watershed transform on the distance map, and the nuclei from the cell types of interest were manually selected.
Confocal image stacks of A. thaliana nuclei were processed and analyzed with programs developed using the Free-D software libraries [67]. Each image stack contained a single nucleus. Images were automatically cropped to limit processing to a bounding box surrounding the nucleus. To separate the nucleus from the background, a preliminary intensity threshold was then computed using the isodata algorithm [101]. This algorithm is sensitive to the relative size of the nucleus within the image. As a result, the threshold was generally too high because of the larger background size. To correct for this bias, the intensity average m and standard-deviation s were computed over the nucleus region defined by the preliminary threshold and the actual threshold was set to m-2s. The resulting binary image generally contained holes, corresponding in particular to the nucleolus, and presented boundary irregularities due to noise. In addition, bumps were also observed on some nuclei at their basal and apical faces, because of blur from chromocenters [39]. Hole filling, opening and closing binary morphological operators [65] were therefore applied to regularize the binary image. The subsequent processing and analyses were confined to this final nucleus mask. A surface model of the nuclear envelope was generated by applying the marching cubes algorithm [102] to the binary mask.
Nuclear size was quantified by nuclear volume. Nuclear shape was quantified using the compactness parameter, which is given by:This parameter takes its maximum value 1.0 for a sphere and decreases toward 0.0 as the shape surface becomes less regular.
Visual image examination revealed that some nuclei were flattened along the Z-axis; thus, a flatness parameter was defined, based on the lengths of the principal axes of the nuclear surface:Symmetrically, an elongation parameter was also defined:The main direction of flattening (resp. elongation) was defined by the coordinate frame axis that was the closest to the shortest (resp. the longest) principal axis of the nucleus.
|
10.1371/journal.pcbi.1003369 | Protein-Protein Interactions in a Crowded Environment: An Analysis via Cross-Docking Simulations and Evolutionary Information | Large-scale analyses of protein-protein interactions based on coarse-grain molecular docking simulations and binding site predictions resulting from evolutionary sequence analysis, are possible and realizable on hundreds of proteins with variate structures and interfaces. We demonstrated this on the 168 proteins of the Mintseris Benchmark 2.0. On the one hand, we evaluated the quality of the interaction signal and the contribution of docking information compared to evolutionary information showing that the combination of the two improves partner identification. On the other hand, since protein interactions usually occur in crowded environments with several competing partners, we realized a thorough analysis of the interactions of proteins with true partners but also with non-partners to evaluate whether proteins in the environment, competing with the true partner, affect its identification. We found three populations of proteins: strongly competing, never competing, and interacting with different levels of strength. Populations and levels of strength are numerically characterized and provide a signature for the behavior of a protein in the crowded environment. We showed that partner identification, to some extent, does not depend on the competing partners present in the environment, that certain biochemical classes of proteins are intrinsically easier to analyze than others, and that small proteins are not more promiscuous than large ones. Our approach brings to light that the knowledge of the binding site can be used to reduce the high computational cost of docking simulations with no consequence in the quality of the results, demonstrating the possibility to apply coarse-grain docking to datasets made of thousands of proteins. Comparison with all available large-scale analyses aimed to partner predictions is realized. We release the complete decoys set issued by coarse-grain docking simulations of both true and false interacting partners, and their evolutionary sequence analysis leading to binding site predictions. Download site: http://www.lgm.upmc.fr/CCDMintseris/
| Protein-protein interactions (PPI) are at the heart of the molecular processes governing life and constitute an increasingly important target for drug design. Given their importance, it is vital to determine which protein interactions have functional relevance and to characterize the protein competition inherent to crowded environments, as the cytoplasm or the cellular organelles. We show that combining coarse-grain molecular cross-docking simulations and binding site predictions based on evolutionary sequence analysis is a viable route to identify true interacting partners for hundreds of proteins with a variate set of protein structures and interfaces. Also, we realize a large-scale analysis of protein binding promiscuity and provide a numerical characterization of partner competition and level of interaction strength for about 28000 false-partner interactions. Finally, we demonstrate that binding site prediction is useful to discriminate native partners, but also to scale up the approach to thousands of protein interactions. This study is based on the large computational effort made by thousands of internautes helping World Community Grid over a period of 7 months. The complete dataset issued by the computation and the analysis is released to the scientific community.
| Protein-protein interactions (PPI) are at the heart of the molecular processes governing life and constitute an increasingly important target for drug design [1]–[4]. Given their importance, it is clearly vital to characterize PPIs and notably to determine which protein interactions are likely to be stable enough to have functional relevance. Computational methods such as molecular docking have rendered possible to successfully predict the conformation of protein-protein complexes when no major conformational rearrangement occurs during the assembly [5]–[11]. However, we [12] and others [13], [14] have demonstrated that docking algorithms are unable to predict binding affinities and thus, presently, cannot distinguish which proteins will actually interact. This leads to ask whether this failure comes from the fact that scoring functions, used to sort the docking solutions, are inefficient for partner identification or whether the difficulty comes from binding promiscuity between proteins in the cell that blurs the interaction signal of the functional partners. In the crowded cell, proteins experience non-specific and unintended interactions with the intracellular environment leading to a severe competition between functional and non-functional partners [15]–[19]. This brings to light the importance of characterizing weak, potentially non-functional, interactions in order to predict functional ones and understand how proteins behave within a crowded environment [16], [20], [21].
In this work, we tackle two distinct but related questions: (i) can a combination of coarse-grain docking and evolutionary information identify true interacting partners among a set of potential ones? (ii) what is the effect of binding promiscuity on a large and variate dataset of protein structures [22]?
Previously, we have shown that knowing the experimental binding site of a protein can help to retrieve its native interacting partner within a set of decoys [12]. On the other hand, recent studies reveal that arbitrary docked partners bind in a non-random mode on protein surfaces [23], [24] suggesting that docking true but also false partners can help to identify protein binding sites. We developed a novel score based on arbitrary docking and evolutionary information to predict protein binding sites. The different docking conformations of a given protein pair are scored according to their associated energy and the agreement between the docked interface and the predicted binding sites. An interaction index is defined, and normalized according to the whole set of proteins tested, in order to discriminate the interacting partners from the set of tested interactions.
We evaluate our method with a complete cross-docking (CC-D) calculation on a set of 168 proteins belonging to the 84 known complexes described in the Mintseris Benchmark 2.0 [25] and covering a large spectrum of different protein interfaces. Enzymes, inhibitors, antibodies, antigens, signaling proteins and others have been considered as well as interfaces that do or do not undergo conformational adjustments during interaction. Docking calculations are made with no knowledge of the experimental complex structure: unbound structures are used. We use a coarse-grain docking algorithm [12], whose energy function relies on both van der Waals and Coulomb potentials. We show that the combination of a coarse-grain docking algorithm with binding sites prediction can significantly contribute to the identification of a reasonably sized set of potentially interacting proteins, that can be further investigated by more precise docking algorithms or laboratory experiments.
The large computational effort necessary to accomplish this work was realized with the help of World Community Grid (WCG), that coordinated thousands of internautes providing their computer time to dock about 300000 conformations per protein pair for the set of 28224 possible pairs in the Mintseris Benchmark 2.0. For each pair, we selected about 2000 decoys. For non-partners, we find weak as well as strong interactions. The decoy set is released and it provides an important reference set of structures that can serve as a proxy for the non-specific protein-protein complexes that occur transiently in the cell or that are avoided by spatial-temporal constraints. These latter are hard to characterize experimentally but they are of biochemical relevance, as highlighted by other studies [26]–[29].
To simulate the variability of crowded environments for a protein in the cell, we study how easily a protein finds its true partner with respect to many random subsets of proteins supposedly competing with it. We realize a thorough analysis of these interactions and we address the question of whether a successful prediction of a protein partner depends on the environment composition or not. We quantify the effect of competing partners in predictions, and we characterize in a quantitative manner three distinguished populations of proteins interacting with a protein : those that strongly compete with the true partner of , those that never compete with it, and those that interact with with variable levels of strength. For each protein , we propose a numerical index that provides the strength of the interaction with all other proteins in the environment, and that gives a signature for .
To our knowledge, this is the first study performing a large-scale CC-D calculation, proposing an analysis of the binding promiscuity of the protein set, and providing to the scientific community the associated dataset of decoys [30], [31] at the same time. Previous large-scale analyses used docking by shape complementarity that quickly scans through several thousands proteins in a matter of seconds [32], [33] but ignore the electrostatic contribution playing however an important role in protein interactions [34]–[37]. We compared our method to two previously done studies [32], [33]. Both of them do not perform a CC-D experiment, but a large-scale analysis of selected protein pairs.
Finally, we checked whether evolutionary information can be used to considerably restrict the number of docking interfaces to be examined and to render molecular computation feasible for a larger scale investigation of PPIs, based on thousands of proteins instead of hundreds. This result makes the protocol proposed here feasible for scaling up the analysis.
The 168 proteins of the Mintseris Benchmark 2.0 [25] form 84 binary complexes known to interact in the cell. They cover three broad biochemical categories and three difficulty categories related to the degree of conformational change at the protein-protein interface. They are classified as Enzyme-Inhibitors (46 proteins), Antibody-Antigen (20), Antigen-Bound Antibody (24), Others (78), and also as Rigid Body (126), Medium (26), Difficult (16). The set is constituted by 51 multimeric proteins and by 117 monomeric ones forming 41 complexes where at least one of the protein is multimeric.
CC-D was realized on the full dataset from unbound structures, leading to 28224 docking simulations. Each calculation explored about 300000 ligand-receptor orientations, corresponding to ligand and receptor complete surfaces, and asked for more than 7 months computational time on WCG. This CC-D scaled up the one introduced in [12], carried out on 6 enzyme-inhibitor complexes.
The docking algorithm simulates the actual docking process in which ligand-receptor pairwise interaction energies are calculated. The energy function we used takes into account van der Waals (modeled by a Lennard-Jones potential) and electrostatic (modeled by a Coulomb potential) terms (see Methods).
For each protein in the dataset, the problem of partner identification is tackled with two main experiments. The first experiment assumes to know the residues belonging to the experimental interface of the proteins. This means that the residues lying at the interface of two proteins in a native complex are supposed to be known while no knowledge of the complex conformation is assumed. The second experiment replaces experimental interfaces by predictions of binding sites based on docking and evolutionary information. The evaluation of the quality of the interaction signal in this PPI large-scale study is of major importance. In particular, the contribution of docking information compared to evolutionary information in partner identification needs to be quantified. To do so, the analysis based on experimental interfaces allows us to evaluate in a precise manner how much a good prediction of the interaction sites improves partners identification, experimental interfaces playing the role of perfect predictions. In the sequel, we also use it to decipher whether a property of protein interactions that has been observed from computational predictions has a biological origin or whether it is a consequence of the noise of the prediction.
We performed a series of tests to check whether the composition of a set of competing partners for a given protein influences partnership prediction. The analysis is performed on both JET+NIP predictions and experimental interfaces (see Figures S10–S16, S17–S23 and Table S1 in Text S1).
A few large-scale studies that wish to identify true interacting partners among a set of potential ones, have been recently proposed. They are computationally demanding and they remain, for this reason, rare. All large-scale studies we compared to have been based on shape complementarity to quickly scan through several thousand ligands in a matter of seconds. These approaches do not include any electrostatic component in their energy model, while electrostatic forces are known to play an important role in PPI.
Notice that, given a protein , no other docking studies besides this one tries to quantify the effect of binding promiscuity of a large and variate dataset of protein structures interacting with .
The docking technique we used is computationally expensive (see “Computational implementation and data analysis” in Methods). To reduce the conformation space to be explored, we predicted the location where the interaction takes place and confined the docking to this region. This is done by predicting binding sites for the receptor protein by using JET [38] and by defining an appropriate cone around the predicted interface (see Methods and Figures S5, S6 in Text S1). When restricting the docking conformational space with JET, we observe a slight decrease of the AUC. By using experimental data, the AUC goes from 0.84 to 0.80 while using predictions, it goes from 0.61 to 0.59 (Table 4), revealing a reduced loss in precision. This shows that using evolutionary information from sequences is a very promising approach to reduce docking computational time.
To evaluate the impact of our restriction on MAXDo execution time, we computed how many docked conformations between protein pairs were dropped. When the 168 proteins are considered together, the average portion of the conformational space that is explored after reduction is of the original space. This value should be understood at the light of protein sizes, as illustrated in Figure S59 in Text S1. In fact, small proteins require to explore about of their original conformational space, while for large ones, the space is reduced to of the initial one. This is because small proteins are rather conserved and JET predicts large patches as their interaction sites, covering a large portion of their protein surface. Notice that this calculation takes into account a reduced number of conformations for the receptor, independently on whether the conformational space of the ligand is completely explored or not. Clearly, the actual computational time depends on the number of conformations that are tested, and if both the conformational spaces of the receptor and of the ligand are reduced, the effect will be quadratic. The small difference in AUC obtained by exploring the reduced space of the receptor compared to the whole (with a fully explored surface of the ligand), is due to the high specificity of JET and to the definition of the cone (see Methods) that takes into account JET's lower PPV.
We have addressed the problem of predicting protein interactions using high-throughput CC-D calculations on a dataset of 168 proteins. We have shown that a simple docking algorithm combined with evolutionary information, can be used to discriminate interacting from non-interacting proteins. The purpose of the method is the in silico large-scale screening of protein structures to find a small set of potential protein partners that could be tested experimentally. The approach reminds the one of drug design aiming to screen large sets of small molecules in order to identify a small set of potential drugs that becomes experimentally testable. These approaches do not pretend to exactly identify a unique solution but rather a set of reasonable candidates, and reduce, in this manner, the amount of experimental time and costs. This means that we are not focused on the correct docking of experimentally known partners, which can be achieved via other more effective but much more computationally demanding methods [43]. However, one can envisage to use such more sophisticated methods on the small set of candidates that our coarse method identifies to propose more precise models of the potential complexes.
We have realized a large-scale PPI analysis by assuming to know the residues forming the experimental interface of the native complexes (no associated experimental conformation is considered) and by using predictions of binding sites. Experimental binding sites can be seen as perfect predictions, and the analysis based on them is realized for two reasons: 1. to understand how much evolutionary information can contribute to PPI reconstruction when coupled with a coarse-grain docking algorithm using an energy function, and 2. to decouple true PPI signal from noise and identify PPI properties that are not consequences of accumulated errors due to predictive algorithms. This second reason allowed us to be confident, for instance, on the promiscuity observed in Figure 5B (bottom black dots) by ensuring that it is not generated by noise in predictions (see Figure 5A).
A few large-scale analyses, that are similar in spirit, have been performed [32], [33]. A comparison of our results with [33], based on the ten protein complexes discussed in detail in [33], reveals a similar performance of the two methods. However, a full comparison with [33] is impossible since they treat only a subset of the Mintseris dataset, use a large background set and do not provide a detailed measure of the performance of their method. On the contrary, our method is tested on all complexes of the Mintseris dataset, a good testing platform for methods dedicated to protein partner prediction due to its numerous structural differences. The global analysis of the two methods (over the subset of 56 proteins; see Table 1 and [33]) highlights that we can reasonably search for protein partners within sets of a few hundred monomers. We demonstrated that improving current predictive methods is possible through a better prediction of binding sites, and we precisely estimated the effect of such predictions.
We could only partially compare to [32] since they do not perform a CC-D of the Mintseris dataset but only cross the 84 receptors against the 84 ligands, that is a fourth of the interactions explored in our analysis. Performances of our method and the one reported in [32] are comparable on the common subset, but notice that contrary to [32], we use unbound structures and we make no use of the non-naive split of the dataset (that is, receptors versus ligands).
The predictive performance of the method is encouraging for the whole Mintseris Benchmark 2.0 and very satisfactory for the enzyme-inhibitor subset (Table 2). For this latter, the AUC reaches a very high value of while the AUC for the whole Mintseris dataset is . Notice that the way we computed the AUC is very strict, since we asked the true partner to be ranked first over the tested dataset. A more relaxed evaluation is reported in Table 1 where we show that a fourth of the 168 proteins in the Mintseris dataset are recognized by looking at the top 17 predictions over the 168 tested partners. If the binding site of the proteins is correctly predicted, the half of the proteins in the dataset are recognized by looking at the top 8 predictions, and two third by looking at the top 17. This is a very encouraging result with respect to the potential applicability of this in silico predictive approach to the reconstruction of PPI networks. In fact, proposing to a biologist a set of less than 20 interactions to test is very reasonable.
The analysis on the average IR for the enzyme-inhibitor subset highlights that an average IR threshold allows the method to propose about 12 partners, a reasonable number of proteins to be selected for experimental tests. In 38 cases over 46 (Figure 6), the true partner is present in the retained subset showing a very high sensitivity. For the whole Mintseris benchmark, for roughly the half of the dataset (82 proteins), the true partner is retrieved with an average IR . Notice that when considering the experimental binding site of each partner, 138 proteins over 168 display an average IR . This means that a precise binding site prediction method will lead to a successful partners discrimination, a problem that could be considered as being much more ambitious than the binding site prediction problem. Again, these results support the feasibility of the approach to identify potential partners but, most of all, they highlight the interest of testing a protein within a large environment, by randomly choosing many small subsets of proteins in the environment, and by selecting as potential partners to be experimentally tested, those proteins that present a stable average IR (black dots, Figure 5) with the protein under study. The selection of 10 potential partners instead of 17 (as suggested by the direct evaluation of the NII matrix in Figure 1 and Table 1) might be crucial for experimental validation. This observation opens a way to new computational schema for partner predictions.
The analysis highlights an important point on the behavior of all proteins with respect to their partners. For each protein, there is a small set of partners that displays a systematic (black points in the bottom of Figure 5AB) very low average IR that lead to ask whether these partners might physically interact and not be false positives. Three reasonable explanations for this set of highly potential partners can be given: (i) partners can interact on a merely physical base but never meet in the cell due to different cellular compartments localization, (ii) partners can interact for functional purposes, possibly not described until now (several different partners are expected to interact with a protein), (iii) partners can interact in the cell not for functional purposes but generating a competition with the functional partner, possibly participating to the regulation of the protein interactions in the cell. Taking into account these possibilities, this set of highly potential partners becomes interesting for further studies. For instance, these interactions would deserve to be experimentally tested to see how strongly they interact, and whether they form a structurally well-defined complex. Also, for a given protein and a set of highly potential partners, one could ask whether general structural (geometrical or physico-chemical) features of the interface exist and in the positive case, classify these interfaces. These further studies could contribute to give important insights into protein partnership discrimination.
For each protein , we defined a signature representing the strength of interaction of with all other proteins. As mentioned above, signatures found for all 's in the Mintseris dataset demonstrate the existence of strong interactions with some proteins, but also the absence of interactions with other proteins, and so on. The spectrum of strengths of interactions suggests the notion of PPI to be revisited so to include the larger panel of potential complex formations between a protein and its potential partners. Several questions could then be asked on proteins presenting similar signatures [44], but they go beyond the aim of this work.
We have shown that evolutionary information can also be used to restrict the conformational space of the docking exploration without an important loss in sensitivity. This result is very important in view of reducing the computational cost of highly time demanding docking calculations (all atom description and precise energy functions) and the perspective of enlarging the dataset size for future CC-D calculations.
To conclude, we are the first to perform a CC-D of a pool of proteins covering a large spectrum of functions and interaction modes, performing it on unbound structures and providing energy values (even though simplified) taking into account electrostatic forces. Our approach is the first combining evolutionary information with CC-D simulations. The evaluation of the performance of these two contributions to the problem of partner identification, suggests that there is still room for improvement in the solution. In particular, we have shown that a precise identification of protein binding sites allows for very satisfactory predictions. Data coming from the CC-D calculations and the evolutionary analysis are provided and they will help the community to evaluate further CC-D studies and methodological developments. In particular, the decoy set constitutes a unique dataset of “negative” partners. For them, we provide about 2000 conformations and an associated coarse-grain energy score. It might be extremely useful to suitably parametrize docking scoring functions, more refined than our coarse-grain scoring function, to discriminate partners. In the context of this study, a subset of these decoy structures filtered by our coarse-grain scoring function could be re-scored for a better partnership evaluation by using a more refined score function better discriminating the interaction signals.
The Docking Benchmark 2.0 [25] is constituted by 168 proteins belonging to 84 known complexes. We used the unbound conformations of the proteins with the exception of 12 antibodies for which the unbound structure is unavailable. For those, the bound structure is used instead. Any reference to the proteins uses either their name or the Protein Data Bank (PDB) code [45] of the experimental complex they belong to with the or extension denoting a receptor or a ligand protein respectively. For example, 1AY7_r and 1AY7_l refer to barnase (receptor) and barstar (ligand) in the barnase-barstar complex 1AY7. The coordinates for the bound and unbound structures of both receptor and ligand proteins are available in the PDB and can be found at http://zlab.bu.edu/zdock/benchmark.shtml.
Molecular docking is performed with the MAXDo (Molecular Association via Cross Docking) algorithm, developed for complete cross-docking (CC-D) studies [12]. Since CC-D involves a much larger number of calculations than simple docking, we chose a rigid-body docking approach using a reduced protein model in order to make rapid conformational searches.
Surface residues are residues with at least of accessible surface. Accessibility is calculated with NACCESS 2.1.1 [53] with a probe size of = 1.4 Å. Interface residues are residues with a change of at least decrease in accessible surface area compared to the unbound protein.
In order to improve the quality of the predictions of protein interaction partners, in our earlier study we developed a normalized interaction index (NII) that takes into account whether a protein-protein interface involves amino acids belonging to a known interaction site [12]. This information can potentially be obtained using predictive tools (see below), but here we use the experimentally determined interfaces of the 84 binary complexes in the Docking Benchmark. We however recall that all our docking trials involve unbound protein conformations. For each protein partner in a given complex , we determine which fraction of the docked interface residues (abbreviated as ) are found in the experimental interface for () and (). Thus defining an overall fraction for the complex as . It is important to notice that the formula can be computed from either experimental interfaces (as defined above) or predicted interfaces (where prediction could be realized, for instance, with evolutionary information; see paragraph below). The notion of “FIR” proposes a new concept for docking evaluation that can be used as an alternative to the usual docking metrics [54] originally designed to evaluate the accuracy of pairwise protein docking models. While the measure denotes the coverage of the experimental interface, that is the sensitivity of the predicted interface, the FIR denotes the PPV of the predicted interface. Also, for the measure, contacts are defined with respect to a 5 Å cutoff on the RMSD of heavy atoms, while for FIR, contacts are defined from a change of solvent accessibility.
For every protein pair , we calculate an energy-weighted optimal interaction index () defined in Eq. (1).
To allow comparison among different partners we defined a normalized index by taking into account all of the four lines/columns that feature either or in the matrix as follows:(3)where is a symmetrized version of the interaction index and it is defined as:(4)where are the 168 proteins of our dataset. values vary between and . Values close to zero imply that two proteins cannot form an interface involving a significant fraction of the experimentally identified residues, or that interfaces involving these residues have poor interaction energies. Values close to one indicate predicted interfaces with good energies and composed of experimentally identified residues.
For each protein , we define as predicted partner of , the protein that leads to .
We consider as true positives () all the predicted pairs that belong to the Docking Benchmark 2.0 and as true negatives () all the pairs that are correctly predicted as non interacting. We define a False Positive Rate () and the True Positive Rate () to be and , where is the set of False Positives (partners incorrectly predicted as interacting) and is the set of False Negatives (partners incorrectly predicted as non interacting). The computation of and for various thresholds enables the Receiver Operating Characteristics (ROC) curve to be drawn. The performance of the prediction is given by the resulting AUC (Area Under the Curve) value. Values of and correspond to random and perfect predictions respectively. AUC calculations were performed with the R package [55]. Also, given a threshold on the NII values, we use five standard measures of performance: sensitivity , specificity , precision or positive predictive value , balanced -score and Matthews correlation coefficient where .
To predict protein partners without using any experimental information, we define a new FIR measure by combining docking and evolutionary information. From FIR values, NII matrices are computed as above.
The interaction rank of a protein pair is defined to be the best rank of the pair among all the pairs that have either or as receptor. This means that given a matrix, we look at the rank of the pair with respect to the values , that is the line indexed by , and at the rank of the pair with respect to the values , that is the line indexed by . The best rank computed for each line is retained for the pair .
To restraint the conformational space of the docking algorithm, we combine JET interface predictions with MAXDo, in such a way that only surface regions containing residues predicted by JET will be analyzed by MAXDo. To do so, for each docked orientation, we computed the center of mass of the ligand and defined the axis linking it to the center of mass of the receptor. We remind that the position of the receptor is fixed. Along this axis, we define an imaginary tube of radius r = 2.9 Å. For each ligand orientation, we check whether the interface of the resulting ligand-receptor complex involves residues predicted by JET or not (Figure S53 in Text S1). Each residue is approximated to a point whose coordinates represent the average of the atom's coordinates. The distance of this point from the axis of the tube, allows to establish whether the residue falls inside the tube or not, and therefore, whether the ligand orientation should be retained or not. Strictly speaking, one should also use the scalar product between the vectors going from the receptor center of mass to the residue and to the ligand center of mass (this product decides whether the residue lies on the side of the ligand-receptor interface). We ask for just one single residue in the orientation interface to be within the tube to retain this latter.
CC-D of the Mintseris' Enzyme-Inhibitors dataset was performed with HEX v6.3 using the shape complementarity based-only score [42]. Docked conformations were clustered using a 3 Å cutoff and the best-scored conformations of the 500 first clusters were retained for the analysis. A protocol similar to that described for MAXDo was applied to evaluate partner prediction based on HEX results, (i) by assuming knowledge of the experimental interfaces and (ii) by crossing docking scores with evolutionary information. All 500 conformations were considered for residue scoring based on docking and for protein interaction index calculation. Parameter values are reported in Table S7 in Text S1.
Given a protein , we searched in the Mintseris dataset for those proteins that have a homolog coming from the same species as . Namely, for each , we searched with Blast (E-value threshold at , alignment coverage ) for the set of sequences that are at least , or or identical to the original sequence. This provides a set of species that we say to be representing . We then checked that the species of is included in the set of the species representing . Notice that the protocol does not necessarily provide the same answer when it is applied to the protein pairs or due to the non-symmetrical Blast result.
We release the first large decoy database comprising not only structures of true complexes but also structures of non-functional complexes potentially forming in the cell. For the 28224 possible protein pairs (involving the 168 proteins) of the Mintseris Benchmark 2.0, we considered about 2000 best ligand orientations (represented on and angles as described above) for each receptor. We provide the associated decoys together with the corresponding energy values. A program to reconstruct the PDB structure of the conformation given and angles is also provided. For each protein in the Mintseris dataset, we also furnish the evolutionary analysis for the detection of the binding sites. The download site is http://www.lgm.upmc.fr/CCDMintseris/
|
10.1371/journal.ppat.1003103 | p12 Tethers the Murine Leukemia Virus Pre-integration Complex to Mitotic Chromosomes | The p12 protein of the murine leukemia virus (MLV) is a constituent of the pre-integration complex (PIC) but its function in this complex remains unknown. We developed an imaging system to monitor MLV PIC trafficking in live cells. This allowed the visualization of PIC docking to mitotic chromosomes and its release upon exit from mitosis. Docking occurred concomitantly with nuclear envelope breakdown and was impaired for PICs of viruses with lethal p12 mutations. Insertion of a heterologous chromatin binding module into p12 of one of these mutants restored PICs attachment to the chromosomes and partially rescued virus replication. Capsid dissociated from wild type PICs in mitotic cells but remained associated with PICs harboring tethering-negative p12 mutants. Altogether, these results explain, in part, MLV restriction to dividing cells and reveal a role for p12 as a factor that tethers MLV PIC to mitotic chromosomes.
| Retroviruses, including the murine leukemia virus (MLV), reverse transcribe their RNA genome to a DNA copy, which travels from the cytoplasm to the nucleus as part of a ‘pre-integration complex’ (PIC), to integrate into cellular chromosomes. The viral p12 protein is a constituent of the MLV PIC, but its function in this complex has remained unknown. We developed a real-time imaging system to detect p12 and MLV PICs in live cells. This revealed that p12 tethers the MLV PIC to mitotic chromosomes. Accordingly, PICs derived from viruses with specific lethal mutations in p12 failed to attach to the chromosomes, and insertion of a heterologous chromatin binding module into p12 restored PICs attachment to the chromosomes and rescued virus replication. In addition, docking of wild type PICs to chromosomes coincided with nuclear envelope breakdown during mitosis, and detachment occurred upon exit from mitosis. Capsid - another viral component of the PIC - dissociated from wild type PICs in mitotic cells but remained associated with PICs harboring tethering-negative p12 mutants, suggesting interplay between these two proteins in regulating targeting of mitotic chromosomes by the PIC. These results highlight steps contributing to the high tropism of MLV to dividing cells.
| To integrate, reverse transcribed retroviral genomes are imported from the cytoplasm to the chromosomes as part of a pre-integration complex (PIC). Differences in retrovirus PIC trafficking influence their ability to infect resting and/or dividing cells. Lentiviruses, including the human immunodeficiency virus (HIV), infect both dividing and resting cells. In contrast, simple oncoretroviruses, such as murine leukemia viruses (MLV), are restricted to dividing cells [1], [2], [3], [4]. The HIV PIC is capable of entering the nucleus through the nuclear pore complexes, allowing integration in chromosomes of resting cells. The MLV PIC is thought to get access to the chromosomes only during mitosis, upon nuclear envelope (NE) disassembly, as inferred from correlating kinetics of cell division and integration [3].
This dissimilarity between HIV and MLV PIC trafficking likely stems from their different composition [4]. Capsid (CA), present in MLV PICs [5], [6], [7] and absent from HIV PICs [8], [9], contributes to the difference in ability to transduce nondividing cells [10]. Also, the lens epithelium-derived growth factor (LEDGF/p75), interacts with HIV PICs [11], [12] tethers the integrase (IN) of HIV and other lentiviruses to chromatin [11], [13], [14], [15], [16], but not the MLV IN [11], [17], for which no equivalent tethering factor has been identified.
p12, a cleavage product of MLV Gag precursor, is thought to influence MLV integration. p12 acts in the budding of assembled Gags, and mutations in this domain hamper particle morphogenesis and release [18], [19], [20]. Viruses with other lethal mutations in p12 showed early infection defects, with normal generation of linear genomic DNA, but no circular DNA forms [18], [19], [21], [22]. The latter forms, thought to be generated by nuclear enzymes, are not substrates for IN-mediated integration but mark nuclear entry of the viral DNA. Their absence in cells infected with some p12 mutants, and the normal in vitro integration activity shown by the PIC of one of these mutants [18], suggest that p12 functions in an unknown way after reverse transcription and before integration.
p12 is a component of the MLV PIC, crucial for the progression of the PIC towards integration: p12, observed as discrete puncta, associates with CA and the viral genomic DNA and this p12- genome association occurs in the cytoplasm and adjoining chromosomes, suggesting that p12 escorts the viral genomic DNA throughout early stages of infection [7]. p12-containing PICs accumulate on mitotic chromosomes, however this accumulation is impaired for p12 proteins with a mutation that rendered the virus integration-defective [7]. These data implied that p12 functions in directing the PIC to integration, yet its precise role has remained unknown. Here we imaged MLV-PICs in live cells and revealed that p12 tethers the MLV PIC to mitotic chromosomes.
To investigate the role of p12 in PIC trafficking we labeled the PIC with p12 fused to enhanced green fluorescent protein (GFP). An in-frame insertion of GFP to the central region of p12 was lethal to the virus (data not shown). Thus, we generated chimeric virions, composed of both wt and modified Gag molecules; the latter containing GFP fused to p12 N-terminus. In the modified Gag (MA-GFP/p12-CA-NC), the GFP sequence was inserted in-frame, downstream of the protease-cleavable matrix (MA)-p12 junction, and upstream of a short, non-cleavable linker fused to p12 (Fig. 1A). MA-GFP/p12-CA-NC retains wt p12-CA and CA-nucleocapsid (NC) cleavage sites, critical for MLV particle formation, maturation and infectivity [23], [24]. Wt virus was co-expressed with MA-GFP/p12-CA-NC, and virions were purified by ultracentrifugation through a 25% sucrose cushion and immunoblotted with anti-GFP antibody. This revealed a major band corresponding to the GFP-p12 fusion (∼36 kDa), a fainter band of MA-GFP/p12-CA-NC precursor (∼88 kDa) and traces of additional cleavage products (Fig. 1B); suggesting that MA-GFP/p12-CA-NC co-assembled with wt Gag and was processed by the viral protease. Importantly, the linker connecting GFP and p12 was protease-resistant as no free GFP was processed from MA-GFP/p12-CA-NC, unlike the control construct (named MA-GFP-p12-CA-NC, Fig. 1B), in which the GFP was flanked by cleavable sites (Fig. 1A).
To test if co-expression of MA-GFP/p12-CA-NC with wt virus affects infectivity, a MLV vector (pQCXIP-gfp-C1), encoding for the puromycin-resistance gene and GFP was expressed together with different ratios of MA-GFP/p12-CA-NC to wt virus. Virions, normalized by reverse transcriptase (RT) activity, were used to infect NIH3T3 cells and GFP+ cells were counted by fluorescence-activate cell sorting (FACS). A 1∶1 molar ratio resulted in only a minor reduction (∼15%) in infectivity compared to the wt virus (with no MA-GFP/p12-CA-NC) control (Fig. 1C). Counting the number of puromycin-resistant colonies in infected cultures gave similar results (data not shown). Thus, 1∶1 molar ratio was used in further experiments.
At 12 hr post-infection (hpi) of U/R cells (human osteosarcoma U2OS cells, expressing the murine receptor for MLV; [7]) with labeled chimeric virions (hereinafter named wt GFP), discrete fluorescent puncta were detected in the cytoplasm of interphase cells, and adjacent to the condensed chromosomes of mitotic cells (Fig. 1D). This appearance was identical to former images of PICs, labeled with Myc-tagged p12 proteins (derived from the replication-competent 1xMycR clone) [7]; in addition, 65±3% of the fluorescent puncta (approximately 100 dots/cell in 5 cells were analyzed) overlapped the chromosomes in mitotic cells, in good agreement with the 70% overlap between Myc-tagged p12 proteins and mitotic chromosomes, as was quantified before [7]. Immunostaining of wt GFP-infected U/R cells with antibodies against CA (a component of the MLV PIC) revealed extensive co-localization between GFP and CA signals (Fig. S1A); and quantification of the overlap between these two signals revealed a 71±4% overlap in interphase cells (data was obtained from five cells, each containing ∼80 fluorescent dots). This number is similar to the extent of overlap measured for Myc-tagged p12 and CA in interphase U/R cells, infected with 1xMycR virus (∼80%; see below). This suggests that GFP-p12 molecules are associated with the PICs to a similar extent as p12 molecules lacking the GFP moiety. Thus, the GFP-p12 labeling system successfully marks the incoming PICs.
To monitor cytoplasmic-nuclear trafficking of PICs in live cells, we labeled the NE of U/R cells by stably expressing lamin A fused to red fluorescent protein (RFP-lamin A; U/R/RFP-laminA cells) [25]. U/R/RFP-laminA cells, arrested before S phase by serum starvation and aphidicolin treatment [3], were infected with wt GFP virions. At 18 hpi, GFP-labeled PICs exhibited undirected and directed movements, including towards the NE, in sharp contrast to the immobility of virions attached to the cover-slip (Movie S1). None of the PICs crossed the intact NE (Movie S1 and see below), demonstrating the physical barrier that the intact NE imposes on nuclear entry of the PICs in interphase cells - a notion previously deduced from measuring integration kinetics, in respect to the cell cycle [3], but never directly shown.
To monitor the PICs in mitotic cells, U/R/RFP-laminA cells were arrested at metaphase with 2-methoxyestradiol (2ME2) and infected with wt GFP virions. 2ME2 impairs microtubule dynamics without gross microtubule depolymerization and arrests cells at the spindle assembly checkpoint [26], [27], [28]. Arrested cells displayed diffuse RFP-lamin A (indicating NE disassembly [29]) and restricted PIC motility (Fig. 2A; Movie S2, part A). In 2ME2-treated cells that did not reach metaphase (with intact NEs), PICs were restricted to the cytoplasm and motile (Fig. 2B; Movie S2, part B), excluding a direct 2ME2-induced inhibition of PIC motility.
To test for the docking of immobile PICs to mitotic chromosomes, we generated U/R/RFP-H2A cells [U/R cells stably expressing RFP fused to histone H2A (RFP-H2A)], infected them with wt GFP virions and imaged unsynchronized, interphase and mitotic cells. Indeed, motionless PICs were attached to the mitotic chromosomes, while cytoplasmic PICs in interphase cells were motile (Fig. 2C, D; Movie S2, parts C, D). Attachment of PICs to mitotic chromosomes was also observed for wt GFP-infected, unsynchronized mouse NIH3T3 cells expressing RFP-H2A (NIH3T3/RFP-H2A; Movie S2, part E); and for wt GFP-infected U/R/RFP-H2A cells that were arrested at M phase by nocodazole treatment (Movie S2, part F). In the latter settings, PICs attachment to mitotic chromosomes could also be observed in cells arrested at mitosis for up to 40 hpi (data not shown; extended time points were not tested because of apparent drug-induced cytotoxicity).
The apparent docking of the PIC to the chromosomes may be the result of a stable association of the PIC components with the viral DNA genome that had integrated into the chromosomes. To test this, we made chimeric virions (named D184A GFP), using MLV with the D184A mutation in the catalytic site of IN, which disrupts its activity [30]. D184A GFP PICs were motile in interphase U/R/RFP-H2A cells, and docked to mitotic chromosomes in dividing cells (Fig. 2E; Movie S2, part G). Thus, IN activity is not required for the docking of MLV PICs to the chromosomes.
To monitor the transition between the cytoplasmic movements of the PICs to their docking to mitotic chromosomes, we viewed infected cells as they entered mitosis. Confluent U/R/RFP-H2A and U/R/RFP-laminA cells were infected with wt GFP particles and 2 hr later, the cells were trypsinized and replated at a lower density. After additional 5 hr we detected infected cells that entered mitosis as judged by the growing condensation of their chromosomes (U/R/RFP-H2A), or by the progressive dissolution of the NE (U/R/RFP-laminA). Docking of the GFP-labeled PICs to the RFP-labeled chromosomes could readily be detected; the time frame between the first observed docking event and the docking of the rest of the PICs was ∼2 min (Fig. 3A; Movie S3, part A). Similarly, the immobilization of PICs occurred concomitantly with the breakdown of the RFP-labeled NE (Fig. 3B, C; Movie S3, part B), in contrast to the cytoplasmic PICs that remained mobile in the same cells (Fig. 3C; Movie S3, part B). Thus, these images are consistent with the idea that the intact NE might act as a physical barrier to the PICs.
To probe if PICs bind exclusively to mitotic chromosomes, provided that the physical barrier of the NE is avoided, we infected 2ME2-arrested U/R/RFP-H2A cells with wt GFP virions and imaged chromosome-docked PICs at 12 hpi. Addition of Reversine, which inhibits the Mps1 kinase, counters the spindle assembly checkpoint [31], and reverses the 2ME2-induced blockage of the cell cycle [32], resulted in the decondensation of the mitotic chromosomes within 1 hr (Fig. 4A–C; Movie S4, parts A, B). Remarkably, in these conditions, PICs regained their movement, which was now restricted to the nucleus (compare Fig. 4B to C, and part A to B in Movie S4). The same was also observed for unsynchronized U/R/RFP-H2A or U/R/RFP-laminA cells that naturally exit mitosis (without 2ME2/Reversine treatment; Fig. 4D–G; Movie S4, parts C–F).
The regaining of PICs movement during the exit from mitosis may reflect their release from the chromatin upon completion of the integration step. However, the same dissociation occurred also with D184A GFP PICs that are unable to discharge their viral genomic DNA due to the lack of IN activity (compare the same D184A GFP-infected U/R/RFP-H2A cells in Movie S2 part G and Movie S4 part G). This result provides further support for the affinity of the MLV PIC towards mitotic, and not interphase, chromosomes. Of note, a minor fraction of PICs (integration-competent wt GFP or integration-incompetent D184A GFP) remained associated with the chromatin following exit from mitosis, (Movie S4; and see an example in Fig. 4E, marked with an arrowhead); demonstrating that integration is not a pre-requisite for such a stable association.
The above imaging employed p12 as a marker of PICs, but falls short of attributing a function to p12. Defined mutations in p12 [such as a five-amino acids alanine block substitution, named PM14; and the S61A and S(61, 65)A mutations] are lethal to the virus and dramatically reduce the levels of circular forms of the viral genomic DNA, suggesting a defect in nuclear entry of this genome [18], [19], [22]. Myc-labeled PM14 PICs are normally distributed in the cytoplasm of interphase cells, but fail to accumulate on mitotic chromosomes [7]. We next generated GFP-labeled mutant chimeric virions (PM14 GFP; composed of PM14 virus and modified Gags with PM14 mutation) and infected and imaged (at 12 hpi) unsynchronized U/R/RFP-H2A cells. In the cytoplasm of interphase cells, PM14 GFP PICs moved similarly to wt GFP PICs (Movie S5, part A). A portion of PM14 PICs reached the chromosomes in mitotic cells; however, despite their proximity to chromosomes, all were motile and none attached to the mitotic chromosomes (Fig. 5A; Movie S5, part B). The failure in docking was also observed in unsynchronized, mitotic mouse NIH3T3/RFP-H2A cells (Movie S5, part C). This strongly implies a role for p12 in the docking of MLV PICs to the chromosomes. To monitor wt and PM14 PICs in the same cell we generated chimeric virions (named wt mCherry), labeled with mCherry instead of GFP (using the MA-mCherry/p12-CA-NC construct; Fig. 1A). In Hoechst-stained 2ME2-arrested U/R cells, PM14 GFP PICs were motile, in contrast to the immobilization of the wt mCherry PICs on condensed chromosomes (Fig. 5B; Movie S5, part D). To quantify the spatial retention of PICs over time we calculated the percentage of overlap between the PICs in the first frame, to the PICs in the second, third and fourth frames of each movie (Fig. 5C). While high retention (∼70 to 80%) was observed for wt GFP and D184A GFP PICs over time; PM14 GFP PICs showed lower and decreasing retention overtime (∼30 to 20%). This low retention was comparable to that of motile wt GFP and D184A GFP PICs upon Reversine-mediated exit from mitosis (Fig. 5C). For PM14 GFP, we also quantified the overlap between the signals of GFP and mitotic chromosomes in fixed cells, and found it to be approximately 16±2% (approximately 100 dots/cell in 8 cells were analyzed). This value is in a good agreement with both the relative low retention exhibited by this mutant in the above real-time analysis, and the low overlap (11%) measured before for Myc-tagged p12 proteins and mitotic chromosomes [7]. This relatively low overlap is in contrast to the 63% overlap, measured for wt GFP and mitotic chromosomes in fixed cells (see above). Reconstitution of serial optical sections into 3D images of mitotic chromosomes of U/R/RFP-H2A cells further showed the close contacts between such chromosomes and wt GFP (Fig. S2A), but not PM14 GFP (Fig. S2B), PICs.
Of note, chimeric PICs resulting from the expression of wt MLV genome and modified Gag containing the PM14 mutation docked to mitotic chromosomes (not shown), demonstrating that the docking activity of wt p12 is dominant over the lack of such activity of PM14 GFP-p12, and providing genetic evidence for the ability of GFP-p12 to mark the incoming MLV PIC.
Viruses carrying the S(61,65)A mutations are phenotypically undistinguishable from the PM14 virus. However, a replication-competent revertant with an additional compensatory mutation (M63I) in p12 exists for this virus [22]. We introduced the S(61,65)A or the S(61,65)A/M63I to the 1xMycR clone and to the modified Gag (Fig. 1A) and co-expressed each cognate pair of constructs to generate chimeric virions [named S(61,65)A GFP or S(61,65)A/M63I GFP, respectively]. S(61,65)A GFP PICs failed to stably anchor to mitotic chromosomes or showed a very short, unstable association with the chromosomes (Fig. 5D; Movie S5, part E). The S(61,65)A/M63I PICs, in contrast, stably docked to mitotic chromosomes, identically to wt PICs (Fig. 5E; Movie S5, part F). A 25 and 65% overlap with mitotic chromosomes was measured for S(61,65)A GFP and S(61,65)A/M63I GFP, respectively; in good accord with the 10 and 65% overlap of PM14 and wt viruses, respectively [7]. Altogether, these results further demonstrate the correlation between the replication competence of the tested virus and the ability of its PIC to dock to the chromosomes; and provide additional demonstration for the connection between the impairment of such docking and the presence of specific mutations in p12.
To evaluate if addition of a foreign chromatin-binding element rescues the docking of mutant p12 PICs, we inserted such a module of the herpes LANA protein (LANA31), into p12 of the PM14 clone, generating the PM14/LANA31 virus. LANA31 consists of 31 residues, 23 of which bind the groove between histones 2A and 2B [33]. This module restores the tethering activity of a mutated LEDGF/p75, resulting in the binding of HIV IN to chromatin [34]. A control virus (named wt/LANA31) was also made by inserting LANA31 into p12 of wt MLV. LANA31 was inserted between the DRD and GNG residues of p12 (Fig. 1A), as MLV replication tolerates insertion of a Myc tag into this location [7]. These viruses were expressed with the MA-GFP/p12-CA-NC modified Gag, resulting in chimeric viruses (named wt/LANA31 GFP and PM14/LANA31 GFP). PICs of PM14/LANA31 GFP stably anchored to mitotic chromosomes in U/R/RFP-H2A cells (Fig. 6B; Movie S6, part A) similarly to wt/LANA31 GFP (Fig. 6A; Movie S6, part B) and wt GFP (Fig. 2C; Movie S2, part C) PICs. Quantification of the overlap between the signals of GFP and mitotic chromosomes in fixed cells, revealed a 73±5% value for PM14/LANA31 GFP (approximately 70 dots/cell in 7 cells were analyzed), which was similar to the overlap found for wt GFP (63%), and which was in contrast to the low overlap (16%) measured for PM14 GFP (see above). In addition, quantification of the spatial retention of PM14/LANA31 GFP PICs over time in mitotic cells showed high retention levels that were almost identical to the retention of wt GFP PICs in mitotic, 2ME2-treated cells (compare Fig. 6C to 5C). Thus, insertion of the LANA31 peptide into p12 rescues chromatin docking of the PICs with PM14 mutations. To evaluate the effect of LANA31 insertion into p12 on virus infectivity we infected NIH3T3 cell cultures and monitored the kinetics of virus spread. Whereas wt virus spread quickly, wt/LANA31 showed much slower spreading (Fig. 6D), indicating that the insertion of LANA31 greatly attenuated virus replication. This is in line with our previous observation, showing that insertion of a peptide with a similar size (30 residues of a triple Myc epitope) in the same location in p12 attenuates virus replication [7]. Furthermore, sequence analysis of p12 of the viruses that spread in the wt/LANA31-infected cells revealed that LANA31 was rapidly deleted from wt/LANA31 virus (data not shown), substantiating the deleterious effect of this sequence on the virus. PM14 virus showed no spreading (here and [19]); the PM14/LANA31 virus showed detectable, slow spreading (Fig. 6D), indicating that insertion of LANA31 into p12 partially restored the infectivity of PM14 mutant. Sequence analysis of p12 of viruses that spread in the PM14/LANA31-infected cells showed a mixture of PM14/LANA31 sequence together with wt sequences (no PM14 and no LANA31; data not shown). This likely reflects the selection that the PM14 mutation enforces on the retention of the LANA31 sequence in p12 and the parallel recombination of the PM14/LANA31 slow virus with endogenous retroviruses of the mouse genome [35] that harbor wt p12 sequences. Such recombination, which likely involves the co-packaging of the endogenous and exogenous viral genomes, and the recombination between these genomes during reverse transcription [36], cannot efficiently occur with PM14 virus that lacks detectable replication, but may occur upon the multiple cycles of infection of the PM14/LANA31 virus.
Further support for the rescue of PM14 infectivity by LANA31 came from single-cycle infection assays, in which NIH3T3 cells were infected with wt, wt/LANA31, PM14 or PM14/LANA31 virus-like particles (VLPs), harboring puromycin or neomycin -resistance markers, and normalized by RT activity (Fig. 6E). Quantification of the number of drug-resistant cell colonies from three independent experiments revealed that LANA31 insertion into wt p12 reduced particles' infectivity to 85±11% of that of wt particles. A major increase in VLPs infectivity, however, was observed when LANA31 was inserted into PM14 p12: whereas PM14 VLPs had only residual infectivity (0.3±0.4%, compared to wt), PM14/LANA31 VLPs showed higher infectivity (16±8%, compared to wt). These results are in accordance with the spreading assays described above. Altogether, insertion of LANA31 into p12 compensates for PM14 effect on both virus replication and anchorage of the PIC to the chromosomes, further emphasizing the role of p12 as the tethering factor for the MLV PIC.
We also tested how the exit from mitosis affects LANA31-mediated docking of the PICs to the chromosomes. In unsynchronized U/R cells co-infected with wt/LANA31 GFP and wt mCherry, and stained with Hoechst; wt mCherry PICs regained their movement upon exit from mitosis, while wt/LANA31 GFP PICs remained motionless (Fig. 6F; Movie S6, part C). This result provides further support for the affinity of the wt PICs towards mitotic but not interphase chromatin.
Altogether, the above results demonstrate a role for p12 as a factor that tethers the MLV PIC to mitotic chromosomes.
CA and Myc-tagged p12 co-localize in the cytoplasm of interphase cells [7]. To further study the spatial relations between CA and p12, we compared interphase with mitotic 1xMycR-infected U/R cells, using immunofluorescence (Fig. 7). In interphase cells, the majority of the PICs were cytoplasmic with a clear overlap (∼80%) between CA and p12 signals, similar to the co-localization observed in particles attached to the glass outside the cells (Fig. 7A and see identical results in [7]). In contrast, in mitotic cells, p12 associated with the condensed mitotic chromosomes and almost no CA could be co-detected in these chromosome-localized p12 spots (∼30% overlap between p12 and CA, Fig. 7B, F; ∼70 and ∼3% overlap with the mitotic chromosomes for p12 and CA, respectively, Fig. 7G). Similar results were obtained when the overlap between CA and GFP-labeled PICs was measured (32±5%, calculated from three inspected cells, each containing ∼80 dots; Fig. S1B). These results, together with the fact that p12 co-localizes with the genomic viral DNA both in cytoplasmic PICs and in PICs attached to mitotic chromosomes [7], imply that during early stages of infection gradual uncoating events occur, involving the sequential dissociation of MA and CA, and the trafficking of p12 proteins as part of the PIC, to the chromosomes in mitotic cells. Importantly, in tethering-negative PM14 and S(61,65)A mutant PICs, a continued CA-p12 association was observed in mitotic cells (Fig. 7 C, D, F), while the S(61,65)A/M63I revertant showed wt levels of dissociation (Fig. 7E, F). This shows that lack of CA dissociation correlates with inability of p12 mutants to tether to mitotic chromosomes.
To follow MLV PICs by live cell microscopy, we generated chimeric particles co-assembled from wt Gag molecules and Gag harboring a viral-protease resistant GFP-p12 fusion. This labeling method was applied as p12 escorts the viral genomic DNA as part of the PIC to the chromosomes [7]; and since the chimeric particles showed only a slight reduction in infectivity. The notion that GFP-p12 labeled the PICs is supported by: (i) the similarity between punctuate distributions of GFP-p12 and Myc-p12 [7] in infected cells; (ii) the cytoplasmic localization of GFP-p12 puncta in interphase cells and their accumulation on mitotic chromosomes, identical to the patterns observed with fluorescence in situ hybridization (FISH)-labeled MLV genomes [3], [7]; (iii) the lack of chromosomal localization of MA-GFP/p12-CA-NC Gag proteins when exclusively assembled as VLPs (without wt Gag) and pseudotyped with the MLV ecotropic envelope (data not shown); (iv) the accumulation on mitotic chromosomes of GFP-p12 harboring the PM14 mutation, only when employed in the context of chimeric PICs containing wt p12 proteins; (v) the extent of overlap between CA and GFP-p12 signals, which was very similar to the overlap of CA and Myc-tagged p12 proteins (derived from the 1xMycR replication-competent virus), in both interphase and mitotic cells. Thus, our live-cell imaging system allows monitoring of early stages of MLV infection, and complements the envelope-based system, developed to detect MLV-cell interactions before entry [37]; and the systems developed to detect HIV PICs [38], [39].
Our main finding is the inability to dock to mitotic chromosomes of PICs derived from PM14 and S(61,65)A p12 mutants, which sharply contrasted to the stable docking of PICs from wt and the S(61,65)A/M63I revertant viruses. This was best visualized upon the co-infection of PM14 and wt viruses, which showed lack of docking, or stable chromosomal association, respectively, in the same cell. Moreover, insertion of a heterologous chromatin-binding module (LANA31) into p12 of the PM14 virus rescued both its infectivity and the anchorage of its PIC to the chromosomes, strongly implying a chromosome-tethering function for p12. The discrepancy between the full restoration of docking to the chromosomes of PM14/LANA31 PICs and the partial restoration of infectivity of this virus likely results from the adverse effect caused by the addition of 31 residues into p12 on other stages of the replication cycle; for example, we observed a reduction in assembly of the wt/LANA31virus (data not shown). A similar-sized insertion (a triple Myc epitope) in the same location also attenuated replication [7]. Of note, while the PM14 mutation in p12 resulted in complete disruption of PIC docking to the mitotic chromosomes, S(61,65)A PICs showed short, unstable associations with the chromosomes; possibly explaining why revertants could be isolated for the latter mutant [22]. Importantly, although the function of p12 as a chromatin tether can be deduced from the clear differences in the ability of the PICs described above to attach to the chromosomes, our microscopic analysis cannot distinguish between functional and nonfunctional PICs in a given cell.
LEDGF/p75 was identified as the factor that tethers the HIV IN to chromatin: it interacts with unknown chromatin ligands and with IN, and is essential for the chromosomal targeting of HIV IN [11], [14], [40]. LEDGF/p75 depletion hampers HIV integration and blocks infection of HIV and other lentiviruses; however, LEDGF/p75 does not interact with MLV IN and accordingly, its depletion does not affect MLV infection [11], [17], [41], [42], [43]. We suggest that different lentiviruses, including HIV, evolved to tether their PICs to the chromosomes through IN-LEDGF/p75 interactions; in contrast, MLV evolved to tether its PIC via a LEDGF/p75-independent way, which involves the use of p12. The notion of different tethering mechanisms of HIV and MLV is further supported by their differences in integration site selection [42], [43], [44], [45], [46], by similarity in the target site selection of a HIV chimeric virus expressing the MLV IN with MLV, and by the increased similarity upon replacement of the HIV Gag of this chimera with MLV Gag [47], [48]. Our data point to p12 as a Gag protein essential for MLV integration; but fall short of dissecting a putative role for p12 in target selection. A likely scenario is that p12 tethering activity is an essential prerequisite for integration, while progression to complete integration depends on additional interactions of IN with cellular factors [49].
What kind of interactions may be involved in the p12-mediated tethering activity? The specific recognition of mitotic chromosomes by p12-labled PICs, and their detachment upon exit from mitosis suggest that the MLV PIC recognizes chromatin features, such as post-translational modifications that are associated with cell cycle progression. For example, phosphorylation of histone H3 at Thr3 (H3T3) and methylation of the adjacent Lys4 (H3K4), serve as a motif for the binding of cellular factors, such as TFIID, to mitotic chromosomes [50]. Intriguingly, strong association between the target site selection of MLV and specific chromatin modifications, including H3K4 methylation, has been described [48]. However, no chromatin binding modules that recognize these modifications [51] have been identified in p12. Notably, p12 displays similarity to histone H5 protein [52] and may directly interact with chromatin. Alternatively, p12 may recruit a cellular factor harboring chromatin recognition domains.
MLV integration peaks in cells after exit from metaphase and decondensation of chromosomes [3]. The release of the GFP-p12 complexes from the decondensed chromosomes in cells that exit mitosis may represent the release of PICs after integration. However, the released GFP-p12 puncta may represent PICs that failed to integrate. The release of D184A PICs upon exit from mitosis exemplifies the independence of this release from IN activity. Moreover, the lack of release of LANA31-tethered PICs further underscores the specific affinity of wt MLV PICs for mitotic chromosomes. Of note, a minority of the wt PICs remained immobilized at the end of mitosis. Although it can be argued that only such PICs mediate active integration, this is unlikely as the same phenomenon was observed for D184A PICs.
Using immunofluorescence analysis, we demonstrated here that CA, a known component of the PIC [5], [6], co-localized with cytoplasmic p12 in interphase cells, but not with chromosome-docked p12 in mitosis. This result concords with the higher ratio between CA and viral genomic DNA, or CA and IN, in MLV PICs extracted from cytoplasmic fractions, compared to nuclear PICs -suggesting that CA is lost from the PIC after its entry into the nucleus [6]. Our results extend this finding and show that CA-PIC/p12 dissociations occur specifically upon mitosis, when no intact NE exists in the cell. Furthermore, the co-localization of p12 and the viral genomic DNA adjacent to the chromosomes [7] provides support for the notion that p12, in contrast to CA, associates with the PIC till the final stages of PIC trafficking. Thus, gradual uncoating events that depend on the cell cycle can be described for MLV: first, MA, which forms the protein layer adjacent to the internal side of the virion membrane, dissolves away after fusion of this membrane with the plasma membrane [7]. Next, CA, which forms an inner protein layer in the virion and is part of the PIC, dissociates from this complex in mitotic cells. In HIV, CA is also mainly absent from the PIC [8], [9]. Moreover, swaps between Gag domains of HIV and MLV demonstrated that CA is the dominant determinant for the difference between HIV and MLV in the ability to transduce nondividing cells and led to the suggestion that the stable association of the MLV CA with the PIC prevents the access of this complex to components of the cellular transport machinery [10]. Thus, the timely dissociation of CA from the PIC in mitotic cells, demonstrated here, may expose this complex to interactions with cellular factors, necessary for the completion of the trafficking and/or p12-mediated docking of the PIC to mitotic chromosomes. This notion is further supported by the correlations between the maintenance of CA association with the p12/PIC in PM14 and S(61,65)A mutants and their inability to dock to the chromosomes; and the reversal of both phenomena in the context of wt and S(61,65)A/M63I revertant. These data are in line with the proposed cooperative effect of p12 and CA at early stages of MLV infection [53].
In summary, we identified p12 - a PIC component essential for integration - as a factor that tethers the MLV PICs to mitotic chromosomes. MLV-based vectors have been used successfully in gene therapy trials in humans, yet with the risk of leukemogenesis [54], [55]. Identification of factors influencing integration, such as p12, should lead to safer MLV-derived vectors [49]. The specific docking of MLV PICs to mitotic chromosomes, the requirement for NE disassembly and the disassociation of CA from the PIC during mitosis, may all contribute to the productive integration of MLV in dividing cells.
Moloney MLV clones wt (pNCS), PM14, 1xMycR, S(61,65)A, S(61,65)A/M63I, and pQCXIP-gfp-C1 vector were described before [7], [19], [22]. LANA31 peptide (of the Kaposi's sarcoma herpesvirus LANA protein; obtained from R. Sarid, Bar-Ilan University), or mutations S(61,65)A and S(61,65)A/M63I, were introduced into the indicated viruses, as described before for the PM14 mutation [7]. Overlapping PCR was used to generate the sequences of MA-GFP/p12-CA-NC, MA-GFP-p12-CA-NC and MA-mCherry/p12-CA-NC (Supporting Information). pEF-H2AmRFP, expressing RFP-histone H2A fusion, and pRFP-lamin A, expressing RFP-Lamin A fusion [25], were provided by M. Brandeis (The Hebrew University of Jerusalem) and H.J. Worman (Columbia University), respectively.
Culture conditions and serum starvation/aphidicolin treatment were described before [7]. 2ME2 (1.3 µg/ml; Sigma M6383) or nocodazole (15 µg/ml; Sigma M1404) were added 16 hr prior to imaging. Reversine (5 µM; Sigma R3904) was added to 2ME2-containing media when indicated. RFP-lamin A and RFP-histone H2A fusions were used to generate cell lines with labeled NE and chromosomes, respectively (Supporting Information).
Generation of, and infection with, the 1xMycR virus were described before [7]. To quantify infectivity of chimeric virions, 293T cells were co-transfected with plasmids expressing the pQCXIP-gfp-C1vector (2 µg), and the indicated molar ratios of wt MLV and MA-GFP/p12-CA-NC. A 1∶1 molar ratio represents 10 µg of pNCS and 5 µg of MA-GFP/p12-CA-NC plasmid. 48 hr post-transfection, supernatants of transfected cultures were filtered (0.45 µ), supplemented with HEPES (pH 7.0; 50 mM final concentration) and virus content was normalized by exogenous RT assay [56]. Supernatants with an equal RT activity were used to infect NIH3T3 cells for 2 hr in the presence of polybrene (hexadimethrine bromide; 8 µg/ml). Two days post-infections the cells were analyzed by FACS for the percentage of GFP+ cells. For live-cell imaging, chimeric particles were generated by co-transfecting 293T cells with a 1∶1 ratio of the indicated virus and the modified Gag as described above. Infections were carried out as described above with the following modifications: ∼8 hr before infection, the cells were plated (∼10% cofluency) in a 4-compartments-cell-view-glass-bottom-dish (35-mm; Greiner Bio One) and were infected with MOI of approximately 10 (based on the comparison of the RT activity of the samples to a standard MLV stock). For imaging, HEPES pH 7.0 (20 mM) was added to growth medium. Imaged samples were maintained at 37°C and supplied with CO2 when imaging exceeded 2 hr. In some experiments, chromosomes were stained with Hoechst 33342 (bisBenzimide H 33342 trihydrochloride, Sigma B2261; 1 µg/ml, 15 min, 37°C). Imaging was with spinning disk confocal (Yokogawa CSU-22 Confocal Head) microscope (Axiovert 200 M, Carl Zeiss MicroImaging), 100× lens (NA 1.45, Zeiss) and Evolve or HQ2 (Photometrics) cameras.
Single-cycle infection assays were carried out with the pQCXIP-GFP-C1 or pQCXIN (Clontech) vectors. For VLPs carrying the wt p12 sequence the following plasmids were co-transfected into 293T cells: vector plasmid (10 µg), VSV-G expression plasmid (2.5 µg), pGag-PolGpt helper plasmid [57] (5 µg), and the plasmid expressing the modified Gag MA-GFP/p12-CA-NC (2.5 µg). The PM14 mutation was inserted to the p12 sequences of pGag-PolGpt helper plasmid (generating pGag-PolGpt/PM14) and of MA-GFP/p12-CA-NC plasmid (generating pMA-GFP/p12-CA-NC/PM14). These two plasmids were used to generate PM14 VLPs as described above for VLPs with the wt p12 sequences. The LANA31 sequence was inserted into pMA-GFP/p12-CA-NC or pMA-GFP/p12-CA-NC/PM14, generating pMA-GFP/p12-CA-NC/LANA31and pMA-GFP/p12-CA-NC/PM14/LANA31, respectively. To generate VLPs with LANA31 module, pairs of pMA-GFP/p12-CA-NC/LANA31 and pGag-PolGpt, or pMA-GFP/p12-CA-NC/PM14/LANA31and pGag-PolGpt/PM14, were co-transfected with VSV-G and vector plasmids, using the same plasmid ratio indicated above. NIH3T3 cells were infected with the resulting VLPs, normalized by exogenous RT assay, and drug-resistant colonies were selected with either puromycin (4 µg/ml), or G418 (1 mg/ml).
For Western blotting, virions were purified through sucrose cushions using ultracentrifugation [7] and detected with anti-GFP monoclonal antibody (Covance, MMS-118R). Immunofluorescence and calculations of the co-localization degree between CA and p12, or CA and GFP-p12, and between the chromosomes and each of these proteins were performed as in [7]. Quantification of the overlap between GFP fluorescence and chromosomes (RFP fluorescence) was performed as was described before for measurements of the overlap between p12 and chromatin signals [7]. Reconstitution of 3D images was performed with SlideBook software, employing MIP (Fig. 1D) or X-Ray (Fig. S2) functions.
For the calculation of the spatial retention of PICs over time, time-lapse sequences were processed [NoNeighbors deconvolution, Laplacian 2D filtering, SlideBook software (Intelligent Imaging Innovations)]. Objects were identified through intensity-based segmentation, with no further selection for specific PICs. For each movie, the second, third or fourth frame were superimposed on the first frame; the areas of overlapping PICs identified, and the signal intensity in overlapping areas was presented as the percentage of total intensity of objects in the first frame.
|
10.1371/journal.pntd.0001345 | Human Cellular Immune Response to the Saliva of Phlebotomus papatasi Is Mediated by IL-10-Producing CD8+ T Cells and Th1-Polarized CD4+ Lymphocytes | The saliva of sand flies strongly enhances the infectivity of Leishmania in mice. Additionally, pre-exposure to saliva can protect mice from disease progression probably through the induction of a cellular immune response.
We analysed the cellular immune response against the saliva of Phlebotomus papatasi in humans and defined the phenotypic characteristics and cytokine production pattern of specific lymphocytes by flow cytometry. Additionally, proliferation and IFN-γ production of activated cells were analysed in magnetically separated CD4+ and CD8+ T cells. A proliferative response of peripheral blood mononuclear cells against the saliva of Phlebotomus papatasi was demonstrated in nearly 30% of naturally exposed individuals. Salivary extracts did not induce any secretion of IFN-γ but triggered the production of IL-10 primarily by CD8+ lymphocytes. In magnetically separated lymphocytes, the saliva induced the proliferation of both CD4+ and CD8+ T cells which was further enhanced after IL-10 blockage. Interestingly, when activated CD4+ lymphocytes were separated from CD8+ cells, they produced high amounts of IFN-γ.
Herein, we demonstrated that the overall effect of Phlebotomus papatasi saliva was dominated by the activation of IL-10-producing CD8+ cells suggesting a possible detrimental effect of pre-exposure to saliva on human leishmaniasis outcome. However, the activation of Th1 lymphocytes by the saliva provides the rationale to better define the nature of the salivary antigens that could be used for vaccine development.
| Cutaneous leishmaniasis affects millions of people worldwide and is caused by protozoa of the genus Leishmania. The parasite is transmitted during sand fly bites. While probing the skin for a blood meal, vectors salivate into the host's skin. Sand fly saliva contains several components that increase hemorrhage and interfere with the host's inflammatory response. Data obtained in mice originally indicate that immunization against saliva protected from leishmaniasis supporting possibility that leishmaniasis could be prevented by a vaccine based on sand fly saliva. Herein we investigated the nature and the importance of the cellular immune response developed against sand fly saliva by individuals at risk of cutaneous leishmaniasis due to Leishmania major. We demonstrated that the immunity against saliva is dominated by the activation of lymphocytes producing a suppressive cytokine called IL-10. These data may preclude the protective effect of sand fly saliva pre-exposure in humans. Further experiments revealed that the production of IL-10 masked the presence of a second kind of lymphocytes producing IFN-γ, a rather protective cytokine. The latter finding highlights the importance of the identification of the proteins activating the latter lymphocytes in order to develop vaccines based on selected proteins from the saliva of sand flies.
| Leishmaniasis includes a heterogeneous group of diseases that are caused by protozoan parasites of the genus Leishmania. The disease ranges from asymptomatic infections to self-limiting cutaneous lesion(s) or fatal visceral forms [1]. Leishmania parasites are transmitted to the vertebrate hosts by the bite of sand flies. During parasite inoculation in the host's skin, the vector injects the saliva that contains a large number of pharmacological components [2]–[4]. Several observations indicate that sand fly saliva is crucial in the establishment of leishmaniasis and disease pathogenesis [5]–[7]. The mechanism by which the vector's saliva enhances leishmania infection remains to be clarified. Sand fly saliva contains potent antihemostatic and vasodilatator compounds as well as potentially immunomodulatory molecules that can directly down-modulate macrophage effector functions and facilitate the establishment of the infection [8], [9]. The exacerbating effects of saliva may also be related to the early release of epidermal interleukin-4 (IL-4) [7]. Alternatively, it could be ascribed to the development of an adaptive immune response that would favor the commitment of a Th2 immunity against Leishmania. In mice, pre-exposure to saliva completely abrogated the effects of the sand fly saliva and protected the host from disease progression [7], [10], [11]. This protective effect correlated with a strong delayed-type hypersensitivity (DTH) response and an early and increased in situ production of IFN-γ and IL-12 [10]. Further experiments demonstrated that immunization with PpSP15 (gi|15963509) from P. papatasi saliva resulted in protection which was not ascribed to a humoral immune response [12]. Altogether, these data strongly supported the possibility that leishmaniasis could be prevented by vaccinating against sand fly saliva and suggest that the protective effect of the saliva might be associated with cell-mediated immunity. In humans, the data about the cellular immune responses are scarce. Herein, we analyzed the cellular immune response against the saliva of Phlebotomus papatasi developed in individuals naturally exposed to sand fly bites and demonstrated that the overall effect of Phlebotomus papatasi saliva was dominated by the activation of peripheral IL-10-producing CD8+ cells. Strikingly, the activation of IFN-γ-producing CD4+ T lymphocytes has been revealed after neutralization of IL-10 production or depletion of CD8 lymphocytes.
All experiments were conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the ethic committee of Institute Pasteur of Tunis. All patients provided written informed consent for the collection of samples and subsequent analysis.
Peripheral blood samples were drawn from 36 donors (Table 1). The sampling has been performed on April, just before the transmission season of cutaneous leishmaniasis in Tunisia (between June and October). Ten donors (B1 to B10; age range 27–52 years, mean 31.9 years) were living in Tunis, a non endemic region for ZCL but in which the presence of P. papatasi has been reported at low frequency [13]. Twenty-six (B11 to B36; age range 15–73 years, mean 40.2 years) were living in Sidi Bouzid, a region located in the center of Tunisia, which is endemic for ZCL caused by Leishmania major and characterized by the presence of P. papatasi at high frequencies [13].
In all in vitro assays, cells were cultured in RPMI 1640 medium supplemented with 10% AB human serum (Sigma, St Louis, MO), 1% sodium pyruvate, 1% non essential aminoacids, 1% HEPES buffer, 5×10−5 M/L β-mercaptoethanol and 40 µg/mL gentamycin, (Invitrogen, Cergy Pontoise, France).
Purified blocking anti-human IL-10 antibody (BD Biosciences, Le Pont de Claix, France) was used in cell culture. The following monoclonal antibodies were used for flow cytometry analysis: FITC-, and Cy-Chrome-conjugated anti-human CD3, CD4 and CD8, PE-conjugated anti-IL-4, IL-10, IFN-γ, TNF-α, granzyme B and control isotypes (BD Biosciences).
Salivary glands from a Tunisian strain of Phlebotomus papatasi were dissected out in phosphate buffer saline and disrupted by 3 freezing/thawing cycles. After centrifugation, the supernatants were stored at –80°C. Just before use, the salivary gland extract (SGE) was prepared by dilution in cell culture medium added with gentamycin (Invitrogen).
Soluble Leishmania antigen (SLA) was prepared from an isolate of L. major (zymodme MON25; MHOM/TN/94/GLC94), obtained from ZCL lesion, as previously described [14].
Peripheral blood mononuclear cells (PBMC) were isolated on a Ficoll-Hypaque gradient. In some experiments, lymphocyte subsets were separated by negative selection using magnetic beads (Miltenyi Biotec, Paris, France). Purity of T cell subsets ranged from 95% to 98%.
Cells were cultured in 96-well plates in cell culture medium at 0.5×106 cells/mL in a final volume of 200 µL and incubated with SGE (1gland/mL) or SLA (10 µg/mL) with or without anti-IL-10 or isotype control in a 5%CO2 humidified atmosphere at 37°C. The optimum condition for PBMC proliferation against SGE was determined in our laboratory in preliminary experiments. We thus stimulated cells with different concentrations of salivary gland extracts (0.25 gland/mL, 0.5 gland/mL, 1 gland/mL and 2 gland/mL) for 3, 4 and 5 days and the optimal results (highest index of proliferation corresponding to the ratio of cpm in stimulated condition / unstimulated one) were obtained with 1 gland/mL during five days. For proliferation studies, the uptake of (3H) thymidine (Amersham, Saclay, France) was measured 18 hours after adding 0.4 µCi/well. Cells were harvested and the radioactivity was counted in a scintillation counter (Rack Beta; LKB Wallace). Results were expressed as a proliferation index: mean counts of triplicates in antigen-stimulated cultures /mean counts of triplicates in unstimulated cultures.
For IL-10 or IFN-γ detection, supernatants of cell culture were collected after 48h or 72h, respectively, centrifuged and stored at –80°C until use. Capture enzyme-linked immunosorbent assay (ELISA) was performed on supernatants using Human IL-10 or IFN-γ ELISA Sets (BD Biosciences) according to manufacturer's instructions. For each cytokine determination, the results were interpolated from a standard curve using recombinant cytokines and expressed in pg/mL or as ratio of IFN-γ concentration in stimulated/unstimulated cultures.
Freshly isolated PBMC were stimulated for 24h to 96h with SGE (1gland/mL) or medium alone in 24-well plates and treated with Golgistop (BD Biosciences) for the last 6 hours of culture. Cells were then washed and incubated with FITC or PE-Cy5.5 conjugated to CD3, CD4 or CD8 antibodies for 20 minutes at 4°C. For intracellular cytokine detection, the cells were fixed and permeabilized using BD Cytoperm/cytofix plus kit (BD Biosciences) according to manufacturer's instructions and labeled with PE-conjugated anti-human IFN-γ, IL-4, IL-10, TNF-α, granzyme B or control isotype (BD Biosciences). Analyses were performed with a FACS Vantage flow cytometer using the CELLQuest software (BD Biosciences).
Specific anti-saliva IgG antibodies were assessed by ELISA and Western blot as previously described [15].
For ELISA, wells were coated with SGE (0.5 gland per well) in 0.1 M carbonate-bicarbonate buffer overnight at 4°C. The wells were then washed in phosphate buffer (PBS) added with 0.1% Tween 20 and incubated with PBS-Tween20–0.5% gelatin for 1 hour at 37°C to block free binding sites. Diluted sera (1∶200) were then incubated for 2 hours at 37°C. Antibody-antigen complexes were detected using peroxidase-conjugated anti-human IgG antibody diluted at 1∶10000 (Sigma) for 1 hour at 37°C and were visualized using Orthophenylendiamine in citrate buffer and hydrogen peroxide. The absorbance was measured using an automated ELISA reader (Awareness Technology Inc) at 492nm wavelength. The cut-off for the assays was the mean optical density obtained with sera of 20 negative controls obtained from the study of Marzouki et al. [15] plus 3 standard deviations.
For Western-blot analysis, the equivalent of 40 to 60 salivary glands were loaded in a single long well and separated on 15% sodium dodecyl sulfate (SDS)-PAGE gel. The separated proteins were then transferred onto a nitrocellulose membrane. The membrane was incubated overnight at 4°C with blocking buffer containing 5% non-fat milk and then cut into 8 to10 strips. Each strip was incubated for one hour at room temperature with diluted serum samples (1∶200). After washing, the strips were incubated with horseradish peroxidase-linked anti-human IgG antibody (Sigma) at 1∶10000 for one hour at room temperature. After five washings, positive bands were visualized using enhanced chemiluminescence (Amersham).
Values obtained in two different groups were compared by the non parametric Mann-Whitney U test using StatView software. Statistics of results obtained as paired data were performed using the Paired-T test. The correlation between different parameters was analyzed using Spearman's rank correlation. The exact Fisher test was used to compare the frequency of individuals exhibiting a positive proliferative response against the sand fly saliva between the stratified groups of donors. Statistical significance was assigned to a value of p<0.05.
The cellular immune response against P. papatasi salivary gland extracts (SGE) was first assessed by studying the proliferative responses of peripheral blood mononuclear cells (PBMC) from our donors. The optimum condition for the proliferation test has been determined previously (namely, stimulation with 1 gland/ml of SGE during five days of culture) (Figure S1). We designated as positive SGE proliferation the cases exhibiting an index of PBMC proliferation against the saliva of P. papatasi above 2. We thus demonstrated that 11 out of 36 donors exhibited positive SGE proliferation (Figure 1). The indices of proliferation obtained in these donors were relatively low with a median of 3.2 but significantly higher than those of donors with negative SGE responses (p<0.0001) (Figure 1). The percentage of individuals with positive SGE proliferation was higher in the group of donors living in endemic area of leishmaniasis than in donors living in northern parts of Tunisia (9 out of 26 donors versus 2 out of 10 donors). The difference was, however, not significant (p>0.05).
Twenty-six donors exhibited positive proliferation against SLA (soluble Leishmania antigen). Except for B7 who has been accidentally exposed to Leishmania parasites while manipulating in a laboratory, the latter results suggest that these donors have been exposed to infected sand fly bites. Consistently, in 18 out of 21 donors with positive proliferation against SLA, a delayed-type hypersensitivity response against Leishmania antigens has been demonstrated (Table 1). Interestingly, a cellular immune response against SGE was more frequently noted in individuals who exhibited a PBMC proliferation against SLA compared to those with no proliferation against SLA (10 out of 26 donors versus 1 out of 10 donors) (Table 1). However, no correlation was found between the level of proliferative responses against SGE and this of proliferation against SLA (soluble Leishmania antigen) in all studied donors (p>0.05) (Table 1 and Figure 1). In donors with a cellular immune response against L. major (positive proliferation against SLA), no association between proliferative responses against SGE and the presence or absence of a past history of ZCL was found (p>0.05) (data not shown).
Cellular immune responses against SGE were further tested by monitoring cytokine secretion in supernatants of stimulated PBMC. IFN-γ was detected at a median concentration of 96 pg/ml in unstimulated PBMC (data not shown). In supernatants of SGE-stimulated cells, concentrations of IFN-γ were not significantly different from that detected in unstimulated conditions (p>0.05) and median levels were comparable in individuals with positive and negative proliferation against SGE (median concentrations 93 pg/ml versus 90 pg/ml respectively, p = 0.63) (Figure 2A). While IL-10 was detected at a median concentration of 28pg/ml in unstimulated PBMC (data not shown), IL-10 concentration in supernatants of stimulated PBMC was significantly increased in individuals showing a specific proliferation against SGE (median concentration of 128 pg/ml, p<0.0001) but not individuals with negative proliferation (median concentration of 23 pg/ml) (Figure 2B). When the threshold of IL-10 induction was defined as the 95th percentile of the values obtained in individuals with negative proliferation, we found that IL-10 was significantly induced in stimulated PBMC from all but one of the individuals showing proliferative response against SGE (Figure 2B). Conversely, in one donor (B13) with negative proliferation against SGE, IL-10 level in supernatants of SGE-stimulated PBMC was quite superior to the threshold (Figure 2B). Interestingly, a significant correlation was observed between proliferation and IL-10 induction in stimulated PBMC from all donors (p = 0.0002) (Figure 2C). Finally, the effects of SGE on IL-10 production did not seem to be related to LPS contamination as no difference in IL-10 production by PBMC stimulated with SGE with or without polymyxin B was observed. Moreover, TNF-α was not induced after SGE stimulation.
To better define the features of the cellular population activated by the salivary gland extract, we focused our study on five representative donors (identified as B2, B6, B11, B12 and B20) from the group of donors showing proliferative response against SGE. These donors were selected as they covered the different range of proliferation positivity and were demographic matched to the remaining of the cohort (2 from the non endemic area and 3 from the endemic area). Intracytoplasmic expression of cytokines (IFN-γ, IL-10, IL-4 and TNF-α) and granzyme B was studied by flow cytometry in freshly isolated peripheral cells obtained from these donors.
In accordance with data obtained above, SGE did not induce IFN-γ production at any of the tested time points (Figure 3A). Similar results were obtained after stimulation by phorbol myristate acetate (PMA) and ionomycin, used to increase the sensitivity of the test (data not shown). Contrastingly, SLA (Soluble Leishmania antigen) induced in donors B2, B11, B12 and B20 (with positive proliferation against such antigen), the production of high amounts of IFN-γ (data not shown).
Similarly, TNF-α secretion was not induced by SGE (Figure 3B). By contrast, IL-10 was detected in 4.8% to 11% of total PBMC that have been stimulated with SGE (Figure 4A and not shown). Surprisingly, IL-10 was produced primarily by CD8+ T lymphocytes and CD3-negative cells. Indeed, SGE induced the production of IL-10 in 2.4% to 9.4% (median 5%) of CD8+ T cells (Figure 4A and 4B). Notably, we also performed flow cytometry analyses in two individuals (B3 and B8) living in non endemic area of leishmaniasis and who did not exhibit neither PBMC proliferation nor IL-10 production after stimulation with SGE. No cytokine production was detected (Figure 4A and not shown).
The detection of specific CD8+ T cells was rather unexpected. The induction of granzyme B was assessed after SGE stimulation to determine whether these cells displayed cytotoxic functions. As shown in Figure 4C, granzyme B was detected in almost 11% of unstimulated CD8+ T lymphocytes. The percentage of granzyme-B-producing-CD8+ cells did not increase after SGE stimulation (median percentage of 7.62% in stimulated versus 7.15% in unstimulated cells, p>0.05). This suggests that SGE did not activate cytotoxic CD8+ T lymphocytes. By contrast, the percentage of IL-4-producing CD8+ T lymphocytes increased after SGE stimulation from a mean percentage of 2.03% to 4.87% further supporting that activated CD8+ T cells are Th2 cells (Figure 4D and not shown). Contrasting with results obtained with IL-10, SGE stimulation did not induce a substantial IL-4 production by CD3 negative cells (Figure 4D).
We then tested whether the high production of IL-10 by CD8+ lymphocytes and CD3-negative cells could account for the relatively low proliferation of PBMC and the absence of IFN-γ production in response to SGE. Blocking anti-IL-10 antibody was used at different concentrations since the amount of IL-10 detected in the supernatants of stimulated lymphocytes was variable from one donor to another. As shown in Figure 5A and 5B, adding a blocking IL-10 antibody significantly enhanced the proliferation and IFN-γ production of stimulated PBMC from individuals with positive proliferation (B2, B6, B11, B12 and B20).
To better define the subsets of T cells that were activated by the saliva of Phlebotomus papatasi, the same experiments were subsequently performed on CD4+ and CD8+ T lymphocytes magnetically separated from PBMC of donors that exhibited proliferative responses to SGE. As shown in Figure 6A, SGE induces the proliferation of both CD4+ and CD8+ T lymphocytes. Blocking the production of IL-10 significantly enhanced the proliferation of CD8+ T lymphocytes. This suggests that the activated CD8+ T cells produce high amounts of IL-10 inhibiting subsequently their own proliferation. By contrast, SGE induced a considerable proliferation of CD4+ T cells that was only slightly enhanced after IL-10 blockage (Figure 6A). SGE did not induce any IFN-γ production by CD8+ T lymphocytes confirming that the specific CD8+ T lymphocytes were not Th1 cells (Figure 6B). However, when the PBMC was depleted from CD8+ T cells and stimulated with SGE, they produced a high amount of IFN-γ and this was further increased when IL-10 was neutralized (Figure 6B). Intracytoplasmic analyses performed in donors B2, B6, B11, B12 and B20 confirmed that the neutralization of IL-10 as well as the depletion of CD8+ cells induces the production of IFN-γ by stimulated CD4+ T lymphocytes (Figure 6C and not shown). The percentage of IFN-γ-producing cells within the stimulated-CD4 T cells varied from 0.93% to 1.54% when IL-10 was blocked and from 0.89% to 1.81% after CD8-depletion. Furthermore, the mean of fluorescence of IFN-γ staining of the activated-CD4 T cells increased after IL-10 blockage or CD8-depletion (Figure 6C and not shown). For instance, in CD4 T cells from CD8-depleted PBMC, the mean fluorescence of IFN-γ staining increased from 21.4 to 34.7 after IL-10 blockage. Altogether, these results demonstrate that SGE induces the activation of CD8+ T cells of Th2 phenotype as well as Th1-polarized CD4+ T lymphocytes that are probably suppressed by IL-10- producing CD8+ T cells.
Previous studies in mice suggested that salivary components can be used to design vaccines able to control Leishmania infection and stressed the role of cellular immunity as correlate for protection [10]–[12]; [16], [17]. In humans, few data have been obtained on cellular immunity against sand fly saliva in volunteers experimentally exposed to the bites of uninfected P. papatasi [18] or Lutzomyia longipalpis [19]. In the present study, we analyzed the cellular immunity developed by naturally exposed individuals against P. papatasi sand fly saliva and demonstrated that salivary gland extract from uninfected sand flies induces the activation of IL-10 and IL-4-producing CD8+ T cells as well as IFN-γ-secreting CD4+ T lymphocytes.
Since there is no standard for the identification of individuals immunized against the saliva of P. papatasi, our strategy consisted of studying individuals who have high risk of contact with the vector of L. major. The studied population comprised individuals living in an endemic area of L. major transmission in Tunisia and thus at risk of being exposed to sand fly bites. We also included individuals who live in non endemic areas of cutaneaous leishmaniasis as the presence of Phlebotomus papatasi has also been reported in these areas [13]. We first assessed the proliferation and cytokine secretion of PBMC after stimulation with salivary gland extracts. A specific proliferative response of PMBC was evident in nearly 30% of donors. A similar percentage (24%) has been recently confirmed in a large number (n = 425) of individuals naturally exposed to sand fly bites (Kammoun et al. in preparation). Interestingly, the proliferative response was more frequently observed in individuals with a cellular immunity against Leishmania, a result that might reflect a more frequent exposure to sand fly bites in endemic areas. The positive cellular response against SGE was also reported in two out of the ten volunteers living in Tunis, corroborating the data reporting the presence of P. papatasi in the northern parts of Tunisia [13]. Except for the donor B7 who has been accidentally exposed to Leishmania antigens while manipulating parasites in the laboratory, 15 individuals exhibit positive proliferation against SLA but did not develop any cellular response to SGE. Such donors, however, developed specific antibodies against saliva as demonstrated by ELISA and/or Western blot analysis (Table 1), thus indicating a previous contact with the saliva of P. papatasi. For such donors, it seems that there was no correlation between the antibody and the cellular response. To our knowledge, the proliferative immune response toward the saliva of sand flies was not previously described in the literature although several works reported the presence of specific antibodies in people naturally exposed to sand fly bites. One possible explanation is the dominance of IL-10 production in response to the saliva of P. papatasi. This IL-10 will inhibit both proliferation and IFN-γ production of CD4 T cells. Furthermore, humans are outbred and therefore very diverse in term of pattern of immune response.
In contrast with results showing an increased production of IFN-γ at the inoculation site in mice sensitized with P. papatasi salivary proteins [10], our data did not detect any IFN-γ production by stimulated PBMC at all tested time-points. Noticeably, monitoring cytokine production by PBMC either on supernatant cell culture or using intra-cytoplasmic analysis revealed an increased synthesis of IL-10. The induction of IL-10 by sand fly saliva was previously reported in mice [20]. The production of IL-10, particularly by innate effectors such as macrophages, may explain the enhancing effect of the saliva on leishmania infectivity. Accordingly, in vitro experiments have demonstrated that the saliva from either P. papatasi or L. longipalpis exhibit immunosuppresive effects on macrophages and reduce the nitric oxid production [8], [9], [21]. The latter effect was ascribed to adenosine [22], an immunomodulatory component of P. papatasi saliva known to induce IL-10 [23]. In humans, however, little is known about the cytokine pattern activated by sand fly saliva. While the saliva of L. longipalpis can activate both IFN-γ and IL-10 synthesis in stimulated PBMC [19], the salivary gland lysates from P. papatasi inhibit the production of IFN-γ from L. major-stimulated PBMC [24], an effect which may be ascribed to IL-10. Interestingly, in our study, blocking IL-10 enhanced the proliferation of stimulated PBMC suggesting that IL-10 production may explain the relatively low rate of PBMC proliferation and perhaps the difficulty of detecting a proliferative response against SGE either in humans or in mice (unpublished data). Accordingly, Rohousova et al. reported an inhibitory effect of P. papatasi saliva on lymphocyte proliferation in mice [25]. Yet, despite the fact that IL-10 suppresses the proliferation of PBMC, we demonstrated that the production of IL-10 was mainly detected in the cells, which proliferated in response to SGE. Such paradoxical data could be easily explained. In fact, the induction of IL-10 by SGE could be illustrated only in patients who were previously exposed to Phlebotomus papatasi bites and so exhibiting proliferative responses to SGE. Interestingly, in one donor who exhibited a moderate production of IL-10 by stimulated PBMC, the cell proliferation was demonstrated only when the effect of IL-10 was blocked. In a current prospective work studying the cellular immune response of four hundred donors living in endemic areas of leishmaniasis in Tunisia, similar results were obtained and in approximately 10% of donors, the proliferative response of SGE-stimulated PBMC was revealed after IL-10 neutralization (Kammoun et al. manuscript in preparation). Interestingly, the five donors included in our study to define the phenotype of activated cells by SGE have been chosen as they covered the different range of proliferation against saliva and were demographic matched to the remaining donors of the cohort. Among them, four were in contact with Leishmania while one was not. The cellular immune response against the saliva was comparably dominated by IL-10 production whatever the donor has been in contact with the parasite or not. Indeed, the level of IL-10 in the supernatant of SGE stimulated PBMC was similar in the 5 donors (data not shown). Although SLA and SGE do not activate the same effector cells, some experiments in which PBMC were co-stimulated with both antigens were performed. Adding SLA did not change the pattern of immune response against SGE that remained dominated by IL-10 (data not shown).
Unexpectedly, our analysis of the activated cells revealed that CD8+ T lymphocytes were a major source of IL-10 synthesis. Detection of T CD8+ specific T cells was rather unexpected and the analysis of IL-4 expression confirmed the Th2 phenotype of these cells. This is the first account in which SGE-specific CD8+ T cells of Th2 phenotype were reported in humans. In mice, Mbow et al. reported a direct enhancing effect of P. papatasi saliva on IL-4 expression in the absence of Leishmania infection [26]. Another study revealed large numbers of IL-10 producing CD4+ and CD8+ T cells in draining lymph node of mice injected with SGE. However, since SGE was co-injected with the parasite, the direct effect of the saliva remains uncertain [20]. In donors with positive proliferation against SGE, a significant percentage of CD3 negative cells (detected on the lymphocyte gate) also produced IL-10 when stimulated with SGE. Preliminary experiments using CD14 and CD16 staining suggest that these cells were not monocytes or NK cells. B lymphocytes could be a possible cellular source of IL-10. Analysis of such hypothesis is under progress as it could bring new insights consistent with the recent literature showing the role of regulatory B lymphocytes in autoimmunity and infectious diseases [27], [28].
Compelling data suggest that the protective effect of pre-exposure to saliva in mice results from the skewing of the anti-leishmania immunity towards a Th1 protective response [10], [12]. Strikingly, our experiments showed that in donors exhibiting proliferative response against SGE, the presence of IFN-γ-producing CD4+ T lymphocytes could be revealed after neutralization of IL-10 production or depletion of CD8 lymphocytes. This indicates that the overall effect of P. papatasi saliva, which is dominated by the production of IL-10, may inhibit the activation of the specific Th1 cells. One may thus expect that the recall of IL-10-producing T lymphocytes in individuals previously sensitized with sand fly saliva would lead to the commitment of the immunity to Leishmania towards a Th2 response. This would be consistent with the lack of a protective effect of the pre-exposure to sand fly saliva in humans. Accordingly, epidemiological data indicate that the incidence of cutaneous leishmaniasis in the Old World is high in endemic areas despite the common occurrence of bites from uninfected sand flies [29]. Furthermore, recent studies indicate that the presence of antibodies against the saliva of P. papatasi in individuals living in endemic areas of ZCL in Tunisia was associated with an increased risk of disease [H. Louzir, personal communication, 15]. In the current prospective work mentioned above, monitoring either the cellular and humoral immune response against sand fly saliva in people living in endemic areas of cutaneous leishmaniasis throughout several transmission seasons could provide some clues regarding the effect of such immunity on the natural history of leishmaniasis. Strikingly, the effect of pre-exposure to saliva on the outcome of leishmaniasis in humans may differ from the reported effect in mice. Even in mice, conflicting results were obtained when the saliva of other species of sand flies was used [30]. Interestingly, a recent report from Rohousova et al. [31] gave an attractive explanation to the conflicting data between the possible non-protective role of sand fly bites suggested by observations from the field and the experimental data on mice. Indeed, the authors showed that the protective effect of pre-exposition to Phlebotomus duboscqi bites was limited to short-term exposure while a long-term exposure regimen to saliva, a scheme close to what occur in naturally exposed individuals to sand fly bites, was not protective. The authors hypothesized that an immunization with a large antigen load tends to skew the immune system towards a Th2 response, which could not be associated with protection.
Saliva is composed of a large broad of proteins. Variations in the protein composition and antigenicity of the saliva from different sand fly species may account for the conflicting results obtained from different studies [32]. Genetic differences between hosts may also underlie the disparity of the immune responses elicited by salivary proteins from the same sand fly species [17], [33] and may perhaps explain the putative differential effects of the P. papatasi saliva between humans and mice. Additionally, the sand fly saliva may elicit a response that differs from that elicited by its separate proteins [16]. For instance, distinct pattern of immune response could be induced by different salivary proteins from P. papatasi resulting in contrasting outcomes of L. major infection in mice [16]. PpSP15 had a protective effect, which correlated with a Th1 anti-Leishmania immune response whereas PpSP44 (gi|15963519) exacerbated the disease. Our data indicate that the saliva of P. papatasi can stimulate human CD4+ and CD8+ T cells with contrasting cytokine profiles thereby suggesting the implication of different salivary components in the activation of these populations. Defining these salivary proteins would be crucial to predict their specific effects on the outcome of leishmaniasis and to determine potential vaccine candidates.
Salivary antigens that are introduced in the host skin are drained to the lymph node where the immune response is triggered after activation of circulating lymphocytes. Reactivation of an acquired immunity against the saliva may lead to protection or exacerbation of L. major infection by skewing the immunity against the concomitantly inoculated Leishmania antigens to a Th1 or Th2 pattern, respectively. Several data obtained in mice suggested that the immunity against sand fly saliva might also act on Leishmania infection by creating an inhospitable in situ environment for the establishment of parasites. Hence, it would be critical to characterize the nature and phenotype of cells that are recruited in situ following exposure to bites of uninfected sand flies.
To our knowledge, this is the first account that describes the different features of the adaptive cellular immunity against the sand fly saliva in humans. When tested in the whole PBMC, salivary gland extracts from P. papatasi induced a low rate of proliferation, which contrasted with significant levels of IL-10 and the absence of IFN-γ. Preliminary results suggest that the CD8+ T cells activated by the saliva of P. papatasi could be γδ T lymphocytes and that the target components might be adenylated antigens as recently described [34], [35]. These data need however further confirmation. Whatever the target salivary component triggering the activation of IL-10 producing cells, this pattern of immune response may favor the Leishmania infection and facilitate the multiplication of the parasite co-injected with the saliva. Our data also demonstrated the activation of specific IFN-γ-producing CD4+ T cells that was revealed after the separation of T cell subsets. This is of great interest as it provides a new rationale for immunological approaches targeting the salivary components activating a Th1 response that could be useful for vaccination against leishmaniasis.
|
10.1371/journal.pgen.1007253 | Large-scale profiling of noncoding RNA function in yeast | Noncoding RNAs (ncRNAs) are emerging as key regulators of cellular function. We have exploited the recently developed barcoded ncRNA gene deletion strain collections in the yeast Saccharomyces cerevisiae to investigate the numerous ncRNAs in yeast with no known function. The ncRNA deletion collection contains deletions of tRNAs, snoRNAs, snRNAs, stable unannotated transcripts (SUTs), cryptic unstable transcripts (CUTs) and other annotated ncRNAs encompassing 532 different individual ncRNA deletions. We have profiled the fitness of the diploid heterozygous ncRNA deletion strain collection in six conditions using batch and continuous liquid culture, as well as the haploid ncRNA deletion strain collections arrayed individually onto solid rich media. These analyses revealed many novel environmental-specific haplo-insufficient and haplo-proficient phenotypes providing key information on the importance of each specific ncRNA in every condition. Co-fitness analysis using fitness data from the heterozygous ncRNA deletion strain collection identified two ncRNA groups required for growth during heat stress and nutrient deprivation. The extensive fitness data for each ncRNA deletion strain has been compiled into an easy to navigate database called Yeast ncRNA Analysis (YNCA). By expanding the original ncRNA deletion strain collection we identified four novel essential ncRNAs; SUT527, SUT075, SUT367 and SUT259/691. We defined the effects of each new essential ncRNA on adjacent gene expression in the heterozygote background identifying both repression and induction of nearby genes. Additionally, we discovered a function for SUT527 in the expression, 3’ end formation and localization of SEC4, an essential protein coding mRNA. Finally, using plasmid complementation we rescued the SUT075 lethal phenotype revealing that this ncRNA acts in trans. Overall, our findings provide important new insights into the function of ncRNAs.
| Genomes from different organisms produce noncoding RNAs that are not translated to make proteins and whose functions are largely unknown. There are approximately 2,000 noncoding RNAs that make up almost 25% of the yeast genome compared to the approximately 6,000 protein coding genes that make up 70% of the yeast genome. With this large number of ncRNAs there is a need for large-scale studies to determine the functional roles of ncRNAs. We take advantage of a recently developed resource of 532 yeast strains in which individual noncoding RNA genes have been deleted. We grow these yeast noncoding RNA deletion strains in different conditions and catalogue how each noncoding RNA contributes to cell growth. Improvement or inhibition of cell growth under particular conditions implicates the deleted RNA in cellular responses to those conditions. We have also investigated in more detail the function of specific noncoding RNAs, revealing examples of how a deletion influences nearby genes, and other examples of noncoding RNAs that regulate genes at distant genomic locations. We have made our extensive data on the fitness of noncoding RNA deletion mutants publically available for searching and bulk download in a new online resource, called Yeast ncRNA Analysis (YNCA). This large-scale analysis of noncoding RNA deletion mutants reveals the importance of many noncoding RNAs in cellular function.
| Eukaryotic cells express a wide variety of RNAs that do not code for proteins but contribute to the many essential functions within cells. The process of protein synthesis by translation requires ribosomal RNAs (rRNAs) to form the ribosomal subunits and transfer RNAs (tRNAs) to bring the amino acids to the ribosome [1,2]. Another class of RNAs called small nucleolar RNA (snoRNAs) predominantly catalyze the modification or processing of other RNAs, but additional novel functions for snoRNAs are emerging [3]. The small nuclear RNAs (snRNAs) of the spliceosome are required for the recognition and removal of introns from pre-messenger RNA [4]. The functions of most of these so called classical noncoding RNAs (ncRNAs) have been known for some time.
More recently, expression analysis of eukaryotic genomes has established that pervasive transcription produces an abundance of ncRNAs whose functions are largely unknown [5–9]. In human cells, where some ncRNA functions are known, there tends to be three mechanistic themes for ncRNA function where ncRNAs act as either decoys to titrate proteins away from their binding sites, scaffolds to bring proteins together or guides to recruit proteins to DNA [10]. A number of methods to probe the functional significance of the numerous ncRNAs in humans have been utilized. For example, ncRNA gene deletion, targeting ncRNAs with RNAi and repression of ncRNA transcription with CRISPR based methods are just a few techniques used to investigate the functions of expressed human ncRNAs [11–15]. Mutations in ncRNAs are also increasingly being associated with human diseases [16–18].
In the yeast Saccharomyces cerevisiae, tiling arrays and strand-specific RNA sequencing analyses have identified novel classes of ncRNAs that are distinct from the classical ncRNAs. Two classes of ncRNAs were initially identified according to their half-life in the cell, the stable unannotated transcripts (SUTs) had a relatively long half-life whereas the cryptic unstable transcripts (CUTs) were RNAs with a short half-life and were revealed only after deletion of the exosome complex exoribonuclease Rrp6 [9,19]. Deletion of the cytoplasmic exonuclease Xrn1, followed by RNA sequencing, revealed another class of ncRNAs termed Xrn1-sensitive unstable transcripts (XUTs) [20,21], some of which overlap with either a SUT or CUT. Subsequently, depletion of the RNA binding factor Nrd1 revealed a fourth class of ncRNA termed Nrd1-unterminated transcripts (NUTs) [22] and deletion of the histone methyltransferase Set2 has identified yet another class of ncRNA called the Set2-repressed antisense transcripts (SRATs) [23]. With the numbers of these yeast ncRNAs in the thousands only a very small proportion have been ascribed a function to date.
Where there are examples of ncRNA function in yeast one emerging theme is that ncRNA transcription can either induce or repress the expression of an adjacent gene [24–27]. One mechanism whereby ncRNA expression can induce or repress nearby gene expression is through chromatin modification [28]. An investigation into the influence of transcription by 180 anti-sense SUTs on the overlapping yeast genes found no direct relationship between antisense SUT transcription and protein abundance from the overlapping reading frame, indicating that the presence of an antisense SUT does not necessarily mean it regulates protein abundance from the sense protein coding gene [29]. Analysis of six intergenic SUTs using the synthetic genetic array (SGA) technology, to identify genetic interactions between deletions of these six SUTs and non-essential protein deletion strains, linked two SUTs to specific cellular functions and provided evidence that they may function in trans [30]. Many of the SUTs and CUTs are associated with specific RNA binding proteins within the yeast cell that are distinct from those bound by mRNAs to presumably allow them to carry out their specific function [31]. There is also evidence from ribosome profiling techniques that some yeast unannotated ncRNAs associate with ribosomes and can be translated into protein, so may not necessarily be noncoding [32–34]. As many of these studies have only investigated the function of a small subset of ncRNAs, a large scale analysis of ncRNA function in yeast would be useful for defining the role in the cell of the remaining ncRNAs.
We have utilized the recently developed collection of ncRNA deletion strains [35], which we have now expanded further, to carry out large-scale functional analysis of ncRNAs in yeast. In total 532 different ncRNA deletions were investigated encompassing tRNAs, snRNAs, snoRNAs, SUTs, CUTs and other annotated ncRNAs that do not overlap protein coding genes. Using both the heterozygous and haploid ncRNA deletion strain collections we have analyzed quantitatively, in a variety of growth conditions and phases, the influence that deletion of each ncRNA has on cellular fitness. This fitness analysis identified novel environmental-dependent haplo-proficient and haplo-insufficient growth phenotypes which provided key information on ncRNA function. Additionally, we have analyzed four essential ncRNAs of unknown function and have determined how deletion of these ncRNAs influenced surrounding gene expression. Moreover, we identified one ncRNA that works in trans and characterized a more detailed function for one of these ncRNAs in regulating the expression, 3’ end formation and localization of an essential protein coding mRNA. Overall, these data significantly expand the information available on the function of ncRNAs in yeast. Finally, the extensive catalog of functional data has been compiled into an easy to use website called YNCA providing an important resource for future ncRNA research.
The ncRNA deletion strain collections, as previously reported, contained 428 heterozygous diploid deletion strains in the reference strain BY4743, 373 haploid (MATa), 370 haploid (MATα) and 331 homozygous diploid ncRNA deletion strains giving a total of 1502 strains for functional analysis of ncRNAs [35]. Each ncRNA, that did not overlap with a protein coding gene, was deleted with the KanMX cassette while simultaneously introducing two unique molecular barcodes to allow identification of each deletion strain. We have now expanded this collection by the addition of 81 heterozygous diploid, 66 haploid (MATa), 67 haploid (MATα) strains and 63 homozygous diploid ncRNA deletion strains to give a total of 1779 strains (S1 and S2 Tables). Within these collections 532 different individual ncRNAs have been deleted in at least one strain background. These new combined collections of strains were utilized for fitness profiling to determine how ncRNA deletion affected the growth of cells under a variety of conditions.
To quantify the impact of ncRNAs on cellular fitness, competition experiments were carried out using the heterozygous deletion collection, with the deletion strains pooled and grown in six different liquid media. Two biological repeats were carried out for each condition. After an initial batch phase, the strains were propagated in continuous culture (steady state), an open system in which the amount of nutrients and pH are kept constant, allowing small fitness differences to be detected [36]. Specifically, cells were grown under carbon-limited and nitrogen-limited conditions at both 30°C and 36°C. Cells were also grown under carbon-limited and nitrogen-limited conditions at 30°C in the presence of 100mM LiCl which is known to inhibit the exoribonuclease Xrn1 and stabilize RNA [37]. Culture samples were removed for analysis at the beginning (initial pool, P) and end of the batch growth (B), at early steady state (ESS), mid steady state (MSS) and late steady state (LSS) time points (Fig 1A) to compare the composition of these populations with each other. Genomic DNA from each sample was isolated and the unique molecular barcodes identifying each deletion strain were amplified for next generation sequencing (Bar-Seq) [38–40] to determine the abundance of each ncRNA deletion strain in the population. As there were two biological repeats a total of four independent barcodes were sequenced for each ncRNA deletion strain. Under-representation of specific deletion strains highlights haplo-insufficient phenotypes, namely ncRNAs that are quantitatively important for phenotypic maintenance. Over-representation of deletion strains (haplo-proficient phenotypes) suggest that lowering the copy number of specific ncRNAs is beneficial in that particular environmental context.
We first compared the population fitness profile between the initial pool and batch stage to identify strains that displayed either haplo-insufficiency or haplo-proficiency in the six different conditions tested (Fig 1B and 1C; S3–S10 Tables). The tRNA, tR(CCU)J, also known as HSX1, displayed an extreme haplo-insufficient phenotype between the pool and batch stages in all six conditions we tested. The tRNA tR(CCU)J is a single copy rare tRNA gene [41] and is clearly required for batch growth of yeast. The reduced fitness of the tR(CCU)J deletion strain from pool to batch indicates that the function of tR(CCU)J is critical when nutrients become limiting. Reduced fitness of the tR(CCU)J deletion strain was also validated in monoculture under nutrient rich (YPD), carbon-limited and nitrogen-limited conditions at 30°C (S1 Fig). Another tRNA, tA(UGC)O, displayed haplo-insufficiency between the pool and the batch stage in both carbon-limited and nitrogen-limited conditions, but only at 36°C (Fig 1B and 1C; S10 Table), suggesting that this tRNA is required for fitness under conditions of heat stress.
A number of the heterozygote deletion strains displayed better growth, haplo-proficiency, between the pool and batch stages. Interestingly, CUT248 deletion was haplo-proficient in all the conditions where nitrogen was limited (Fig 1C; S10 Table) which was confirmed in monoculture (S1 Fig). CUT248 is located near DPS1 (Fig 2A) which is known to be up-regulated during yeast fermentation in the presence of diammonium phosphate [42]. Analysis of DPS1 expression by quantitative real-time PCR (qRT-PCR) confirms that deletion of CUT248 induces an increase in DPS1 expression in rich media and nitrogen-limiting conditions (Fig 2A). CUT248 therefore appears to repress DPS1 transcription. Lowering the amount of CUT248 in a diploid background, allows increased DPS1 expression, which could be beneficial for growth in nitrogen-limited conditions. In contrast to the deletion of CUT248, overexpression of the CUT248 RNA sequence from a plasmid in a wild-type haploid strain BY4741 results in a slow growth phenotype (S2 Fig) suggesting further that the levels of CUT248 are important for cellular fitness. A noticeable influence of temperature on fitness can be observed in the strain carrying the SUT340 deletion during the pool to batch transition (Fig 1B and 1C). SUT340 displays strong haplo-insufficiency at 36°C in both carbon-limited and nitrogen-limited conditions but not in any of the conditions at 30°C, revealing that this ncRNA with no known function is required for growth at high temperature. Additionally, RNA stabilization through inhibition of Xrn1 with LiCl triggers haplo-proficiency of CUT873 and tT(AGU)J specifically in carbon-limited conditions between the pool and batch stages (Fig 1B; S10 Table). We have also tested the deletion mutants tA(UGC)O, SUT340, CUT873 and tT(AGU)J in monoculture using the same conditions in which haplo-insufficient and haplo-proficient phenotypes were observed, and reconfirmed their phenotypes. The tA(UGC)O and SUT340 deletion mutant strains which were haplo-insufficient are both significantly less fit than the WT strain when grown in monoculture (S3A Fig). Similarly, the CUT873 and tT(AGU)J deletion mutant strains which were identified as being haplo-proficient are both significantly fitter than the WT strain when grown in monoculture (S3B Fig).
Deletion of the overlapping ncRNAs SUT233/CUT707 results in haplo-insufficiency in four of the six conditions in the pool to batch transition (Fig 1B and 1C; S10 Table) which was confirmed in monoculture (S1 Fig). SUT233 lies upstream of the gene HAP4 which codes for a transcription factor involved in the diauxic shift in yeast [43–45]. CUT707 lies upstream of KTI12 which codes for a protein that in yeast associates with the elongator complex required for tRNA modification [46,47]. HAP4 expression is increased during the diauxic shift to allow the upregulation of the glyoxylate cycle with HAP4 inducing the expression of approximately 88% of the proteins made during the diauxic shift [43,45]. Analysis of HAP4 and KTI12 expression by qRT-PCR confirms that in nitrogen-limiting conditions deletion of SUT233/CUT707 reduces expression of both HAP4 and KTI12 (Fig 2B). To show the number of haplo-insufficient and haplo-proficient ncRNA deletion strains in common between conditions in the pool to batch experiments UpSet diagrams of intersecting sets have been provided (S4 Fig).
We next compared the fitness of the heterozygote ncRNA deletion strains between the early LSS and ESS stages (Fig 3; S3–S9 and S11 Tables). By keeping nutrients, pH and growth rate constant we were able to quantify smaller differences in fitness in response to changes in temperature. For example, deletion of SUT089 displayed haplo-proficiency in both carbon-limited and nitrogen-limited conditions at 30°C. However, this haplo-proficiency of SUT089 was significantly buffered in both carbon-limited and nitrogen-limited conditions at 36°C. Another striking example of temperature affecting the fitness of a heterozygous diploid ncRNA deletion strain is the large increase in fitness of SUT467 in nitrogen-limited conditions when temperature is increased from 30°C to 36°C. We have also found that SUT471 is haplo-proficient in all six conditions, therefore its presence clearly limits growth in continuous culture conditions. Under continuous culture conditions tR(CCU)J, which displayed severe haplo-insufficiency in the pool to batch growth phase, did not display any significant growth defect (Fig 3A and 3B; S11 Table). Therefore, analysis of deletion strains under continuous culture conditions clearly reveals additional phenotypes not seen in traditional batch culture where nutrients become limiting. To show the number of haplo-insufficient and haplo-proficient ncRNA deletion strains in common between conditions in the ESS to LSS experiments UpSet diagrams of intersecting sets have been provided (S5 Fig).
Co-expression analysis has been used widely to infer functional relationships between protein encoding genes [48–51]. Here we apply a similar approach to our fitness data from eight different data sets to look for ncRNA deletion strains with similar fitness profiles and uncover phenotypic networks in the heterozygous ncRNA deletion collection. Four clusters were identified for a total of 226 deletion mutants which accounts for approximately 40% of the original dataset (Fig 4; S12 and S13 Tables). Our results indicate that deletion strains within each cluster followed the same fitness pattern throughout the eight testing conditions. Cluster 1 and 2 are the biggest containing 149 and 65 strains, respectively. Within these clusters the ncRNA deletion strains are separated into smaller sub-groups, sub-cluster 1 and sub-cluster 2, based on direction of fitness changes. The other clusters are relatively small (8 and 4 strains) and consist mainly of tRNAs and SUTs (S13 Table).
Cluster 1 encompasses strains with primarily specific response to temperature (Fig 4A). As shown in the heat map, this response to temperature is particularly evident in the initial pool (P) to batch (B) transition where, in any media considered, a change in fitness can be seen when the temperature is raised from 30°C to 36°C. Lowering the dosage of some ncRNAs either increases (Fig 4A, sub-cluster 1) or decreases (Fig 4A, sub-cluster 2) cell fitness with increasing temperature. The results suggest that ncRNAs in this cluster are involved in the optimal growth during heat stress and general nutrient deprivation. For example, SUT643 in sub-cluster 1 may have a function in transcriptional regulation of the neighbouring gene IME1, which is essential for meiosis, and is required for repression of HSP82 [52–54]. Our data indicate that lowering the dosage of SUT643 has a positive impact on yeast growth at high temperature, suggesting that the repression on HSP82 is partially lifted (the quantitative fitness profile for SUT643 is shown in S6A Fig).
Cluster 2 encompasses strains with specific response to growth phases, such as transition from P to B (batch phase with nutrient depletion) and from ESS and LSS stage (continuous culture phase with constant nutrients and pH), suggesting that ncRNA deletion strains in this cluster become important when nutrient levels are not constant (Fig 4B). In this case, lowering the dosage of some ncRNAs either increases (Fig 4B, sub-cluster 1) or decreases (Fig 4B, sub-cluster 2) cell fitness with nutrient depletions. When cells are about to reach stationary phase, there is a decline in overall transcriptional activities and several changes in cellular metabolism occur to store complex carbohydrates such as glycogen and trehalose [55–57]. Based on our data, sub-cluster 2 (Fig 4B) encompasses ncRNAs which are crucial for survival during the pool to batch stage. We found that some ncRNAs in this sub-cluster 2 are located next to genes that are highly correlated with transition to stationary phase. For example, SUT471 is located downstream of SNF11 and upstream of TPS2. TPS2 encodes for a phosphatase in the last step of the trehalose pathway, important for carbon storage and is activated sequentially after diauxic shift and is suppressed fully before entering stationary phase [58,59]. SNF11 encodes for a subunit of the SWI/SNF chromatin remodelling complex, which is involved in transcriptional regulation of several genes at the onset of stationary phase [60,61]. Another example is SUT509 which is located downstream of the medium chain fatty acyl-CoA synthetase gene FAA2 which has a transcriptional profile similar to that of the gene TPS2. The effect of SUT471 and SUT509 deletion on the neighbouring genes may, therefore, be responsible for their haplo-insufficiency in the pool to batch transition (the quantitative fitness profiles for SUT471 and SUT509 are shown in S6B and S6C Fig). We further analyzed all four clusters for representation of different ncRNA classes and found no bias in the distribution of SUTs or CUTs (FDR > 0.05). In addition, we could not identify any common biological functions using the gene ontology terms of neighbouring genes for the ncRNAs that comprise the different clusters.
To determine the influence of complete removal of a ncRNA on cell fitness we individually arrayed each strain of the haploid deletion collections on rich media (YPD) plates at 30°C and assessed colony size compared to the wild-type strain. The haploid deletion collections exhibited significant variation in fitness on YPD and this variation was detected across all types of ncRNA (Fig 5; S14 and S15 Tables). The deletion overlapping both SUT233 and CUT707 (Fig 2B), which displayed significant haplo-insufficiency as a diploid heterozygotic deletion in most conditions in the pool to batch growth (Fig 1), is the least fit in the haploid deletion collection (Fig 5). Deletion mutants of tL(CAA)A and SUT339 are respectively, the second and third least fit strains in the haploid collection on YPD media. The tL(CAA)A tRNA is part of a family of tRNAs for the leucine CAA codon and deletion of tL(CAA)A has previously been shown to significantly impair growth on YPD, whereas other members of this tRNA family do not display a severe fitness defect [62] (S14 Table). Our data support the idea that there are major and minor copies in tRNA gene families and loss of different members of a tRNA family affect cellular fitness differently [62]. From the genomic location of SUT339 there is no immediately obvious reason how its deletion is affecting fitness. The SNR75, tD(GUC)J3 and tE(UUC)B deletion strains are the top three fittest strains in this plate assay showing increased growth. SUT471 deletion also had a significantly positive effect on fitness in the haploid background (Fig 5). This positive effect on fitness is consistent with SUT471 being haplo-proficient in all of the continuous culture conditions (Fig 3), supporting the theory that SUT471 expression limits cell growth. ncRNA deletions showing little fitness change in the heterozygote background, but significant effects in the haploid background, have also been identified here (S15 Table). These data demonstrate that useful fitness data can be obtained from the plate array method of phenotyping on solid media. Moreover, some of the most dramatic phenotypes which were scored via colony size (Fig 5) are also seen in our continuous culture experiments (Figs 1 and 3). For example, the SUT004, SUT107, CUT356, SNR10 and tQ(UUG)L deletion mutants which were haplo-insufficient in at least one of the continuous culture conditions, also displayed significantly impaired fitness in the haploid fitness screen. Expanding this array method to a variety of other growth conditions should further our understanding of ncRNA function.
The heterozygote ncRNA deletion strains were induced to sporulate and the haploid spores dissected to reveal whether individual ncRNA gene deletion was essential for growth. Three percent of the ncRNA gene deletions (17 of 532) were found to be essential in nutrient rich conditions (YPD), with thirteen of these (i.e. snRNAs, snoRNAs, tRNAs) already known to be essential (S2 Table). Four novel essential ncRNAs were identified, SUT075, SUT367, SUT527 and SUT259/691, and were found to be essential in separate biological replicates of the deletion strains (S7 Fig). One of these essential ncRNAs, SUT527 (also annotated as RUF20), overlaps by 140 base pairs with the 3’ untranslated region (UTR) of the essential gene SEC4, a GTPase required for vesicle-mediated exocytic secretion and autophagy [63,64] (Fig 6A). To determine whether SUT527 essentiality was derived from its overlap with the 3’ UTR of the essential SEC4, two shorter deletions of SUT527 were constructed with 40bp overlap and no overlap with the SEC4 3’ UTR. The shorter SUT527 deletion, that still overlapped the SEC4 3’ UTR, resulted in a non-viable phenotype, whereas the strain containing a SUT527 deletion with no overlap with the SEC4 3’ UTR was viable (S8 Fig). This viability indicates that SUT527 essentiality is derived from the overlap with the 3’ UTR of the essential gene SEC4 and that deletion in this region does not generally cause silencing of SEC4 transcription. Transformation of the original SUT527 diploid deletion strain with a plasmid containing an approximately 1.4kb DNA fragment containing the SEC4 sequence known to complement SEC4 function [64] restored strain viability after sporulation and tetrad dissection. To understand whether the essential phenotype was caused by the deletion of the SEC4 3’ UTR in itself or caused by the interaction of the SUT527 RNA with the SEC4 3’ UTR, we reduced SUT527 expression in a haploid strain using a regulated Tet promoter [65]. We found that SEC4 mRNA expression was greatly decreased (Fig 6B) and SEC4 3’ UTR formation was affected when SUT527 expression was suppressed (Fig 6C). The SEC4 3’ UTR is required for localization of SEC4 mRNA [66]. Fluorescent in situ hybridization (FISH) revealed that SUT527 displayed a similar punctate localization to SEC4 mRNA and SEC4 mRNA was mislocalized when SUT527 expression was switched off (Fig 6D and 6E). Analysis of data sets from a global sequence analysis of small RNAs from S. cerevisiae strains engineered for RNAi to reveal the presence of dsRNAs [21], identified small RNAs produced from SUT527 in the region of overlap with the SEC4 3’ UTR (S9 Fig). The presence of these small RNAs in the region of overlap between SEC4 and SUT527 indicates that in vivo there is dsRNA formation between SUT527 and the SEC4 3’ UTR. In fact, FISH of SEC4 mRNA and SUT527 in the same cells with different coloured detection probes revealed that in cells approximately 10% of the SEC4 mRNA (red) and SUT527 RNA (green) puncta were found next to each other (Fig 6F). We have therefore defined a molecular function for the ncRNA SUT527 and suggest that the physical interaction between SUT527 and the 3’ UTR of SEC4 influences SEC4 3’ end formation and mRNA localization.
Of the four essential ncRNAs identified here (SUT527, SUT075, SUT367 and SUT259/691) SUT527, SUT075 and SUT367 are located adjacent to essential genes. Deletion of the ncRNA with the KanMX cassette could potentially remove essential regulatory elements for a nearby essential gene, or the expression from the KanMX module may influence the expression of a nearby essential gene. To determine the influence of deleting a single copy of SUT527, SUT075, SUT367 or SUT259/691 on the expression of nearby genes, qRT-PCR was used to analyze expression of nearby genes in the diploid deletion strains compared to the wild-type diploid strain. Analysis of the diploid SUT527 deletion strain revealed greatly reduced expression of SEC4 as expected (Fig 7A). Deletion of the overlapping SUT259/691 ncRNAs increased the expression of the upstream and downstream non-essential genes EMP46 and GAL2 which are both transcribed in the same direction as the KanMX (Fig 7B). SUT690 is located between EMP46 and SUT259/691. It is plausible that SUT690 might be the target of SUT259/691 regulation. However, analysis of SUT690 expression in the ΔSUT259/691 strain reveals that SUT690 expression levels are unchanged (S10 Fig). The deletion of either EMP46 or GAL2 alone does not result in a lethal phenotype, but since a double deletion mutant of EMP46 and GAL2 shows positive epistasis [67], it is possible that overexpression of both genes gives the opposite effect and hampers fitness (see Discussion). To test this hypothesis, we have cloned both EMP46 and GAL2 into the pBEVY-GA plasmid, containing a bi-directional GAL1/10 promoter. Overexpression plasmids with EMP46, GAL2 or both EMP46 and GAL2 were created and transformed into the BY4743 background strain. The comparative fitness of these overexpression strains were then examined using spot assays. Solitary overexpression of GAL2 or EMP46 was not lethal, however they resulted in impaired fitness, particularly the overexpression of GAL2 (Fig 8). The simultaneous overexpression of GAL2 and EMP46 resulted in no cell growth and is therefore lethal (Fig 8). This lack of growth supports the idea that the lethality, observed in the SUT259/691 knockout strain, is a result of a combined increase in EMP46 and GAL2 expression. The partial deletion of SUT075 caused a large decrease in the expression of the essential gene PRP3 which is transcribed in the opposite direction to SUT075 and the KanMX expression (Fig 7C). The decreased expression of the essential PRP3 may be the explanation for SUT075 lethality. Deletion of SUT367 caused an increase in the expression of the essential gene RPL3 which is transcribed in the same direction downstream of SUT367 (Fig 7D). Interestingly, RPL3 is one of the few ribosomal protein genes in yeast that is neither duplicated nor contains an intron, both properties that are associated with increased ribosomal protein gene expression [68]. Therefore, increased expression of RPL3 may be detrimental to cells providing a reason for SUT367 lethality (see Discussion). Overall, we observed that deletion of ncRNAs can both positively and negatively influence the expression of nearby genes and that in some cases can explain the lethality.
To investigate further the essentiality of ncRNAs SUT527, SUT075, SUT367 and SUT259/691, we overexpressed these SUTs from plasmids to discover any that could recover the lethal phenotype of the corresponding deletion strain and identify trans ncRNA effects. We constructed centromeric plasmids with each SUT expressed in either the sense or antisense orientation from the GAL1 promoter. These plasmids were then transformed into the corresponding heterozygote diploid deletion strains. These strains were then sporulated and tetrads dissected. Successful generation of viable haploid strains, containing the deleted essential SUT, would indicate the ability of the overexpressed SUT to function in trans. The SUT367, SUT527 and SUT259/691 sense or antisense plasmids were unable to reverse lethality of the corresponding ncRNA deletion in the haploid background following sporulation and tetrad dissection. However, deletion of SUT075 was no longer lethal when the sense orientation SUT075 plasmid was present (Fig 9A and 9B) but was still lethal with the antisense SUT075. This complementation of the SUT075 deletion strain lethal phenotype by the ectopic expression of SUT075 RNA indicates that SUT075 functions in trans. It is, therefore, plausible that SUT075 regulates expression of distal genes in the genome, not just the neighbouring PRP3 gene. The deletion of one copy of SUT075 in the diploid background (Fig 7C) significantly reduced expression of the adjacent PRP3 gene. Expression of PRP3 was measured in the heterozygote diploid SUT075 deletion strain with the sense SUT075 plasmid expressing the SUT075 ncRNA, to determine if the rescue of the lethal phenotype in the haploid progeny (Fig 9A and 9B) was the result of PRP3 expression levels being returned to normal. PRP3 expression was found to be 8.3 fold greater in the heterozygote diploid SUT075 deletion strain, when the SUT075 expression plasmid was present (Fig 9C). Recovery of PRP3 expression, to levels greater than in wild type cells, suggests that the GAL1 promoter is stronger than the native SUT075 promoter. Overall, these data confirm that expression of SUT075 from a plasmid is able to up-regulate PRP3 expression in trans and reverse lethality in strains deleted for SUT075.
To ease the use and access of our extensive functional fitness data for future research, we have built a publicly accessible online resource called the Yeast ncRNA Analysis (YNCA) (http://sgjlab.org/ynca/) to host the heterozygote and haploid deletion fitness profiles for each of the deleted ncRNAs. Data are searchable by neighbouring genomic features, ncRNA type, essentiality, chromosomal position and growth phenotype for each growth medium used, as well as searchable by ncRNA name as classified in Xu et al. (2009) [9]. The types of searchable neighbouring genomic features are known open reading frames, tRNA genes, snoRNA genes, centromeric and telomeric regions, autonomous replicating sequences (ARS), long terminal repeats (LTR), pseudogenes, LTR retrotransposons and transposon internal genes. The user can download both raw experimental values and statistical significance values from a results table specific to the search performed. The list of barcode TAGs associated with each strain is also available on the website. We plan to progressively expand the YNCA database to include homozygote deletion strain fitness data under a variety of conditions and results from future analyses.
By utilizing the newly developed ncRNA deletion strain collections in the yeast Saccharomyces cerevisiae we have carried out large scale profiling of ncRNA function under a variety of growth conditions and phases. The extensive functional fitness data can be accessed via the database YNCA (http://sgjlab.org/ynca/) where the influence of individual ncRNA deletion strains on cellular fitness has been catalogued in an easy to navigate and searchable website. This large scale functional profiling has now provided valuable functional information on the deletion of 532 different ncRNAs which includes tRNAs, snoRNAs, snRNAs, SUTs, CUTs and various other annotated ncRNAs. We have also investigated in more detail four novel essential ncRNAs and determined the mechanisms by which they result in a lethal phenotype when deleted.
Yeast strains deleted for individual tRNA genes have been previously constructed with these deletion strains tested for both growth rate and growth yield under a number of conditions [62]. While there is significant overlap between our collection and that of Bloom-Ackermann et. al., there are strains that are unique to each collection (S16 Table). Where there is overlap between collections we have observed similar growth phenotypes of tRNA deletion strains. For example, our observation that the tR(CCU)J deletion strain displays decreased fitness in all the conditions we tested during the pool to batch growth phase was also observed for the growth rate of the tR(CCU)J deletion strain in four of the six growth conditions used by Bloom-Ackermann et al [62]. Within tRNA families major and minor tRNAs have been identified where deletion of the major tRNA influences the ability of a deletion strain to grow under different conditions more than one of the minor tRNAs in the same family [62]. For instance, the tRNAs tR(UCU)E and tR(UCU)M2 were identified as being major tRNAs that are influenced the most by different growth conditions in their family [62]. In the six conditions tested here we have also found that the tR(UCU)E deletion displays decreased fitness in all six conditions in the pool to batch transition. However, we have identified tR(UCU)B in the same family, a deletion novel to our collection, that also displays decreased fitness in all six conditions in the pool to batch transition and tR(UCU)G1 as less fit in four of the six conditions in the pool to batch transition (S4–S10 Tables). In contrast, we did not observe a significant decrease in fitness with deletion of tR(UCU)M2 in any of the six conditions in the pool to batch transition. Analysis of the tRNA deletion strains in the tR(UCU) family under continuous growth conditions did not identify any tRNAs in the tR(UCU) family that displayed a consistently significant decrease in fitness, when deleted, in any of the conditions we tested. These results suggest that tRNA levels are more important in conditions where nutrients become limited.
By using continuous growth conditions, we have uncovered additional phenotypes for ncRNA deletion strains that are not observed under growth conditions where nutrients become limiting. Specifically, we have observed changes in fitness associated with temperature changes that were not observed in the pool to batch growth phase. Additionally, phenotypes observed in the pool to batch stage where nutrients are limited were not observed in continuous culture. By observing the fitness of the deletions strains by two distinct methods of cell culture we have produced an extensive catalog of fitness data for each of the heterozygous diploid deletion strains. Combined with our analysis of the haploid deletion collections arrayed on solid media, overall we present the most extensive analysis of ncRNA requirements for cellular fitness to date. These data have been compiled into a database called YNCA. YNCA uses the more sensitive ESS-LSS fitness change for search based on the heterozygote fitness profile, but displays both pool-batch and ESS-LSS fold change values in the detailed ncRNA-specific page. Both sets of data are downloadable. For each growth medium, the user can retrieve strains which display any or one selected fitness phenotype (after statistical analysis; gain/loss of fitness or haplo-proficient/insufficient) or obtain a list of all strains with available experimental data, for raw data download. Given that only a few examples of CUTs and SUTs have a known function and that the analysis of ncRNA function sometimes focuses on the regulation of neighbouring genes, the YNCA website offers the option to search by nearby genomic feature, hence facilitating the selection of candidate ncRNAs as gene-specific or feature-specific regulators.
In the construction and analysis of the ncRNA deletion strains we have identified ncRNAs that are essential for cell growth. Many of these essential ncRNAs have been previously identified and annotated as essential ncRNAs. The five snRNAs (U1, U2, U4, U5 and U6) required for pre-mRNA splicing, the snoRNAs snR128 (U14) and snR30 (U17), the RNA component of RNase MRP (NME1), the RNA component of nuclear RNase P (RPR1) and tRNAs tR(CCG)L, tR(CCU)J, tS(CGA)C and tT(CGU)K have all been previously shown to be essential. Besides the known essential ncRNAs we have identified four novel ncRNAs that are essential when deleted. These four novel essential ncRNAs are SUT075, SUT367, SUT527 and SUT259/691. The essentiality of SUT527 is caused by its overlap with the 3’ UTR of the essential protein coding gene SEC4, as making smaller deletions that did not overlap the annotated 3’ UTR of SEC4 did not result in a lethal phenotype. It appears that the overlap of SUT527 with the 3’ UTR of SEC4 is required for both the stability of the SEC4 mRNA and for the localization of SEC4 mRNA. SEC4 mRNA localization is determined by its 3’ UTR [66]. There is evidence that SEC4 3’ UTR/SUT527 RNA duplexes are formed within cells [21] and we have observed that the SEC4 mRNA and SUT527 RNA localize in close proximity. A cytoplasmic function for other SUTs is very likely as a proportion of SUTs are transported to the cytoplasm where they have been proposed to exert their function [31].
The ncRNA SUT367 was found to be essential when deleted, but analysis of the nearby essential gene RPL3 in the diploid heterozygous deletion strain revealed that RPL3 expression is increased (Fig 7D). Large scale screens have previously identified that overexpression of RPL3 causes growth impairment, disrupts the cell cycle [69] and induces chromosome instability [70]. The mechanisms by which deletion of SUT367 leads to RPL3 overexpression or how overexpression of a ribosomal protein gene leads to chromosome instability/cell cycle disruption and lethality is not clear. However, other ribosomal protein genes have also been identified to cause chromosome instability/cell cycle disruption leading to cell lethality when overexpressed [69,70]. We show that deletion of SUT367 prevents spores from germinating after meiosis, and it is plausible that the resulting overexpression of RPL3 is responsible for the inability of the spores to grow.
A deletion of SUT259/691 is lethal and this deletion results in the overexpression of two adjacent nonessential protein coding genes EMP46 and GAL2, but not the adjacent SUT690, in the diploid heterozygote (Fig 7B, S10 Fig). Individual overexpression of either EMP46 or GAL2 displays a slow growth phenotype on their own ([71] and Fig 8 of this a manuscript). When EMP46 and GAL2 are overexpressed simultaneously the cells are unable to grow (Fig 8). Therefore, SUT259/691 are essential for the regulation of EMP46 and GAL2, and when deleted cause an overexpression of these genes which causes lethality.
The SUT075 is expressed in the opposite direction to the essential gene PRP3 with the deletion we made of SUT075 being 230 nucleotides away from the start of PRP3 and 143 nucleotides away from the stop codon of the non-essential gene JIP4 (Fig 7C). We successfully used complementation to determine that expressing the full length SUT075 RNA from a plasmid in trans could rescue the essential phenotype of SUT075 deletion. Therefore, we have identified another example of a ncRNA that works in trans. The action of SUT075 is in part locally as the trans expression of SUT075 increases the expression of the adjacent essential gene PRP3, but there is also the possibility that SUT075 acts elsewhere in the cell. As transcription of SUT075 produces an RNA that works in trans we investigated yeast ribosome profiling data and found that SUT075 does not associate with ribosomes so is unlikely to be translated into protein [34]. To date only a few examples of ncRNAs working in trans have been identified. The Ty1 RTL CUT ncRNA has been found to regulate Ty1 expression in trans [72] and the ncRNA PHO84 can work in trans to silence genes [73]. Recent work has found two new trans acting SUTs suggesting that trans acting ncRNAs may be more prevalent than previously thought [30]. Our identification of SUT075 as another trans acting ncRNA supports the view that there may be more trans acting ncRNAs identified in the future. Overall, analysis of these ncRNAs, initially identified as being essential, has revealed that the transcription of ncRNA can both positively and negatively influence the expression of adjacent genes or produce an RNA that can function on its own, indicating that regions of the genome identified as producing SUTs and CUTs are functional and do not represent just transcriptional noise.
By exploiting the yeast ncRNA deletion collections we produced a large array of phenotypic data which is a useful resource for providing a snapshot of ncRNA function in the cell. By expanding the number of conditions investigated it is hoped that a picture can be built of how ncRNAs contribute to the fitness of cells in different environments. Here by exploring individual examples of ncRNA we have determined the molecular function of SUT527 and also showed that SUT075 works in trans, expanding the repertoire of cellular functions that require ncRNAs. As the characterization of the numerous ncRNAs continues, the use of ncRNA deletion collections in large-scale functional and interaction studies will ultimately provide information on how ncRNAs fit into the functional framework of the cell.
All S. cerevisiae strains and primers used are listed in S17 Table and S18 Table. Individual deletion strains or the collection of deletions strains are available on request.
Methods for the construction of the deletion strain collections have been previously described [35]. In preparation for chemostat continuous culturing of the heterozygote collection, a pool of the deletion strains was prepared. A -80°C stock of heterozygous ncRNA deletion strains were grown overnight at 30°C in liquid YPD and the OD600 of each strain in the microtitre plate was read using a FLUOstar OPTIMA plate reader (BMG Labtech). Subsequently each strain was normalised to an OD600 of 0.1 and pooled.
The diploid heterozygote ncRNA deletion collection pool was grown in chemically defined F1 medium limited for glucose (carbon limitation) or nitrogen at 30°C and 36°C. The pooled heterozygous deletion strains were also grown in F1 medium limited for glucose or nitrogen at 30°C in the presence of 100mM LiCl. The diploid heterozygote ncRNA deletion collection pool was grown in batch culture for 24hrs and then switched to continuous culture where it took about 42 hrs to reach steady state. Steady state growth conditions were maintained for 30 generations. Samples were taken at the Pool (P), Batch (B), Early Steady State (ESS, 48 hours after switching to continuous culture), Mid Steady State (MSS, after 20 generations) and Late Steady State (LSS, after 30 generations) stages then processed for Illumina sequencing of the barcodes to determine the abundance of each strain. Two biological repeats were carried out for each condition. Details of growth medium and continuous culture in chemostats are as previously described [36].
Genomic DNA was isolated from samples using the Wizard Genomic DNA Purification Kit (Promega) according to the manufacturer’s protocol. UPTAGs and DOWNTAGs were amplified with primers compatible with multiplexed Illumina sequencing. For the UPTAGs the forward primer was 5’AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGCTCTTCCGATCTGATGTCCACGAGGTCTCT and the reverse primer was 5’CAAGCAGAAGACGGCATACGAGATNNNNNNGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTGTCGACCTGCAGCGTACG. For the DOWNTAGS the forward primer was 5’-AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCTCGAGCTCGAATTCATCGAT and the reverse primer was 5’CAAGCAGAAGACGGCATACGAGATNNNNNNGTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTCGGTGTCGGTCTCGTAG. The NNNNNN represents the 6-mer indexing tag used for multiplexing the different samples. Amplified TAGs were quantified with the KAPA library quantification kit (KAPA Biosysytems) and 10nM of the TAG libraries was used for Illumina sequencing.
Sequenced reads were trimmed to contain just the TAG sequence using Trimmomatic [74]. Trimmed reads were mapped to a database of the TAG sequences using Bowtie2 [75]. A TAG was deemed to be identified if the trimmed sequenced read aligned to the full length of that TAG with a maximum of 1 mismatch. Summed counts for each of the two TAGS for a deletion strain were used as input for DESeq [76]. The Log2 fold change was determined between different growth stages and the changes with a p value of < 0.05 and 1.5 fold change were identified.
Two primers, RUF20_P1 and RUF20_P2 (S18 Table) were designed to amplify the KanMX-TetO7 cassette from plasmid pCM325 [65]. The resulting RUF20-KanMX-TetO7 cassette was transformed into the strain CML476 [65]. The two ends of the cassette were homologous to the start and 500bp upstream of the SUT527 gene to replace 500bp of the SUT527 promoter. The successful transformants were selected on 200μg/ml YPD-G418 plates, confirmed by PCR and the resulting strain was named CML/RUF20/tetO7.
For RT-PCR, mRNA was isolated from 200μg yeast total RNA prepared by the hot phenol method [77] with the SIGMA GenElute mRNA Miniprep kit according to the manufacturers protocol and eluted in 100μl of elution buffer. The polyadenylated mRNA sample was treated with 10 units of RQ1 DNase (Promega) in 1X DNase Buffer (Promega) and 200 units RNasin (Promega) at 30°C for 30 minutes. The reaction was then stopped by adding 2mM of EDTA and incubation at 65°C for 10 minutes. An equal amount of citrate buffered phenol (pH 5.3) was added followed by vortexing for 1 minute and centrifugation at 15600g for 2 minutes. The polyadenylated mRNA was then precipitated from the aqueous phase by adding 0.1 volume of 3M sodium acetate (pH 5.3), 2.5 volumes of 100% ethanol and 10μg of glycogen. The sample was precipitated at -20°C for 30 minutes. The RNA was collected by centrifugation for 5 minutes and washed with 96% ethanol, pelleted again at 15600g and air dried. The pelleted sample was resuspended in 16.25μl of water to be used for the first strand cDNA synthesis using the OneTaq RT-PCR Kit (New England Biolabs). The procedure for cDNA synthesis and PCR amplification was based on the manufacturer’s instructions. Primers for primer walking of the SEC4 and TUB2 coding sequence and 3’ UTR are listed in S18 Table.
To determine the expression of genes, cells were grown to an OD600 of 0.5 prior to RNA extraction using the Qiagen RNeasy Mini Kit. RNA concentrations were determined with a NanoDrop Lite Spectrophotometer. The GoScript Reverse Transcriptase was used for cDNA synthesis with 200ng of RNA. Quantitative RT-PCR was performed on the cDNA using iTaq universal SYBR green Supermix (BioRad) in a CFX Connect Real-time PCR Detection System (BioRad). qPCR cycling conditions were as follows: initial denaturation 95°C for 3 mins; 35 cycles of 95°C for 45 secs, 58°C for 45 secs and 72°C for 3 mins; final extension of 72°C for 5 mins. ACT1 was used as a reference gene. The Ct values were used to measure the expression of each gene according to the 2-ΔCt method [78]. Sequences for the oligonucleotides used can be found in the S18 Table. Using the ΔΔCт method and ACT1 as a reference gene, the fold change (2^) in expression, relative to the wild-type was calculated. Error bars are calculated using each of the three independent biological samples. P values were calculated using the Welch two sample t-test.
The open reading frame plus 500bp upstream and 250bp downstream of SEC4 (129943–131331), which contains the approximately 1.4kb BamHI/EcoRI fragment that complements SEC4 function [64] was amplified with Phusion DNA polymerase (New England Biolabs) and primers SEC4F-Bam and SEC4B-Eco (S18 Table). The open reading frame plus 500bp upstream and 500bp downstream of SUT527 (13146–14586) was amplified with Phusion DNA polymerase and primers RUF20F-Bam and RUF20B-Xba (S18 Table). PCR products were then cloned into pRS413 to produce plasmids pRS413-SEC4 and pRS413-RUF20. The correct SEC4 and SUT527 (RUF20) sequences were confirmed by sequencing.
Plasmids pRS413-SEC4 and pRS413-RUF20 were used as templates for production of transcription templates for SEC4 and SUT527 (RUF20) probes by PCR for digoxigenin or dinitrophenol labelling using primer pairs SEC4T7/SEC4B_prob and RUF20FT7/RUF20B-prob (S18 Table). Digoxigenin and dinitrophenol labelled probes were made using 1μg of purified SEC4 or SUT527 (RUF20) PCR template with 1X DIG RNA labelling mix (Roche) or an identical RNA labelling mix containing DNP-11-UTP in place of DIG-11-UTP in a transcription reaction at 37°C for 2 hours using T7 RNA polymerase (Promega) according to the manufacturer’s instructions. Two units of RQ1 DNase (Promega) were then added and the mixture incubated at 37°C for 15 minutes. The sample was then purified using the Qiagen RNA Easy kit following the manufacturer’s instructions. The RNA concentration was measured and 10μg of RNA probe was used for the hybridization step.
Coverslips No.1 glass 22mm X 22mm (Fisher) were boiled for 30min in 250ml water with 0.1N HCl. Cover slips were then rinsed 10X with deionised water and stored in 70% ethanol. Flamed coverslips were coated with 200μl of 1X poly-L-lysine solution (Sigma) for 2min then excess poly-L-lysine removed and the coverslips air-dried. Coverslips were washed three times with 250μl of water for 10 minutes and air dried. Slides were stored in single wells of a six-well tissue culture dish at room temperature after air drying.
For cell fixation, cells were grown at 30°C in 50ml YPD with our without doxycycline (600μg/mL) to OD600nm = 0.5 and fixed in 4% formaldehyde (Sigma) for 45min at room temperature. Cells were then centrifuged at 2,400g for 5min at 4°C then resuspended in 1ml buffer B (16mM KH2PO4, 83mM K2HPO4, 5.4% Sorbitol). Cells were then washed three times with buffer B. Washed cells were resuspended in 1ml freshly-prepared spheroblast buffer (Buffer B with 20mM Vanadyl Ribonucleoside Complex (NEB), 250 units lyticase and 0.002% β-mercaptoethanol) and incubated at 30°C for 15 minutes. Cells were washed twice with 1ml ice cold buffer B and spun at low speed 2000g for 1 minute. Cells were resuspended in 1ml buffer B and 150μl of the cells were placed on coated coverslips and incubated at 4°C for 30 minutes to allow adherence of the cells to the coverslips. Cells were then washed with 5ml ice cold Buffer B and 5ml of 70% ethanol was added, cells were then stored at -20°C.
The stored coverslips were immersed in 1ml of the hybridization mix (50% formamide, 5X SSC, 1mg/ml yeast tRNA, 100μg/ml heparin, 1X Denhardts, 0.1% Tween 20, 0.1% CHAPS, 5mM EDTA) in a six-well tissue culture dish. The dish was then sealed with parafilm and incubated at 50°C for 1 hour. Next, the hybridization mix was removed and another 2ml of the hybridization mix was added with 10μg probe (either DIG-labelled probe alone for SEC4 or SUT527/RUF20 for single detection or DIG-labelled probe for SEC4 and DNP-labelled probe for SUT527/RUF20 for colocalization) then incubated overnight at 50°C. Coverslips were washed with 2ml 0.2X SSC three times. Then 2ml of blocking buffer (1X PBS, 0.1% TritonX-100 and 10% horse serum) was added to the coverslips and incubated at room temperature for 1 hour. For single localization of DIG-labelled probes coverslips were incubated for 2 hrs with HRP conjugated anti-digoxigenin monoclonal antibody (Jackson Immuno Research) diluted to 1:500 with 250μl blocking buffer. Coverslips were then washed three times with 1ml blocking buffer and incubated for 2 hours with Alexa Fluor 488-conjugated anti-HRP antibody (Jackson Immuno Research) diluted 1:100 with 250μl blocking buffer. For combined co-localization detection of DIG- and DNP-labelled probes coverslips were incubated for 2 hrs with goat anti-DIG antibodies (Vector Laboratories) and rabbit anti-DNP (Vector Laboratories) diluted to 1:500 with 250μl blocking buffer. Coverslips were then washed three times with 1ml blocking buffer and incubated for 2 hours with mouse anti-rabbit Alexa Fluor 488 antibody (Jackson Immuno Research) and mouse anti-goat Alexa Fluor 647 antibody (Jackson Immuno Research) each diluted 1:100 with 250μl blocking buffer. Coverslips were then washed three times with 1ml blocking buffer and the coverslips were placed on a slide with a drop of ProLong Gold antifade reagent with DAPI (Molecular Probes by Life Technologies) and allowed to set.
For single localization slides were visualised with a Nikon Eclipse E600 microscope using a 100x/0.5–1.3 NA differential interference contrast oil Iris Apo objective. The images were captured using a Nikon DS-QilMc camera and NIS-Elements BR 3.2 software. To obtain the quantitative data on RNA localisation in each strain, 100 cells were scored and analyzed for the localization on whether RNA signals were localized to the cell membrane or not. Cells were scored from three technical repeats. For colocalization images were collected on a Zeiss Axioimager.D2 upright microscope using an Olympus UPlanFL 100x/1.30 Oil Ph3 0.17 objective and captured using a Coolsnap HQ2 camera (Photometrics) through Micromanager software v1.4.23. Specific band pass filter sets were used to prevent bleed through from one channel to the next. Images were then processed and analyzed using Image J.
In order to investigate the growth effects of the ncRNA deletions, strains were grown under rich (YPD) and minimal (chemically defined F1 with carbon or nitrogen limitations) media conditions at 30°C. F1 medium was prepared in accordance to Delneri [36]. Carbon and Nitrogen limited F1 media were modified to contain 0.25% glucose (w/v) and 0.46 g/liter (NH4)2SO4, respectively. Growth measurements at OD595 were recorded using a BMG FLUOstar OPTIMA Microplate Reader, as previously described by Naseeb and Delneri [79] for up to 70hr incubation time. Cells were grown at 30°C from an OD600 0.1 and readings taken every 5min. Three technical replicates of three independent biological samples were used for each deletion mutant strain and six technical replicates for the wild type strain. Graphs were produced using the grofit package of the R program. Area under curve (AUC) measurements for the tA(UGC)O, SUT340, CUT873 and tT(AGU)J deletion mutants were calculated as per Norris et al [80], using the grofit::gcFitSpline R package.
To account for plate and batch effects, two biological replicates (MATa and MATα) and four technical replicates of each haploid deletion mutant strain were prepared. Three technical replicates of each plate were performed. Strains were removed from -80°C storage and grown to saturation at 30°C in YPD, in 384 well microtitre plates. Using a Singer Rotor HDA, the 384 well cell cultures were stamped onto YPD plates and incubated at 30°C for 2 days. Plates were then imaged using a Bio-Rad Gel Doc XR system and images processed using SGAtools [81]. The average of the normalized colony size values for replicates of each biological were then combined and used for analysis. We assumed normal distribution on the dataset and used the standard EM algorithm to determine means and standard deviations from the mixture of strains with normal growth and others with reduced fitness using Mixtools [82]. The P values were calculated from the parameters that are closer to the wild-type and fitness differences considered significant with p < 0.05.
Sequencing data were normalized and converted to Log2 fold change to allow comparison between the pool and batch stage and between the early and late steady state across different media and temperatures using DESeq2 [83]. To classify the deletion strains based on the impact of growing conditions, we applied generalized linear model with normal approximation and selected those with significant response to our testing variables (P value ≤ 0.05). As a result, fitness profiles were simplified and clustered using the ad hoc partitioning around medroid method implemented in the R package cluster [84]. Finally, data analysis was conducted to evaluate enrichment of SUT/CUT using exact binomial test. False discovery rate (FDR) was calculated using R. The biological functions of neighboring genes to ncRNAs in each cluster were identified using GO Term Finder in SGD.
The PGAL1 promoter was amplified from pAV1901 [85] using Gal1.for and Gal1.rev primers and cloned into the SalI site in pRS416 [86]. Next the CYC1 terminator was amplified from p426-GPD [87] using Cyc1.for and Cyc1.rev primers and cloned into the BamHI site creating pRS416Gal1Cyc1. The sense and antisense ncRNA expression plasmids were created using the primer pairs in S18 Table. Phusion DNA polymerase (New England Biolabs) was used in all amplifications according to the manufacturers protocol using yeast genomic DNA from BY4742 as template. All sense and antisense ncRNA expression plasmids were cloned into pRS416Gal1Cyc1 via the HindIII restriction site using the Gibson cloning technique [88]. All constructs were verified by sequencing.
Each ncRNA overexpression plasmid was transformed into the corresponding heterozygote diploid deletion strain and wild-type BY4743. Cells containing the overexpression plasmids were selected for on SD media lacking uracil (0.67% Bacto yeast nitrogen base without amino acids, 2% glucose, 2% agar, 0.192% Yeast synthetic drop-out medium supplement without uracil). Strains were then sporulated in liquid sporulation medium lacking uracil (1% potassium acetate, 0.005% zinc acetate, 0.002% histidine and 0.003% leucine). Cultures were incubated for 5 days at 25°C followed by three days incubation at 30°C. Tetrad dissection was performed on SD media plates containing 2% Galactose and lacking uracil, using a Singer instruments MSM 400 microdissector. After 4 days incubation at 30°C, tetrad dissection plates were replica plated on to SD media containing 2% Galactose, 300mg/L G418 and lacking uracil. Haploids growing on the final plates were considered to contain the original ncRNA deletion cassette and the ncRNA overexpression plasmid.
DNA was extracted (QIAamp DNA Mini Kit) from these haploids for PCR confirmation. The presence of the ncRNA overexpression plasmid was confirmed using universal pRS416 primers (pRS416 F Primer ‘CATGGAGGGCACAGTTAAGC’ and pRS416 R Primer ‘ACCACATCATCCACGGTTCT’). Deletion of SUT075 was confirmed using a primer specific to the kanamycin cassette (kanC3 ‘CCTCGACATCATCTGCCCAGAT’) and a primer flanking the insertion site (SUT075 confD ‘TGCAGGGAACAGATTTTAGATTT’). PCR reaction mix contained: 0.5μM of each primer, 100ng of DNA template, 12.5μl MyTaq Red Mix (Bioline) and water to 25μl. Cycling conditions: initial denaturation at 95°C for 10min followed by 35 cycles of 95°C for 30sec; 57°C for 30sec; 72°C for 90sec and a final elongation of 72°C for 5min. PCR products were run on a 1.5% agarose gel.
EMP46 and GAL2 overexpression plasmids were constructed following the same methodology as the ncRNA overexpression above, with a few adjustments. The pBEVY-GA plasmid, containing a bi-directional GAL1/10 promoter, was used [89]. EMP46 was inserted at the upstream site via the BamHI site and GAL2 was inserted at the downstream site via the XmaI site. Three plasmids were constructed containing: 1) GAL2; 2) EMP46 or 3) GAL2 and EMP46. These plasmids were transformed separately into BY4743 and selected on SD media lacking uracil (as above). Cultures of these overexpression strains were then serially diluted tenfold and stamped (using the spot assay function of Singer Instrument’s ROTOR) onto SD media containing either 2% Galactose (promoter activate) or 2% Glucose (promoter inactivate). Plates were then imaged using a Bio-Rad Gel Doc XR system.
Cultures of BY4743 (+empty pRS416), ΔSUT075 (+empty pRS416) and the ΔSUT075 (+ sense SUT075 recovery plasmid) heterozygote diploid strains were grown to an OD600 0.5 in liquid SD media containing 2% galactose and lacking Uracil. Three biological replicates of each strain were cultured. RNA was extracted and qRT-PCR was performed using the PRP3 forward and reverse primers (S18 Table), following the methods previously described (Quantitative Real Time-PCR).
YNCA was developed in RStudio [90], version 1.0.143, with the use of the packages shiny [91] and rmarkdown [92]. Local server hosting relies on the open source version of Shiny-server. The underlying server-side data processing is written in R [93], version 3.4.0. Lists and positions of chromosomal features in S. cerevisiae are taken from the Saccharomyces Genome Database (www.yeastgenome.org). The type of features included are: known opening reading frames, tRNA genes, snoRNA genes, centromeric and telomeric regions, autonomous replicating sequences (ARS), long terminal repeats (LTR), pseudogenes, LTR retrotransposons and transposon internal genes.
|
10.1371/journal.pntd.0001799 | The Interactive Roles of Aedes aegypti Super-Production and Human Density in Dengue Transmission | A. aegypti production and human density may vary considerably in dengue endemic areas. Understanding how interactions between these factors influence the risk of transmission could improve the effectiveness of the allocation of vector control resources. To evaluate the combined impacts of variation in A. aegypti production and human density we integrated field data with simulation modeling.
Using data from seven censuses of A. aegypti pupae (2007–2009) and from demographic surveys, we developed an agent-based transmission model of the dengue transmission cycle across houses in 16 dengue-endemic urban ‘patches’ (1–3 city blocks each) of Armenia, Colombia. Our field data showed that 92% of pupae concentrated in only 5% of houses, defined as super-producers. Average secondary infections (R0) depended on infrequent, but highly explosive transmission events. These super-spreading events occurred almost exclusively when the introduced infectious person infected mosquitoes that were produced in super-productive containers. Increased human density favored R0, and when the likelihood of human introduction of virus was incorporated into risk, a strong interaction arose between vector production and human density. Simulated intervention of super-productive containers was substantially more effective in reducing dengue risk at higher human densities.
These results show significant interactions between human population density and the natural regulatory pattern of A. aegypti in the dynamics of dengue transmission. The large epidemiological significance of super-productive containers suggests that they have the potential to influence dengue viral adaptation to mosquitoes. Human population density plays a major role in dengue transmission, due to its potential impact on human-A. aegypti contact, both within a person's home and when visiting others. The large variation in population density within typical dengue endemic cities suggests that it should be a major consideration in dengue control policy.
| In the urban dengue system the life history of the mosquito vector, Aedes aegypti, transpires mainly inside and around human residences. In this study we integrated field data from an endemic city of Colombia into a simulation model to assess how natural variation in A. aegypti production and household human density influence dengue transmission. Contrary to traditional models, we show that the basic reproductive rate of dengue (Ro) is more likely to be positively correlated with human density. Moreover, the natural regulatory pattern of A. aegypti production, where a few super-productive houses dominate vector recruitment, caused a "super-spreading" pattern, whereby the large majority of viral introductions did not generate secondary infections, and Ro depended on sporadic, highly explosive transmission events. These events were dependent on the introduced infectious human infecting mosquitoes produced in super-productive vessels. When the likelihood of human introduction was incorporated into our risk indicator, a significant interaction emerged between human density and A. aegypti super production, such that removal of these containers had a much larger impact on reducing dengue in areas of higher human density. These results show that knowledge of interactions between human population density, social interactions and the natural regulatory pattern of A. aegypti can improve the design of dengue control efforts.
| In the latter half of the 20th century dengue emerged as the most prevalent urban vector borne disease of humans, readily propagating among urban populations of humans and Aedes aegypti mosquitoes. Intervention against A. aegypti domestic container habitats, known as source reduction, is central to the dengue prevention activities of most health departments of endemic cities [1]–[4]. However, few programs have the resources necessary to intervene effectively in all areas infested by A. aegypti [5]. Therefore, dengue prevention could benefit from an understanding of the areas in which the impact of source reduction would be maximized.
Most evidence indicates that urban A. aegypti populations are regulated by mortality that occurs in the egg/larval stages [6]–[9]. This results in the common finding that most infested containers produce few pupae, whereas the majority of the adult vector population derives from only a few containers and houses, called super-producers [4], [8], [10], [11]. The elimination of super-productive containers forms the conceptual basis for targeting source reduction programs [2], [3], [12]. This approach is grounded in modeling studies that show that the targeted elimination of containers above a threshold pupal abundance can significantly reduce the risk of dengue [13], [14]. However, this consistent regulatory pattern of A. aegypti also causes the majority of vectors to emerge in the same location, generating significant spatial heterogeneity in adult vector distributions [15]–[18]. However, because most models of the impact of source reduction on dengue assume homogenous mixing between humans and mosquitoes, little is known about how the phenomenon super-production per se affects transmission dynamics. This knowledge is critical to understanding how the natural regulation of A. aegypti influences the dynamics of dengue.
Traditionally, field estimates of the entomological risk of mosquito-borne disease have focused on the ratio of vectors to humans, in order to estimate the rate at which humans receive infectious bites [19]. This rationale has been used to assess the entomological risk of dengue through surveys of human and pupal abundance in order to estimate the metric A. aegypti pupae per person [14]. Unlike other mosquito-human systems A. aegypti rests, feeds and oviposits largely inside houses, generating a close physical proximity to humans. This co-habitation likely also explains why A. aegypti, exhibits adaptations such as asynchronous ovarian development and the preferential use of human blood rather than sugar as an energy sourcefor reproduction [20]–[22]. Consequently, female A. aegypti repeatedly bite humans even when seeking to lay eggs. This differs from other mosquito disease systems with a larger distance between blood-seeking and oviposition sites, in which host-vector contact is strongly determined by the time required for the mosquito to process the blood meal, oviposit and once again encounter hosts [19]. Greater human density may therefore increase the number of humans that each A. aegypti encounters, a process that is overlooked when risk is measured by the vector to host ratio. Because few targeted source reduction programs consider heterogeneities in human density when determining geographic targets, improved knowledge of the combined epidemiological impacts of super-productive vessels and human density can play an important role in optimizing dengue surveillance and control.
A second important feature of the urban dengue system is the capacity of human movement across urban areas to propagate infection spatially [23]–[26]. As human density increases, so do the number of different people that a person encounters, increasing both the number of mosquitoes that potentially bite a person and the number of unique people that a mosquito may bite. Moreover, because each person has a unique social contact network, greater human density in a particular area will increase the frequency of dengue infected visitors. Therefore, human density may influence the risk of dengue epidemics in a given area by affecting both the average number of secondary cases (Ro) and the frequency of viral introduction. As our measure of risk, we used the average secondary human infection rate for a given per-capita rate of dengue introduction into an immunologically naïve population [27], an index that we term epidemic potential. Because both the rates of secondary infection and viral introduction may increase with human density, epidemic potential captures the human density-dependence of dengue risk better than R0 alone.
Human and mosquito population sizes influence dengue transmission by two distinct processes: (1) human transmission to mosquitoes and (2) mosquito transmission to humans [28]. The first depends on the number of unique mosquitoes that bite each person, whereas the second is determined by the number of unique people that each infected mosquito bites. A greater human density may decrease the number of mosquito bites received by each person but increase the number of people that each mosquito bites, complicating efforts to estimate dengue risk for a given population. Variation in the rate of vector production will directly impact human transmission to mosquitoes, but only indirectly affect mosquito transmission to humans through the abundance of infectious mosquitoes.
Additionally, the characteristic aggregation of A. aegypti across houses suggests a low probability of a high-impact event. That is, if an infectious person contacts a house where mosquitoes aggregate, many potentially infected mosquitoes may result. However, when mosquitoes are aggregated in only a few houses, it is more likely that a randomly introduced human infection will contact a house with few mosquitoes, resulting in a small number of secondary cases. Understanding the balance between human density and these opposing influences of mosquito aggregation is essential for entomological risk assessment and for the optimization of source reduction strategies.
In this paper we integrate field data with simulation modeling in order to develop a better understanding of how the interaction between human and mosquito densities facilitates dengue transmission and to provide guidelines for designing and evaluating targeted prevention programs. Using field-collected snapshots of the distribution of A. aegypti pupae and humans across houses surveyed in Armenia, Colombia, we evaluated the impact of human and A. aegypti pupal densities on the simulated number of secondary human infections and the epidemic potential. We determined how A. aegypti production and human densities affected the propagation of dengue across houses and identified the field indices that most correlate with entomological risk. In addition, we used a vector control simulation to determine how human density modulates the long-term impact of targeted control of highly productive A. aegypti habitats.
The field data used in this study was collected in Las Colinas and La Fachada, two highly endemic neighborhoods of the city of Armenia, Colombia. Armenia had the highest number of reported dengue cases of any Colombian city between 2001 and 2008, and the highest cumulative incidence between 2001 and 2011, according to surveillance records of the Instituto Nacional de Salud (INS) of the Colombian Ministry of Health. La Fachada and Las Colinas have elevations of 1335 and 1329 m, respectively and we have observed mean ambient temperatures of 24–25°C through limited surveillance using thermal sensors. In each neighborhood (Fig. 1) we randomly selected eight study patches comprising between 41 and 142 houses (1–3 adjacent blocks) for a total of 16 patches across the city. Over a 26-month period seven censuses of water-holding vessels, including counting of all A. aegypti pupae, were conducted in each patch by public vector control technicians under supervision of Armenia's Health Secretary and INS investigators. This provided us with a dataset of 112 sample distributions of mosquito pupae and humans (seven surveys for each of the 16 patches), with minimal variation in the types of productive containers, housing structure, climate, housing density, and vegetation across samples. An exhaustive human demographic census was conducted in 2009.
We conducted three separate analyses using our model. First, to evaluate the effects of host densities on the generation of secondary cases we ran the model for 300 iterations for each of the 112 patch-surveys (33,600 total iterations), while introducing a single infectious human in every iteration. Second, to simulate dengue introduction from an external area we again ran the model across the range of scenarios while stochastically introducing dengue by assigning a probability of becoming infectious (intro) to each person. Finally, to evaluate targeted vector control interventions in variable human densities, we simulated the elimination of all containers containing a threshold number of pupae in each of the 16 patches, while varying human density by a fixed proportion of the field-observed value.
Over the seven surveys, we found 1,707 vessels infested with A. aegypti and counted a total of 32,058 pupae. Inspection percentage of the 1364 premises in the study patches averaged 70.4% across the seven surveys, with 79% of all premises inspected in at least four of the seven surveys. Our 16 surveyed patches ranged from 170 to 587 people and from 41 to 142 houses. Human densities in study patches ranged from 3.2 to 4.5 residents per house (Table 1). Over the 112 patch-surveys (7 surveys of each patch), the mean number of A. aegypti pupae per house ranged from 0.017 to 30.9. Pupal production was highly aggregated, with 92% of the total pupae found in only 5% of the house-surveys (those containing at least 16 pupae). Roughly 80% of the variation in vector production could be explained by pupal abundance in the most productive container in each of the 112 patch-surveys (Fig. 2).
Across our simulations, 72% of the 112 patch-survey scenarios had R0>1 (average number of secondary infections across 300 model iterations). R0 across patches (secondary infections averaged across all iterations of all seven surveys) varied between 0.88 and 3.87, and was above 1 in 14 of 16 patches (Table 1) Variation in R0 across patches was highly correlated to the frequency of viral introductions that generated greater than 20 secondary infections (R2 = 0.95, Fig. 3), which represented only 10 percent of the total model iterations. In all patches the large majority of viral introductions did not result in secondary transmission (Table 1). This indicates that secondary transmission was largely driven by the occurrence of highly explosive transmission events.
We correlated human-to-mosquito and mosquito-to-human transmission with secondary infections across each of the 33,600 iterations (300 iterations for each of the 112 patch-surveys). Human-to-mosquito transmission explained roughly 20% more of the variation in secondary infections (Fig. 4a) than mosquito-to-human transmission (Fig. 4b). Many introductions that resulted in highly explosive transmission occurred when infectious mosquitoes transmitted to relatively few humans (Fig. 4b). We used the model to evaluate pupal abundance in containers that produced infected mosquitoes. Introduced human infections could only infect large number of mosquitoes when these were produced in containers with high pupal abundance (Fig. 5a). For example, in 10% of model iterations (with at least one infected mosquito) the introduced human infection infected greater than 10 mosquitoes. These mosquitoes were never produced in containers with a mean pupal count less than 16 (corresponding to less than 5% of house-surveys) (Fig. 5a). By contrast, transmission to humans per infectious mosquito was not associated with the number of pupae in the containers that produced the infectious mosquitoes. When infected mosquitoes were produced in containers with few pupae, very few model iterations produced high levels of secondary human infection (Fig. 5b). For example, infected mosquitoes generated ≥20 secondary human infections in only 0.4% of the trials in which they were produced in containers with on average less than 16 pupae.
Because of model stochasticity and limited variability in human density across patches, human density effects were analyzed across quartiles of patch-wide human density (Fig. 6a) and the density of residents in the I0 house (Fig. 6b). Both mosquito-to-human and human-to-mosquito transmission were significantly lower in density quartile 1 (Fig. 6a). By contrast, increased number of residents in the Io house had little impact on mosquito to human spreading, but significantly reduced human to mosquito spreading as resident density increased between quartiles 2 and 4.
A multinomial negative binomial model revealed that the number of humans per house, the average daily number of biting vectors in the I0 house, and the total number of pupae all increased dengue transmission, while the number of resident humans in the Io house negatively affected transmission (Table 2). Secondary infections were more closely associated with I0 vectors (Z = 28.3) than with total pupae (Z = 15.0). Total human population size did not significantly affect transmission when the other covariates were included (Table 2). These results suggest that the aggregated distribution of pupae may cause explosive dengue transmission through the concentration of biting mosquitoes in the I0 house. They also suggest that although increased human residents dilute each individual contact with mosquitoes at the household level, transmission across houses at the patch level is favored by human density. Overall, these data show that vector production and human density can influence the dengue R0 through both household and community level processes.
We evaluated the association of entomological surveillance indicators with epidemic potential, averaged over 1000 iterations in which the virus was successfully introduced. All metrics were significantly associated with epidemic potential, but the product of pupal abundance and human density (pupae x humans per house) was a substantially stronger predictor than the others. In particular, pupae x humans per house had over three times the log-likelihood of predicting epidemic potential compared to pupae per house or pupae per humans (Table 3). This indicates a strong interaction between human density and pupal production.
We simulated the input of eliminating pupae in houses above a threshold pupal density on epidemic potential under a range of human population densities. This enabled us to eliminate any potential colinearities between human density and vector production in the field data. The substantial differences in the slopes of the curves in Figure 7 indicate that a greater reduction in epidemic potential was achieved at higher human densities as the pupal abundance of containers targeted for control rose (Fig. 7). This re-confirms the interaction between vector production and human density. Targeted source reduction yielded roughly 2.5 times the decrease in epidemic potential when human density was doubled (Fig. 7). However, this magnification occurred when houses with high pupal densities were intervened (Fig. 7). By contrast, increasing the target threshold from 1 to 16 pupae did not significantly affect the epidemic potential in any of the human densities (Fig. 7). These findings indicate that the interaction of vector production and human density was due to the combined impacts of A. aegypti super-production on both the spatial aggregation and the overall rate of vector recruitment. If the interaction was caused only by overall vector recruitment, larger differences in the slopes (Fig. 7) would be observed when less productive habitats were targeted.
Understanding how the natural regulation of A. aegypti production and human density influence the dynamics of dengue propagation is critical for optimizing vector control programs. Previous modeling studies demonstrate that because most A. aegypti are produced in only a few containers, elimination of a small subset of containers is sufficient to drive the dengue R0 below 1 [13], [14]. These studies are based on the classical assumption in mosquito borne disease modeling of homogenous mixing of hosts and vectors [19]. This assumption leads to the conclusion that A. aegypti 's vectorial capacity decreases as human abundance increases relative to vector abundance [14]. In this paper, through the use of a spatially explicit agent-based model, we were able to relax this assumption in order to assess how observed variations in pupal production and human density reflect variation in the intensity of dengue transmission in a community of houses. Our model re-confirmed the potential for source reduction to substantially reduce dengue by targeting only containers that produce above a threshold number of pupae, but generated surprisingly contrasting results with regards to how pupal production interacts with human density. Rather than an inverse relationship as predicted by the traditional vectorial capacity equation [19], we found that increased human density can favor Ro through both human-to-mosquito and mosquito-to-human transmission. Moreover, when viral introduction was accounted for, human density amplified the effect of A. aegypti super-production on dengue risk. By parameterizing vector dynamics with seven seasonal pupal surveys, we show that long-term decreases in vector production can achieve substantially larger reductions in epidemic potential when concentrated in areas of higher human density.
In infectious disease ecology, super-spreading occurs when an individual host causes an inordinate number of secondary infections as compared to the majority of hosts [33]. Our model indicates that super-spreading may play a critical role in dengue transmission. Although the large majority of dengue introductions did not generate secondary human infections, 95% of the variation in R0 was explained by the frequency of introductions that generated at least 20 secondary human infections. These results are supported by the spatiotemporal clustering of dengue infection at the city-block level [34] and our preliminary data from a pilot study of dengue infection clusters in these same Armenia neighborhoods (Padmanabha et al, in revision). Our model directly links dengue super-spreading to the dominant role of super-productive vessels in vector recruitment. Introductions that resulted in greater than 20 secondary human infections occurred by and large, only when infected mosquitoes were produced in houses among the top 5 percentile of vector production. Moreover, the density of biting A. aegypti in the residence of the introduced human infection was the strongest predictor of secondary transmission. This suggests that by concentrating the large majority of the emergent vector population in only a few houses, super-production facilitates human infection of large numbers of mosquitoes, albeit infrequently. Thus, while homogenous mixing models establish that dengue transmission requires a threshold vector density [13], [14], our model mechanistically links the regulation of A. aegypti production with the propagation of dengue across houses.
These mechanistic details proved critical to achieving a more complete understanding of the effects of human density on dengue risk. Vector production influenced the transmission cycle through human-to-mosquito transmission. Due to the observed variation in super-productivity across patch surveys (Fig. 2), human-to-mosquito transmission had a much larger impact on variation in secondary infections than mosquito-to-human transmission. In addition to vector production, the introduction of infectious humans also acts through human-to-mosquito transmission. Thus, when we incorporated a likelihood of human introduction into our risk indicator, areas with higher human density had an increased likelihood of having an introduced human infection reside in a house where mosquitoes concentrated, thereby generating an interaction between vector production and human density. Moreover, intra-patch human social interactions caused increased human density to increase the probability that each house, including those where mosquitoes concentrated, received a visit from the introduced infection. This is the reason why increased human density favored human-to-mosquito infection, even though the household size of introduced infections reduced their average contact rate with mosquitoes.
These interactive relationships between human density and vector production may have direct implications for dengue risk assessment and resource allocation for vector control. Previous models assuming homogenous mixing of hosts and vectors predict a threshold value of the index pupae per person required for dengue epidemics [14]. We found that the index pupae x humans per house had a stronger correlation with epidemic potential than pupae per person, a measure that does not account for the human density dependence of both the dengue Ro and viral introduction. Moreover, targeted vector control in areas of high human density may reduce epidemic potential by decreasing the abundance of mosquitoes in areas where dengue is most likely to be introduced. This suggests an opportunity for multi-level targeting of source reduction efforts. Specifically human density could be used to determine in which neighborhoods to focus vector control, and the likelihood of A. aegypti super-production could be used to focus efforts within targeted neighborhoods.
Our model reinforces the need to better understand the dynamics of human exposure to mosquitoes outside the house. Given the short lifespan and limited dispersal of urban A. aegypti, human social networks are likely to drive the constant re-introduction of dengue into patches. In the absence of data on the type, duration, and location of social contacts that lead to A. aegypti exposure, we sought a balance between simplicity and realism in our assumptions regarding the frequency of contact with mosquitoes outside one's home. As such, we excluded major public centers, such as schools, offices or parks, because the residential blocks in our field study were devoid of these. Although these areas are potentially important in transmission dynamics [23], A. aegypti contact with visitors is likely to be higher for social contacts that involve household visits, due to A. aegypti's endophilic nature and domestic oviposition sites. For example, visits between neighbors, and especially between children [who are more likely to be susceptible to dengue in highly endemic areas], may have more relevance in terms of exposure to A. aegypti. Moreover, we have observed that in Colombian cities, working class and marginal areas, such as La Fachada and Las Colinas, have more interactions among neighbors than in affluent areas. This was a major motivation for including a distance dependent component to the intra-patch social contacts in our model. Accordingly, we consider conservative our assumption of daily exposure to A. aegypti bites in exactly one other house within the patch. An increase in intra-patch social contacts with A. aegypti exposure is likely to heighten the human density dependence of dengue transmission. Future expansion of our work would benefit from an improved understanding of (1) how A. aegypti biting habits influence human exposure in non-residential premises and (2) how housing density, age and social class affect the geography and centrality of social networks that involve exposure to A. aegypti.
Recently it has been shown that larval environmental conditions, including resource availability and thermal conditions, can affect the vector competence of Aedes spp [35], [36]. All things being equal, our study and others [13], [14] indicate that most secondary infections are generated by super-productive habitats. When these habitats are absent explosive transmission was nearly impossible in our model. Because most viral introductions do not generate secondary infections explosive transmission events were critical to the dengue R0. This suggests that it would be highly beneficial for dengue virus to efficiently infect and disseminate better in mosquitoes produced in super-productive habitats. However, there is a lack of understanding of exactly what eco-physiological conditions are associated with super-productive habitats. Human behavior, particularly emptying frequency and water usage, unquestionably plays an important role [7], [8].It is also conceivable that because super-productive containers necessarily have large L4 cohorts, which can exert significant competitive pressure on one another [9],the physiology of adult mosquitoes has particular adaptations to resource poor environments. Mild resource competition in the larval environment has been shown to favor dengue infection and dissemination in A. aegypti [36]. Our results suggest that this finding could be an outcome of viral adaptation to mosquitoes produced in super-productive vessels. Furthermore, our model can be used to explore the eco-epidemiological implications of such evolution. We speculate that it would intensify the interactive effect between super-production and human density in favoring dengue risk.
In summary, we found that variation in A. aegypti production across socio-ecologically similar urban patches can generate large variation in secondary transmission and the epidemic potential of dengue, with human population density magnifying these effects. Our results suggest that super-spreading plays an important role in dengue transmission and occurs when a viremic human is bitten by a large number of mosquitoes that were produced in a super-productive vessel. Human density, in turn, can potentiate the epidemiological significance of super-productive A. aegypti habitats. These results re-affirm the importance of spatial heterogeneity in fine-scale dengue dynamics [37]. Moreover, because both human density and the frequency of A. aegypti spuper-production may vary widely within rapidly urbanizing developing countries, our results may be useful for stratifying risk. By contrast, while the dengue system is theoretically very sensitive to A. aegypti survival and biting rates [38], [39], there is little evidence to suggest that either of these processes will significantly vary across areas in the same city with similar climatic conditions. We suggest that by mechanistically evaluating the epidemiological impacts of observed socio-ecological variation, further modeling studies can contribute to the development of a comprehensive framework for stratifying epidemic risk and optimizing dengue prevention resources.
|
10.1371/journal.ppat.1003690 | Structure of a Bimodular Botulinum Neurotoxin Complex Provides Insights into Its Oral Toxicity | Botulinum neurotoxins (BoNTs) are produced by Clostridium botulinum and cause the fatal disease botulism, a flaccid paralysis of the muscle. BoNTs are released together with several auxiliary proteins as progenitor toxin complexes (PTCs) to become highly potent oral poisons. Here, we report the structure of a ∼760 kDa 14-subunit large PTC of serotype A (L-PTC/A) and reveal insight into its absorption mechanism. Using a combination of X-ray crystallography, electron microscopy, and functional studies, we found that L-PTC/A consists of two structurally and functionally independent sub-complexes. A hetero-dimeric 290 kDa complex protects BoNT, while a hetero-dodecameric 470 kDa complex facilitates its absorption in the harsh environment of the gastrointestinal tract. BoNT absorption is mediated by nine glycan-binding sites on the dodecameric sub-complex that forms multivalent interactions with carbohydrate receptors on intestinal epithelial cells. We identified monosaccharides that blocked oral BoNT intoxication in mice, which suggests a new strategy for the development of preventive countermeasures for BoNTs based on carbohydrate receptor mimicry.
| Food-borne botulinum neurotoxin (BoNT) poisoning results in fatal muscle paralysis. But how can BoNT–a large protein released by the bacteria clostridia–survive the hostile gastrointestinal (GI) tract to gain access to neurons that control muscle contraction? Here, we report the complete structure of a bimodular ∼760 kDa BoNT/A large progenitor toxin complex (L-PTC), which is composed of BoNT and four non-toxic bacterial proteins. The architecture of this bacterial machinery mimics an Apollo lunar module, whereby the “ascent stage” (a ∼290 kDa module) protects BoNT from destruction in the GI tract and the 3-arm “descent stage” (a ∼470 kDa module) mediates absorption of BoNT by binding to host carbohydrate receptors in the small intestine. This new finding has helped us identify the carbohydrate-binding sites and the monosaccharide IPTG as a prototypical oral inhibitor, which extends survival following lethal BoNT/A intoxication of mice. Hence, pre-treatment with small molecule inhibitors based on carbohydrate receptor mimicry can provide temporary protection against BoNT entry into the circulation.
| The seven botulinum neurotoxin serotypes (BoNT/A–G) produced by Clostridium botulinum are the causative agents of the neuroparalytic syndrome of botulism and pose a serious threat for bioterrorism [1]. Conversely, BoNT/A is a highly effective therapy for treating neurological disorders [2]. The naturally occurring BoNTs are released together with up to four non-toxic neurotoxin-associated proteins (NAPs) (also called associated non-toxic proteins, ANTPs) in the form of progenitor toxin complexes (PTCs) with different molecular compositions [3]. Such PTCs are highly potent food poisons, e.g., the PTC of BoNT/A displays an oral LD50 of ∼35 µg/kg body weight [4]. While BoNT is sensitive to denaturation by the acidic environment and digestive proteases present in the gastrointestinal (GI) tract [5], the PTCs of different serotypes exhibit ∼360–16,000-fold greater oral toxicity than free BoNT [4], [6], [7], [8]. The NAPs are encoded together with the bont gene in one of two different gene clusters, the HA cluster or the orfX cluster [9]. Both clusters encode the non-toxic non-hemagglutinin (NTNHA) protein, which adopts a BoNT-like structure despite its lack of neurotoxicity [5]. The HA gene cluster also encodes three hemagglutinins (HA70, HA17, and HA33; also called HA3, HA2, and HA1, respectively), which together with BoNT and NTNHA constitute the large PTC (L-PTC) [10]. The structure and function of the corresponding orfX proteins are largely unknown [11].
Structural information of HAs is available for serotypes C and D, such as the crystal structures of HA33 of serotype C (HA33-C) [12], [13], a complex composed of HA17 and HA33 of serotype D [14], and HA70 of serotype C (HA70-C) [15], [16]. However, BoNT/C and D rarely cause human botulism but are known to cause the syndrome in cattle, poultry, and wild birds. For BoNT/A, the major cause of human botulism, only the structure of HA33 (HA33-A), which displays an amino-acid identity of ∼38% to HA33-C and D, has been solved [17]. We have recently determined the crystal structure of the BoNT/A–NTNHA complex [5]. However, it remains largely unclear how the HAs assemble with one another and how they interact with BoNT and NTNHA.
Various structural models have been proposed for the L-PTC. One recent paper suggested a complex composed of BoNT∶NTNHA∶HA70∶HA17∶HA33 in a 1∶1∶2∶2∶3 ratio for L-PTC/A [18], whereas earlier studies suggested a stoichiometry of 1∶1∶3–5∶5–6∶8–9 or 1∶1∶3∶3∶4 for L-PTC/A, or 1∶1∶2∶4∶4 for L-PTC/D [19], [20], [21]. In comparison, electron microscopy (EM) studies on L-PTC/A, B and D supported a stoichiometry of 1∶1∶3∶3∶6 [14], [22].
The functional roles of NAPs are also not well defined. We have recently shown that NTNHA shields BoNT against low-pH denaturation and proteolytic attack in the GI tract by forming the minimally functional PTC (M-PTC), and releases it during entry into the general circulation [5], [23]. However, it is not clear whether HAs further protect the toxin. At the same time, the L-PTC may contribute to BoNT internalization into the host bloodstream through interactions with intestinal cell surface glycans [24], [25], [26]. The HAs of BoNT/A and B could disrupt the human epithelial intercellular junction through species-specific interaction with E-cadherin, presumably facilitating BoNT transport via the paracellular route [27], [28], [29]. Defining the L-PTC structure would permit a more complete understanding of the complex's role in toxin shielding and delivery, and would help to describe the molecular mechanism underlying these important actions.
Here, we report the structure of a ∼760 kDa L-PTC/A using a combination of X-ray crystallography, single-particle EM and three-dimensional reconstruction (3D-EM). We found that L-PTC/A consists of two structurally and functionally independent sub-complexes, the M-PTC and the HA complex. The HA complex is composed of HA70, HA17, and HA33 in a 3∶3∶6 stoichiometry and adopts an extended three-blade architecture, whereas the M-PTC is situated on top of the HA complex platform. BoNT/A absorption is mainly mediated by nine glycan-binding sites on the HA complex that together form multivalent interactions with host carbohydrate receptors on intestinal epithelial cells. HA complex-mediated toxin absorption can be blocked in vitro by carbohydrate receptor mimics. The monosaccharide IPTG also inhibits oral BoNT/A intoxication in mice, providing the first approach for a possible preventive treatment prior to deliberate BoNT poisoning.
The high toxicity of BoNT/A prevents imaging of the fully active toxin by cryo-EM. So, we began our analysis with negative-staining EM and determined the 3D molecular envelope of L-PTC/A at ∼31 Å resolution (Fig. 1A–C and Fig. S1 in Text S1). The M-PTC moiety was clearly identified in the EM density map based on its crystal structure [5]. Beneath the M-PTC, the HAs adopt a symmetric three-blade architecture that is ∼100 Å tall and ∼260 Å wide between the tips of neighboring blades. Surprisingly, the M-PTC and the HA complex are relatively independent of each other and associate only through two small interfaces (Fig. 1D and Fig. S2 in Text S1). This arrangement contrasts with the extensive interactions between BoNT/A and NTNHA that are required for mutual protection in the GI tract [5], suggesting that the HA complex might play a minimal role in BoNT protection. We did not observe the LL-PTC under EM, which has been proposed to be a dimer of the L-PTC with a molecular weight of ∼900 kDa that might only occur at high concentrations [20], [30].
To determine the molecular architecture of the HA complex, we produced highly homogeneous recombinant proteins of HA70, HA33, HA70–HA17, HA17–HA33, HA70D3(residues Pro378–Asn626)–HA17–HA33 (termed the mini-HA complex), and the complete HA70–HA17–HA33 (the HA complex). HA17 formed inclusion bodies and heterogeneous soluble aggregates when expressed and purified alone. This is probably due to the large hydrophobic patches on its surface, which are protected by its binding partners within the HA complex. We then systematically analyzed the solution association of these individual proteins and their complexes using analytical ultracentrifugation (AUC), which was performed at pH 2.3 and 7.6 to mimic the physiological conditions in the GI tract (Table S1 in Text S1). Our data indicate that the HA complex assembles at both pHs as a hetero-dodecamer consisting of HA70, HA17, and HA33 in a 3∶3∶6 ratio to yield a ∼470 kDa complex. Specifically, homo-trimeric HA70 forms the core of the complex with each C-terminal HA70D3 domain binding to one HA17, which in turn simultaneously coordinates two HA33s.
We next separated the HA complex into two major components: the central hub composed of homo-trimeric HA70 and the blade composed of HA70D3–HA17–HA33. Their crystal structures were determined at 2.9 Å and 3.7 Å, respectively (Fig. 2C–D, Table S2 in Text S1, Fig. S3–S4 in Text S1). We also obtained a high-resolution structure of the blade by combining the structures of HA70D3–HA17 and HA17–HA33, which were determined at 2.4 Å and 2.1 Å, respectively (Fig. 2A–B and Fig. S5–S6 in Text S1). Each HA adopts an almost identical conformation in the independently solved structures, despite differences in crystal packing, suggesting that they represent physiologically relevant conformations.
HA70 consists of three domains (D1–3) (Fig. S3 in Text S1). The D1 and D2 domains, which adopt similar structures, mediate the trimerization of HA70 with each protomer contributing ∼3,100 Å2 of solvent-accessible area for interactions. The D3 domain, sitting at the tip of the trimer, is composed of two similar jelly-roll-like β-sandwich structures. The linker between D1 and D2 (residues Thr190–Ser205) is degraded and not visible in the crystal structure, which is reminiscent of the post-translational nicking of HA70 into ∼25 and ∼45 kDa fragments that occurs physiologically [20].
HA17 has a compact β-trefoil fold and connects HA70 and HA33. Based on the crystal structure of the HA70D3–HA17 complex, the interactions between HA70 and HA17 bury a solvent-accessible area of ∼795 Å2 (per molecule) (Fig. 3A and Fig. S5 in Text S1). The structure of HA70D3 is almost identical to its equivalent domain in the full-length HA70 with a root-mean-square deviation (rmsd) of ∼0.93 Å over 232 Cα atoms. The major HA70–HA17 interactions are composed of 13 pairs of hydrogen bonds and salt bridges. In addition, HA70-Phe547 is buried in a hydrophobic region in HA17 composed of Ile18, Ile92, Ala93, Thr96, and Met140 (Fig. 3A and Fig. S5 in Text S1).
HA17 simultaneously binds to two HA33 molecules that form a dumbbell-like shape composed of two β-trefoil domains linked by an α-helix. The two pairs of HA17–HA33 interfaces bury a solvent-accessible area of ∼666 Å2 and ∼410 Å2 (per molecule), respectively (Fig. 3B and Fig. S6 in Text S1). The two HA33-binding interfaces on HA17 are adjacent but non-overlapping. HA17 contributes seven and four pairs of hydrogen bonds and salt bridges to bind the two HA33 molecules, respectively. Complementing these hydrophilic interactions, the two HA33s contain a hydrophobic surface (Trp75/Leu116/Leu129) that interacts with two neighboring hydrophobic patches on the HA17 surface (Phe75/Pro125/L127 and Leu108/Pro130/Phe132) (Fig. 3B).
The two molecules of HA33 in each blade of the HA complex are almost identical (rmsd of ∼0.35 Å over 286 Cα atoms) and bury a solvent-accessible area of ∼961 Å2 (per molecule) between them (Fig. 3C). All the interacting residues are in the N-terminal domain of HA33, whereas the interface consists of hydrophilic interactions on the periphery and a hydrophobic core in the center (Fig. S6C in Text S1). Due to the two-fold symmetry between the two molecules, intra-HA33 interactions are generally symmetric.
Finally, we assembled the subunit crystal structures to create a complete structure of the HA complex (Fig. 2E). The 12-subunit HA complex is stabilized by numerous protein–protein interactions that include interactions among the HA70s of the central trimer, between HA70 and HA17, between HA17 and the two HA33 molecules, and between the two HA33s in each blade. The assembled HA complex structure could be unambiguously docked into the 3D-EM density of the native L-PTC/A (correlation coefficient, CC∼87.7%) (Fig. 1), whereas a small difference was observed in the C-terminal domain of HA33 due to its structural flexibility. We also performed an independent 3D-EM reconstruction of our recombinant, in vitro-reconstituted HA complex at ∼31 Å resolution (CC∼93.1%) (Fig. 2E), and found it to be almost identical to the HA complex present in the L-PTC.
The crystal structure of the M-PTC was unambiguously docked into the 3D-EM density of the native L-PTC (CC∼87.3%), which is situated on top of the HA complex, yielding a ∼760 kDa 14-subunit complex (Fig. 1 and Fig. S2 in Text S1). BoNT/A interacts with the HAs only through its receptor-binding domain (HC domain).The interface is likely composed of Gly399 and Ile400 in HA70 and Val1128, Gly1129, Glu1210, Pro1212, and Asp1213 in HC (pairwise Cα–Cα distance within 15 Å) (Fig. 1D and Fig. S7A in Text S1). Gly399 and Ile400 of HA70 are located in a loop that has weak electron density in the crystal structures, suggesting high flexibility. Moreover, the potentially interacting residues in HC are located in two flexible loops and not conserved among various BoNT serotypes (Fig. S7B in Text S1). Thus, the BoNT/A–HA70 interface in the L-PTC may be formed by induced fit.
The major interface between the M-PTC and the HAs is mediated by NTNHA. The potential interface residues in NTNHA, which are within 12 Å Cα–Cα distance of the HAs, are located in loop Gly308–Gly313 and the residues flanking loop Gly116–Ala148 (nLoop) [5]. The corresponding interface residues in the HA complex are located around the center of the HA70 trimer (Fig. 1D). The nLoop displays no visible electron density in the structure of the M-PTC and is spontaneously nicked in the free NTNHA or the M-PTC [5], [31], [32], [33], [34]. However, the nLoop remains intact in the L-PTC, suggesting it may be buried by the HA complex [30], [31], [35]. We found that the synthetic nLoop peptide binds to HA70 with high affinity (Kd∼340 nM) at a stoichiometry of one nLoop to one HA70 trimer (Table S3 in Text S1). This finding unambiguously established the orientation of the pseudo 2-fold symmetric M-PTC on top of the HA complex. The nLoop of NTNHA binds strongly to HA70 at pH 7.6, which is in contrast to the M-PTC that dissembles and releases BoNT/A at neutral or basic pH [5], [30]. This suggests that BoNT may be the only component released from the L-PTC in response to the pH change encountered upon entering the circulation [30].
The HA complex and the M-PTC are stable at low pH (e.g., pH 2.3) and are resistant to digestive proteases, as shown by in vitro cleavage by trypsin and pepsin (Fig. S8 in Text S1) [5]. The loose association between these two complexes suggests that they may have distinct functions during oral intoxication. The penetration of BoNT through an epithelial cell barrier to reach the general circulation is the first essential step of oral BoNT intoxication, which prompted us to investigate the role of HAs in BoNT/A absorption from the GI tract. For this study, we used the well-characterized intestinal epithelial cell line Caco-2. Although derived from a human colon adenocarcinoma, Caco-2 cells differentiate to form a polarized epithelial cell monolayer that provides a physical and biochemical barrier to the passage of ions and small molecules, resembling the uptake and barrier properties of the small intestinal epithelial layer [36], [37], [38], [39]. Caco-2 cells have been extensively used to investigate their permeability upon infection, e.g. by rotavirus [40] or enteropathogenic E. coli [41], and transcytosis upon intoxication with cholera toxin [42] or BoNT [43], [44], [45]. Furthermore, it was demonstrated that the transepithelial electrical resistance (TER) of Caco-2 cell monolayers was reduced by the L-PTC of BoNT/A and B. Although the mechanism by which this may occur is unclear, BoNT absorption has been proposed to occur via the paracellular route [27], [28], [29].
We found that treatment of Caco-2 cells with the in vitro-reconstituted HA complex markedly reduced the TER. This effect was more marked when the HA complex was applied to the cell monolayer from the basolateral side than from the apical side, which needed ∼17 nM and ∼58 nM to achieve a 90% and 70% decrease in TER after 12 hours, respectively (Fig. S9A–B in Text S1). Remarkably, the potency of the isolated HA complex was similar to that of the intact L-PTC (Fig. 4A–B). In contrast, there was no effect on Caco-2 TER by BoNT/A, NTNHA, the M-PTC, or by the subunits of the HA complex, including HA70, HA33, the HA17–HA33 complex, and the mini-HA complex (Fig. 4C–D). Taken together, these data suggest that the fully assembled HA complex is the functional unit of the L-PTC that facilitates intestinal absorption of BoNT, and acts by compromising the integrity of the epithelial cell layer.
Many human receptors for microbial pathogens or toxins are glycoconjugates. The L-PTC is known to initiate toxin absorption by binding to intestinal cell surface glycans [24], [25], [26]. We therefore performed a comprehensive thermodynamic analysis to characterize the interactions between HAs and several common monosaccharides and oligosaccharides (Fig. S10 and Table S3 in Text S1). We found that HA33 bound to lactose (Lac), N-acetyllactosamine (LacNAc), and galactose (Gal) with dissociation constants (Kd) of ∼1.0 mM, ∼1.0 mM, and ∼1.8 mM, respectively, and that it also bound to isopropyl β-D-1-thiogalactopyranoside (IPTG) [46], a non-metabolizable galactose analog, with a Kd of ∼0.8 mM. HA70 bound to α2,3- and α2,6-sialyllactose (SiaLac), both with Kd of ∼0.5 mM, and displayed a lower affinity for N-acetylneuraminic acid (Neu5Ac) (Kd∼7.8 mM). There was no overlap between the carbohydrate selectivity of HA70 and HA33.
To determine the physiological relevance of these HA–glycan interactions during toxin absorption, we examined their ability to interfere with the HA complex-mediated disruption of Caco-2 TER. Lac, Gal, and IPTG markedly inhibited the TER reduction induced by the HA complex, and showed higher potencies when applied to the apical than to the basolateral compartment (Fig. 5A–B and Fig. S11A–D in Text S1). In contrast, α2,3- and α2,6-SiaLac, and to a lesser extent Neu5Ac, inhibited the decrease in TER only when applied apically, albeit more weakly than Lac (Fig. 5A–B and Fig. S11E–F in Text S1). We then examined the transport of the HA complex across the Caco-2 monolayer using a fluorescence-labeled HA complex (HA*) (Fig. 5C). Lac and IPTG efficiently inhibited the transport of HA* when it was applied to the apical or basolateral chamber. Blocking the transport of HA* via α2,3- and α2,6-SiaLac was more potent toward the basolateral compartment than toward the apical side. Neu5Ac at 50 mM did not inhibit transport of HA* from either side of the Caco-2 cell monolayer. These data are consistent with the ability of these carbohydrates to inhibit TER reduction induced by the HA complex. Collectively, these results suggest that the binding of HAs to Neu5Ac- and Gal-containing glycans on epithelial cells is essential for the transport of BoNT across the intestinal wall. Moreover, the carbohydrate receptors may play a more important role in the initial L-PTC absorption in the intestinal lumen, whereas other host receptors (e.g., E-cadherin) are involved once it gains access to the basolateral side.
To fully understand the binding specificity, we determined the crystal structures of HA70 in a complex with α2,3- or α2,6-SiaLac (Fig. 6 and Table S4 in Text S1). We found that α2,3- and α2,6-SiaLac bound to the same region in the D3 domain of HA70, where the terminal Neu5Ac in both glycans mediates the majority of the HA70–glycan interactions through six pairs of hydrogen bonds (Fig. 6A and Fig. S12 in Text S1). Mutating the Neu5Ac-binding residues (e.g. T527P, R528A, or T527P/R528A) completely abolished the binding (Table S3 in Text S1). The Neu5Ac-binding mode in HA70-A is also conserved in HA70-C (Fig. S13 in Text S1) [15], [16], suggesting HA70 is unlikely to be a major determinant of the host tropism of various BoNT serotypes.
In contrast to the well-defined conformation of Neu5Ac, the Gal–Glc moiety seems to have a larger structural flexibility and is not essential to HA70–glycan recognition. Specifically, α2,3-SiaLac adopts a linear conformation, which is likely stabilized by a Glc-mediated crystal contact with its symmetry mate. However, α2,6-SiaLac binding to the same site adopts a folded conformation in which there is no crystal contact and Glc has no visible electron density (Fig. S12B in Text S1). Furthermore, these conformations are also different than that observed in the structures of α2,3- and α2,6-SiaLac when they bound to HA70-C, despite the conserved Neu5Ac-binding mode [16]. The different glycan conformations and the weak electron densities for Gal–Glc observed here are probably due to the intrinsic flexibility of SiaLac in solution [47]. The ability of HA70 to bind SiaLac with different glycosidic linkages contrasts with the binding profile of influenza virus HA. Neu5Ac binds to a deep pocket in influenza HA, which restricts the composition and topology of glycans that can bind to influenza HA [48], [49], [50]. In contrast, the Neu5Ac-binding site in HA70 is on a flat surface, allowing more freedom for additional glycan binding beyond the terminal Neu5Ac.
We also determined the crystal structures of the HA17–HA33 complex bound with Gal, Lac, or LacNAc (Table S4 in Text S1). All three bind to an identical site in HA33, where the HA33–glycan interactions are mediated only by the Gal moiety through extensive hydrogen bonding and a crucial stacking interaction between Phe278 and the hexose ring of Gal (Fig. 6B). The HA33–Gal interaction is well-defined and identical for the two HA33 molecules in one asymmetric unit (AU). The Glc or GlcNAc moiety does not directly interact with HA33. One Glc/GlcNAc in the AU is involved in a crystal packing and shows clear electron density, while the density for the other copy is weakly defined; the latter is likely caused by the weak HA33–glycan binding affinity and intrinsic structural flexibility of HA33 that will be discussed later (Fig. S12D–F in Text S1). To further confirm the structural findings, we mutated the Gal-binding residues in HA33 (e.g., D263A or F278A) and found that these mutations almost completely abrogated the Lac binding (Table S3 in Text S1).
Gal binds at an equivalent site in HA33 of L-PTC/C (HA33-C) (Fig. S14 in Text S1) but with ∼15-fold lower binding affinity than with HA33-A [13], which is likely caused by the replacement of Phe278 in HA33-A with Asp271 in HA33-C. In addition, HA33-C binds Neu5Ac in an adjacent binding site [51]. However, HA33-A does not bind Neu5Ac-containing sugars because the key Neu5Ac-binding residues in HA33-C, Trp176 and Arg183, are replaced in HA33-A with Tyr180 and Asn187, respectively (Table S3 in Text S1). These differences between HA33-A and HA33-C indicate that the known host susceptibility to different BoNT serotypes may be determined in part by the interaction between HA33 and host glycan receptors.
To further analyze the functional role of BoNT's glycan receptors, we “knocked-down” specific glycan binding to the HA complex using structure-based mutagenesis. The HA33-DAFA complex (composed of the wild-type (WT)-HA70, WT-HA17, and HA33-D263A/F278A) did not bind to Gal, whereas the HA70-TPRA complex (composed of the HA70-T527P/R528A, WT-HA17, and WT-HA33) failed to bind to Neu5Ac (Table S3 in Text S1). We found that the HA33-DAFA complex did not reduce TER when applied from either side of the Caco-2 cell monolayer. Furthermore, the loss of the Gal-binding site prevented the transport of HA33-DAFA through the Caco-2 monolayer (Fig. 5C), indicating the crucial role of the carbohydrate interaction during transcytosis. The HA70-TPRA complex maintained activity only when applied from the basolateral side, which was inhibited by Lac (Fig. 6D–E). These data suggest that there are at least two steps at which HA–glycan interactions play an important role in toxin absorption. Both Neu5Ac- and Gal-containing glycans are important for the initial L-PTC absorption in the intestinal lumen, but Gal-containing receptors on the basolateral surface of the epithelial cells may also participate, presumably in facilitating transport via the paracellular route [27], [28], [29].
To determine whether the glycans could interfere with BoNT absorption in vivo, we examined the effect of the monosaccharides Neu5Ac, Gal, and IPTG on the oral toxicity of L-PTC/A in mice [4]. Concomitant oral administration of L-PTC/A and IPTG at 500 mM significantly extended the median survival time (MST) of animals to ∼91 hours compared with ∼39 hours for the control group. Furthermore, IPTG was effective when it was administered one hour prior to treatment with L-PTC/A with an increase of MST to ∼62 hours. Some improvement in survival was also evident with IPTG treatment one hour after intoxication with L-PTC/A, with an increase of MST to ∼55 hours (Fig. 6F). Since IPTG does not affect the neurotoxicity of BoNT/A based on the mice phrenic nerve hemidiaphragm assay, this finding suggests that receptor mimics could block BoNT/A intestinal absorption at an early point of oral intoxication. Gal and Neu5Ac (up to ∼500 mM) did not confer significant protection, most likely due to their low binding affinity and/or metabolism (Fig. S15 in Text S1).
Here, we report the complete structure of a 14-subunit ∼760 kDa L-PTC/A, which is achieved by building novel crystal structures of each subunit into 3D-EM reconstruction. To our knowledge, this is the largest bacterial toxin complex known to date. The L-PTC/A adopts a unique bimodular architecture, whereas BoNT/A and NTNHA form a compact M-PTC and three HA proteins adopt an extended three-arm shape. Our results conclude the same stoichiometry and a similar overall architecture as suggested by recent EM studies of L-PTC/A, B, and D [14], [22]. Furthermore, our complementary crystallographic, EM, and biochemical studies have revealed for the first time that both BoNT/A and NTNHA are involved in interactions with the HA complex, and that the two modules associate through two small interfaces, in contrast to numerous protein–protein interactions within each module.
Aside from a small interface involving the BoNT/A receptor-binding domain, the majority of the interactions between the M-PTC and the HAs are mediated by the NTNHA nLoop. In spite of the overall structural similarity between BoNT/A and NTNHA, the nLoop is a unique feature of NTNHA, which is fully exposed on the M-PTC surface [5]. The nLoop is conserved in the NTNHAs that shield BoNT/A1, B, C, D, and G, and assemble with HAs into the L-PTC. However, the nLoop is missing in NTNHAs that assemble with BoNT/A2, E, and F, which do not have accompanying HA proteins and only form the HA-negative M-PTC [11], [52], [53]. We have found that one molecule of the synthetic nLoop peptide binds to the trimeric HA70 with a high affinity, clearly suggesting that the nLoop is bridging the M-PTC and the HA complex. This is consistent with previous reports that the nLoop is intact in the context of the L-PTC but spontaneously nicked in the free NTNHA or the M-PTC [5], [30], [31], [32], [33], [34].
Structural and sequence analyses suggest that the 12-subunit architecture of the HA complex is likely conserved across different BoNT serotypes [14], [22]. For example, pairwise structural comparisons yield rmsd of ∼1.28 Å (582 Cα atoms) and ∼1.20 Å (137 Cα atoms) for HA70-A/HA70-C and HA17-A/HA17-D, respectively; they are ∼0.87 Å (129 Cα atoms) and ∼1.23 Å (134 Cα atoms) for the two domains of HA33 between serotypes A and D and similarly between HA33-A and HA33-C [13], [14], [15], [16]. Moreover, the protein–protein interactions within the HA70 trimer and between HA17 and HA33 are largely conserved among our crystal structures of serotype A and the known crystal structures of serotypes C and D.
Despite the largely rigid structure of the HA complex, HA33 seems to have an intrinsic structural flexibility. The N-terminal domain of HA33 is fixed in the HA complex through extensive inter-HA33 and HA17–HA33 interactions, but its C-terminal domain is largely unrestricted. When comparing two HA33-A structures that were determined in different crystal forms, we found that the N- and C-terminal domains of HA33-A twist against each other by ∼14° (Fig. S16 in Text S1) [17]. A more significant conformational change is observed between HA33-A and C (∼61°) and HA33-A and D (∼65°) (Fig. S16 in Text S1) [13], [14]. In the context of the assembled HA complex, such a conformational change leads to a shift up to ∼23 Å for the C-terminal Gal-binding site in HA33. We suggest that HA33 could require such structural flexibility to achieve its multivalent host-receptor binding in the intestine.
The loose linkage between the M-PTC and the HA complex clearly suggests divided functions. We previously reported that the M-PTC's compact structure protects BoNT against digestive enzymes and the extreme acidic environment of the GI tract [5], [23]. We now show that the HA complex is mainly responsible for BoNT absorption in the small intestine, through binding to specific host carbohydrate receptors. This new finding permitted the identification of IPTG as a prototypical oral inhibitor that extends survival following lethal oral BoNT/A intoxication of mice. Multivalent interactions involving nine binding sites for Neu5Ac- and Gal-containing glycans increase the overall avidity of binding between the L-PTC and glycans on the epithelial cell surface, and thus compensate for the modest glycan-binding affinities at individual binding sites (Fig. 6C). Similarly, the potency of carbohydrate receptor mimics could be improved by optimizing the HA–glycan interactions as revealed here or by introducing new HA–inhibitor interactions at individual binding sites based on rational design, as well as by designing multivalent inhibitors. Although such inhibitors cannot be used to treat fully developed food-borne botulism, they could provide temporary protection upon pre-treatment and could also be useful for cases of intestinal colonization with C. botulinum spores such as in cases of infant or adult intestinal botulism. Our results also suggest that the L-PTC could be exploited for alternative applications. For example, protein-based therapeutics could be coupled to the modified non-toxic L-PTC to allow oral delivery by improving drug stability, absorption efficiency, and bioavailability.
The Institutional Animal Care and Use Committee of the United States Department of Agriculture, Western Regional Research Center approved the experimental and husbandry procedures used in these studies (protocol # 12-2). All animal experiments were conducted under the guidelines of the U.S. Government Principles for the Utilization and Care of Vertebrate Animals Used in Testing, Research and Training.
The sequences corresponding to full-length HA70 (residues M1–N626), HA70D3 (residues P378–N626), full-length HA17 (residues M1–I146), and full-length HA33 (residues M1–P293) from BoNT/A1-producing C. botulinum strain 62A were cloned into expression vectors pQE30, pGEX-6p-1, pRSFDuet-1, and pET28a, respectively. In addition, HA17 and HA33 were cloned into the bicistronic pRSFDuet-1 for co-expression.
To facilitate protein purification, a 6×His tag followed by a thrombin cleavage site was introduced to the N-termini of HA70, HA17, and HA33. HA70D3 was cloned into pGEX-6p-1 following the N-terminal GST and a PreScission cleavage site. For HA17 and HA33 in the pRSFDuet-1 vector, HA17 was produced with an N-terminal 6×His tag followed by a PreScission cleavage site, while HA33 had no affinity tag. All HA33 or HA70 mutations were generated by QuikChange site-directed mutagenesis (Stratagene).
Four different protein expression schemes were used to produce the individual HAs or HA complexes. (1) HA70 (pQE30), HA70D3 (pGEX-6p-1), and HA33 (pET28a) were expressed alone; (2) HA70 (pQE30) and HA17 (pRSFDuet-1) were co-transformed into bacteria and co-expressed; (3) HA70D3 (pGEX-6p-1) and HA17 (pRSFDuet-1) were co-transformed into bacteria and co-expressed; and (4) HA17 and HA33 were co-expressed using the bicistronic pRSFDuet-1 vector.
All recombinant proteins were expressed in the E. coli strain BL21-RIL (DE3) (Novagen). Bacteria were grown at 37°C in LB medium in the presence of the appropriate selecting antibiotics. Expression was induced with 1 mM isopropyl-β-D-thiogalactopyranoside (IPTG) when OD600 had reached 0.7. The temperature was then decreased to 18°C and expression was continued for ∼16 hours. The cells were harvested by centrifugation and stored at −20°C until use.
For purification of His-tagged proteins (HA70, HA33, the HA70–HA17 complex, and the HA17–HA33 complex), proteins were bound to a Ni-NTA (nitrilotriacetic acid, Qiagen) affinity column in a buffer containing 50 mM Tris (pH 8.0) and 400 mM NaCl, and subsequently eluted in the same buffer containing 300 mM imidazole. The eluted fractions of each protein were pooled and dialyzed overnight at 4°C against a buffer composed of 20 mM Tris (pH 8.0) and 50 mM NaCl, then the His-tag was removed with thrombin (for HA70, HA33, and the HA70–HA17 complex) or PreScission protease (for the HA17–HA33 complex). GST-tagged HA70D3 and the HA70D3–HA17 complex were purified using Glutathione Sepharose 4B resins (GE Healthcare) in phosphate-buffered saline, and eluted from the resins after on-column cleavage using PreScission protease.
The following three schemes were used to further purify the proteins. (1) HA70 and the HA70–HA17 complex was purified by MonoQ ion-exchange chromatography (GE Healthcare) in a buffer containing 20 mM Tris (pH 8.0) and eluted with a NaCl gradient, followed by Superdex 200 size-exclusion chromatography (GE Healthcare) in 20 mM Tris (pH 8.0) and 50 mM NaCl. (2) HA33 and the HA17–HA33 complex were purified by MonoS ion-exchange chromatography in a buffer containing 20 mM sodium acetate (pH 5.0) and eluted with a NaCl gradient, followed by Superdex 200 chromatography in 20 mM Tris (pH 8.0) and 50 mM NaCl. (3) HA70D3 and the HA70D3–HA17 complex were purified by MonoQ ion-exchange chromatography in 20 mM Tris (pH 8.0) followed by Superdex 200 chromatography in 20 mM Tris (pH 8.0) and 50 mM NaCl for HA70D3 or 20 mM sodium citrate (pH 5.0) and 100 mM NaCl for the HA70D3–HA17 complex. Each protein or protein complex was concentrated to ∼3–6 mg/ml using Amicon Ultra centrifugal filters (Millipore) and stored at −80°C until used for further characterization or crystallization.
The purified HA70 was labeled with Alexa Fluor® 488 carboxylic acid, succinimidyl ester (Life Technologies) according to the manufacturer's instructions. The labeled HA70 was further purified by Superdex 200 chromatography in 20 mM Tris (pH 8.0) and 50 mM NaCl. The calculated dye to protein ratio was ∼2 moles of dye per mole of monomeric HA70.
The HA17–HA33, the HA70–HA17, and the HA70D3–HA17 complexes were produced by co-expression and co-purification as described above. To assemble the mini-HA complex (HA70D3–HA17–HA33), the purified HA33 and the HA70D3–HA17 complex were mixed at a molar ratio of ∼2.5∶1 and incubated at 4°C overnight. The excess HA33 was removed by Superdex 200 chromatography with 20 mM Tris (pH 7.6) and 50 mM NaCl. The fully assembled HA complex was reconstituted by mixing the purified HA70 and the HA17–HA33 complex at a molar ratio of ∼1∶1.3. The mixture was incubated at 4°C overnight and the excess HA17–HA33 complex was removed from the mature HA complex by Superdex 200 chromatography with 20 mM Tris (pH 7.6) and 50 mM NaCl. The fluorescence-labeled HA complex was prepared with Alexa Fluor® 488-labeled HA70 and unlabeled HA17–HA33 complex (HA*) or HA17–HA33DAFA complex (HA33DAFA*) using a similar protocol.
Sedimentation equilibrium (SE) experiments were performed in a ProteomeLab XL-I (BeckmanCoulter) analytical ultracentrifuge. Purified HA samples were dialyzed extensively against a buffer containing 50 mM Tris (pH 7.6) and various NaCl concentrations, or 50 mM citric acid (pH 2.3) and various NaCl concentrations. Protein samples at concentrations of 0.4, 0.2, and 0.1 unit of OD280 were loaded in 6-channel equilibrium cells and centrifuged at 20°C in an An-50 Ti 8-place rotor at the first speed indicated until equilibrium was achieved and thereafter at the second speed. HA33 was analyzed at rotor speeds of 19,000 and 22,000 rpm. The HA17–HA33 and the HA70D3–HA17–HA33 complexes were analyzed at 12,000 and 14,000 rpm. The HA70–HA17 and the HA70–HA17–HA33 complexes were run at speeds of 6,000 and 8,000 rpm. For each sample, data sets for the two different speeds were analyzed independently using HeteroAnalysis software (by J.L. Cole and J.W. Lary, University of Connecticut). Three independent experiments were performed for each sample.
The AUC data showed that HA33 is predominantly monomeric in solution at pH 2.3 or pH 7.6. HA17–HA33 forms a tight complex at pH 2.3 or pH 7.6, and the data were best fit to a model composed of one HA17 and two HA33 molecules. The HA70–HA17 complex precipitated at pH 2.3 and was therefore analyzed only at pH 7.6. The best fits for HA70–HA17 clearly suggested a complex composed of three HA70 and three HA17 molecules. The data for the HA70D3–HA17–HA33 complex were best fit to a model composed of one HA70D3, one HA17, and two HA33 molecules. HA70–HA17–HA33 forms a tight complex containing three HA70, three HA17, and six HA33. Weak dimerization was observed for the mini-HA complex (Kd of ∼23.1 µM) and the full HA complex (Kd of ∼10.9 µM) at pH 7.6 in the presence of 100 mM NaCl, but was not observed at higher ionic strength. The weak oligomerization Kd suggests that the mini-HA and the full HA complex are monomeric under physiological conditions.
The L-PTC of BoNT/A was obtained from List Biological Laboratories, Inc. (Campbell, California) and Miprolab GmbH (Göttingen, Germany). The recombinant HA complex was reconstituted in vitro as described above. Negatively stained EM specimens were prepared following a previously described protocol [54]. Briefly, 3 µl of the L-PTC (∼0.02 mg/ml in 20 mM MES, pH 6.2, and 100 mM NaCl) or the HA complex (∼0.01 mg/ml in 20 mM Tris, pH 7.6, and 50 mM NaCl) was placed on a freshly glow-discharged carbon-coated EM grid, blotted with filter paper after 40 seconds, washed with two drops of deionized water, and then stained with two drops of freshly prepared 1% uranyl formate, which also served to fix the proteins.
Particle images were acquired using a 4k×4k TVIPS CCD camera on a Tecnai F20 electron microscope (FEI) equipped with a field emission electron source operated at 200 kV, at a nominal magnification of ∼70,000, resulting in a calibrated pixel size of 4.28 Å/pixel on the object scale after binning. The defocus values were set in the range of 1.5–3.2 µm. The electron dosage was ∼40 electrons/Å2. Image quality was monitored on the basis of power spectra quality. Particle boxing, CTF correction, initial model generation, 3D refinement, and resolution assessment were all carried out with the EMAN2 package [55]. Particles were semi-automatically boxed out and subjected to reference-free class-averaging using EMAN2. The standard EMAN2 initial model generation program (e2initialmodel.py) was used to obtain initial templates for refinement. With the use of this methodology, models were constructed from a series of randomly generated Gaussian blobs and refined against reference-free-generated 2D class averages. The resulting models were ranked on the basis of the agreement of the projection with the class average. The top initial templates were used as starting models for the subsequent refinement with EMAN2. For the L-PTC, no symmetry was imposed throughout the 3D reconstruction and refinement, while for the HA complex, a C3 symmetry was imposed. Refinement was terminated when no significant changes could be visually detected. A data set of 15,140 particles was used for the final reconstructed map of the L-PTC, for which the resolution was estimated to be ∼30.8 Å based on the resolution criteria of Fourier shell correlation (FSC) at 0.5 cutoff. A data set of 3,746 particles was used for the final reconstructed map of the HA complex, for which the resolution was estimated to be ∼30.6 Å. The density maps were filtered to 30 Å with the low-pass filter in EMAN2. Handedness of the maps was determined on the basis of the HA complex structure derived from crystal structures.
Visualization and rigid-body docking of atomic models into the 3D-EM density maps were performed using UCSF Chimera [56]. The Chimera Fit in Map utility, which maximizes the cross-correlation coefficient between the 3D-EM density map and the calculated density map (filtered to 30 Å) of the atomic structures, was used to optimize the docking of atomic structures into 3D-EM maps. After fitting refinement, the positions with highest correlation coefficient (cc) values were chosen. For the HA complex, the atomic structure derived from crystal structures of HA70 and the HA70D3–HA17–HA33 complex fitted the 3D-EM map very well (cc = 93.1%). There was a slight deviation at the C-terminal domain of HA33 that is located at the tip of the complex, which may be due to the structural flexibility of HA33. For the L-PTC, the densities for the M-PTC and the HA complex could be clearly identified in the 3D-EM reconstruction map, and the density for the HA complex was manually 3-fold averaged using Chimera and EMAN2. Atomic structures of the M-PTC and the HA complex were then docked separately into their 3D-EM densities, with highest cc values of 87.3% and 87.7%, respectively. The docked M-PTC and the HA complex were then merged to generate the complete pseudo-atomic model for the L-PTC.
Initial crystallization screens were performed using a Phoenix crystallization robot (Art Robbins Instruments) and high-throughput crystallization screen kits (Hampton Research, Qiagen, or Emerald BioSystems), followed by extensive manual optimization. The best single crystals were grown at 18°C by the hanging-drop vapor-diffusion method in a 1∶1 (v/v) ratio of protein and reservoir, as follows. (1) HA70 was crystallized with a reservoir solution composed of 0.1 M sodium acetate (pH 4.4) and 1.5 M ammonium chloride. Carbohydrate complexes were obtained when HA70 crystals were soaked in the mother liquor supplemented with 100 mM α2,3-SiaLac, α2,6-SiaLac, or Neu5Ac at 18°C overnight. (2) The HA17–HA33 complex was crystallized using a reservoir of 0.1 M MES (pH 6.2), 0.1 M MgCl2, and 5% (w/v) PEG [poly(ethylene glycol)] 8K. Micro-seeding was used to improve crystal quality. Carbohydrate complexes were obtained when crystals of the HA17–HA33 complex were soaked with 100 mM Gal, Lac, or LacNAc at 18°C overnight. (3) The HA70D3–HA17 complex was crystallized using 0.1 M sodium acetate (pH 4.8), 12% (w/v) PEG MME 2K, and 0.1 M CsCl. (4) The HA70D3–HA17–HA33 complex was crystallized using 0.1 M Tris (pH 8.2), 0.1 M NaCl, and 6% (w/v) PEG 20K.
The crystals of HA70 and its carbohydrate complexes were cryoprotected in their original mother liquor supplemented with 20% (v/v) ethylene glycerol and flash-frozen in liquid nitrogen. Crystals for all the other samples were cryoprotected in 22% (v/v) glycerol with their mother liquors and flash-frozen in liquid nitrogen. X-ray diffraction data were collected at the Stanford Synchrotron Radiation Lightsource (SSRL) or Advanced Photon Source (APS). The data were processed with HKL2000 [57] or iMOSFLM [58]. Data collection statistics are summarized in Tables S2 and S4 in Text S1.
The structure of HA70 of BoNT/A was determined by molecular replacement software Phaser [59] using the HA70 of BoNT/C (PDB code 2ZS6) [15] as the search model. The D3 domain of HA70 of BoNT/A, together with HA17 of BoNT/D (PDB code 2E4M) [14] and HA33 of BoNT/A (PDB code 1YBI) [17], were used as the search models to solve the structure of the HA70D3–HA17–HA33 complex by Phaser. The structures of the HA17–HA33 and the HA70D3–HA17 complexes were determined by Phaser using partial structures of the HA70D3–HA17–HA33 complex as search models.
The manual model building and refinements were performed in COOT [60] and PHENIX [61] in an iterative manner. The carbohydrates were modeled into the corresponding structure during the refinement based on the Fo-Fc electron density maps. The refinement progress was monitored with the free R value using a 5% randomly selected test set [62]. The structures were validated through the MolProbity web server [63] and showed excellent stereochemistry. Structural refinement statistics are listed in Tables S2 and S4 in Text S1. The coordinate and diffraction data for all the structures reported here will be deposited in the Protein Data Bank. The conformational change of HA33 was measured by DynDom [64]. All structure figures were prepared with PyMol (http://www.pymol.org).
The calorimetry titration experiments were performed at 23°C on an ITC200 calorimeter from Microcal/GE Life Sciences (Northampton, MA). The HA samples were used as the titrand in the cell and the carbohydrates were used as titrants in the syringe. To control for heat of dilution effects, protein samples were dialyzed extensively against the titration buffer (50 mM Tris, pH 7.6, and 100 mM NaCl) prior to each titration. Carbohydrates and nLoop peptide were dissolved in the same buffer. The pH of the acidic Neu5Ac solution was carefully adjusted to pH 7.6. The following concentrations were used for pair-wise titrations: HA33 (200 µM) vs. carbohydrates (Gal, Lac, LacNAc, IPTG, or α2,6-SiaLac) (50 mM); HA70D3 (200 µM) vs. α2,3-, or α2,6-SiaLac (40 mM); HA70D3 (160 µM) vs. Neu5Ac (80 mM); and HA70 (30 µM) vs. nLoop (400 µM). The data were analyzed using the Origin software package provided by the ITC manufacturer. The thermodynamic values reported are the average of three independent experiments (Table S3 in Text S1).
The recombinant HA70–HA17–HA33 complex, HA70, the HA17–HA33 complex, and the M-PTC were subjected to limited proteolysis with trypsin and pepsin overnight at room temperature. The trypsin digestions were performed at two different pHs in buffers containing 50 mM sodium phosphate (pH 6.0 or 7.5) and 300 mM NaCl, or in the Krebs-Ringer's solution (119 mM NaCl, 2.5 mM KCl, 1.0 mM NaH2PO4, 2.5 mM CaCl2, 1.3 mM MgCl2, 20 mM Hepes, and 11 mM D-glucose). The trypsin∶sample ratios (w/w) were 1∶10 (pH 6.0) or 1∶20 (pH 7.5). The digestions were stopped by adding 1 mM PMSF and boiling the samples in reducing SDS-loading buffer for 10 minutes. The pepsin digestions were performed at a 1∶100 ratio (w/w) of pepsin∶sample in a buffer containing 50 mM citrate acid (pH 2.6, an optimal pH for the pepsin reaction) and 300 mM NaCl. Pepsin digestions were terminated by addition of a 1 M Tris-HCl (pH 8.0) stock solution to give a final concentration of 200 mM and samples were then boiled in the reducing SDS-loading buffer. All samples were subjected to SDS-PAGE.
Cell culture: Caco-2 cells were obtained from the German Cancer Research Center (Heidelberg, Germany). Cells were cultured in Dulbecco modified Eagle medium (DMEM, Gibco® | Life Technologies, Darmstadt) supplemented with 10% fetal bovine serum, 100 U of penicillin per ml, and 100 mg of streptomycin per ml for up to six months. The cells were subcultured twice a week and seeded on BD Falcon Cell Culture Inserts (#353494, growth area 0.9 cm2, pore size 0.4 µm) at a density of approximately 105 cells cm−2 for flux studies and determination of transepithelial electrical resistance (TER).
Measurement of TER: All TER experiments were conducted in 0.5 ml and 1.5 ml of Iscoves Modified Dulbeccos Medium without phenol red (IMDM, Gibco® | Life Technologies, Darmstadt) in the apical and basolateral reservoir, respectively. TER was determined with an epithelial volt-ohm meter (World Precision Instruments, Berlin, Germany) equipped with an Endohm 12 chamber for filter inserts. Filters with cell monolayers were used at day 11 after seeding which is seven days of post confluency. Only filters with an initial resistance of ≥300 Ω cm−2 were used. For analysis of independent experiments subsequent results were expressed as percentages of the corresponding resistance of each data set determined immediately after administration of samples. Values are expressed as means of ≥3 independent experiments with duplicate samples ± standard deviations.
Carbohydrate inhibition assays: Lac, Gal, IPTG, Neu5Ac, α2,6- and α2,3-SiaLac were dissolved in IMDM, sterile filtered and stored at −20°C. Neu5Ac stock solution was adjusted to pH 7.4. The wild type HA complex (HA wt), fluorescence-labeled HA complex (HA*), or the HA70-TPRA complex were pre-incubated with the corresponding carbohydrate over night at 4°C in IMDM and diluted to the final concentration with IMDM prior to administration. The TER upon administration of each carbohydrate in the highest concentrations used was checked in the absence of HA and was virtually identical to that of the control without sugars.
Transport measurement: For paracellular transport studies, filters were incubated in IMDM added to the apical (0.5 ml) and basolateral (1.5 ml) reservoir. As marker substance Alexa Fluor® 488 labeled HA* or HA33-DAFA* was administered to the apical or basolateral reservoirs at final concentrations of 58 nM and 17 nM, respectively. After 24 hour of incubation, 200 µl of samples were taken from the apical and the basolateral reservoir. The marker substance was measured in a BioTek Synergy 4 fluorescence spectrophotometer at 495 nm excitation and 519 nm emission wavelengths.
The mouse protection assay was performed following a previously described protocol [4]. Briefly, random sets of 10–20 female Swiss Webster mice (20–23 g) were used per dose. Mice were treated by oral gavage with 100 µl containing 1.9 µg of L-PTC/A (Metabiologics) in phosphate–gelatin buffer (10 mM phosphate buffer, pH 6.2, and 2% gelatin), with or without the indicated concentrations of IPTG, Neu5Ac, or Gal. Mice were also administered 100 µl of 500 mM IPTG by gavage 1 hour prior or after treatment with 100 µl containing 1.9 µg of L-PTC/A in phosphate gelatin buffer. The acidic Neu5Ac was adjusted to pH 6.2 for administration. Mice were monitored for botulism symptoms for up to 14 days post-intoxication. Median survival and p-values were determined with the GraphPad Prism 5 program (San Diego, CA).
Atomic coordinates and structure factors for HA70, HA70D3–HA17–HA33, HA70D3–HA17, HA17–HA33, HA17–HA33–Lac, HA17–HA33–Gal, HA17–HA33–LacNAc, HA70–α2,3-SiaLac, and HA70–α2,6-SiaLac have been deposited with the Protein Data Bank under accession codes 4LO4, 4LO7, 4LO8, 4LO0, 4LO2, 4LO1, 4LO3, 4LO5, 4LO6, respectively. EM 3D reconstructions for the L-PTC and the HA complex have been deposited with the Electron Microscopy Data Bank (EMDB) under accession codes EMD-2417 and EMD-2416, respectively.
|
10.1371/journal.ppat.1000769 | The Syk Kinase SmTK4 of Schistosoma mansoni Is Involved in the Regulation of Spermatogenesis and Oogenesis | The signal transduction protein SmTK4 from Schistosoma mansoni belongs to the family of Syk kinases. In vertebrates, Syk kinases are known to play specialized roles in signaling pathways in cells of the hematopoietic system. Although Syk kinases were identified in some invertebrates, their role in this group of animals has not yet been elucidated. Since SmTK4 is the first Syk kinase from a parasitic helminth, shown to be predominantly expressed in the testes and ovary of adult worms, we investigated its function. To unravel signaling cascades in which SmTK4 is involved, yeast two-/three-hybrid library screenings were performed with either the tandem SH2-domain, or with the linker region including the tyrosine kinase domain of SmTK4. Besides the Src kinase SmTK3 we identified a new Src kinase (SmTK6) acting upstream of SmTK4 and a MAPK-activating protein, as well as mapmodulin acting downstream. Their identities and colocalization studies pointed to a role of SmTK4 in a signaling cascade regulating the proliferation and/or differentiation of cells in the gonads of schistosomes. To confirm this decisive role we performed biochemical and molecular approaches to knock down SmTK4 combined with a novel protocol for confocal laser scanning microscopy for morphological analyses. Using the Syk kinase-specific inhibitor Piceatannol or by RNAi treatment of adult schistosomes in vitro, corresponding phenotypes were detected in the testes and ovary. In the Xenopus oocyte system it was finally confirmed that Piceatannol suppressed the activity of the catalytic kinase domain of SmTK4. Our findings demonstrate a pivotal role of SmTK4 in gametogenesis, a new function for Syk kinases in eukaryotes.
| Parasitic blood flukes of the genus Schistosoma cause schistosomiasis, one of the most important infectious diseases for humans and animals worldwide. Besides their medical importance, schistosomes possess unique biological features. Among these is the sexual maturation of the female, which requires a constant pairing contact with the male. Pairing induces mitogenic activity and differentiation processes in the female that lead to gonad development. This is a prerequisite for egg production, which is closely connected with the pathological consequences of the disease since the eggs are trapped in different host organs, inducing inflammatory processes. Although these correlations are long known, the molecular basis of differentiation processes in female gonads are poorly understood. In the context of identification of signal transduction proteins controlling female reproductive development we identified SmTK4 of S. mansoni, the first Syk-family kinase of a parasite. By using biochemical and molecular approaches in combination with in vitro culture and a novel microscopical technique, we demonstrate in this study the pivotal role of the signaling protein SmTK4 in spermatogenesis and oogenesis of S. mansoni. This is a new attribute for Syk kinases of eukaryotes promoting SmTK4 as a candidate target for blocking transmission and disease progression.
| Helminth parasites of the genus Schistosoma are the causative agents of schistosomiasis, one of the most prevalent parasitic diseases for humans and animals worldwide [1],[2]. More than 200 million people suffer from the pathological consequences of this disease, which originate from the massive egg production of schistosomes. The eggs cause inflammatory reactions in the gut, bladder, spleen and liver leading to granuloma formation [1],[3]. Praziquantel is the only drug applicable to all schistosome species and is commonly used to treat patients, but treatment does not prevent reinfection. In the light of the absence of a vaccine and the probability of emerging resistance, a search for alternative treatments is a commonly accepted need for further research [4],[5]. In this respect great international efforts are ongoing to analyze the genome of this blood fluke, its transcriptome, proteome, and glycome [6]–[10].
Besides their medical importance, schistosomes exhibit a nearly unique biological phenomenon–the pairing-dependent induction and maintenance of the sexual maturation of the female. During a constant pairing contact, the male activates signal transduction pathways in the female leading to the proliferation and differentiation of cells in the reproductive organs, such as, the ovary and vitellarium [11]–[14]. This is a prerequisite for the female to produce about 300 eggs each day [15]. One half reaches the outside of the definitive host to deliver miracidia continuing the life cycle. The remaining eggs are deposited in the host tissue causing pathogenesis. An egg from a mature female consists of one fertilized oocyte, originating in the ovary, and 30–40 surrounding vitelline cells produced in the vitellarium. Since growth and differentiation of vitelline cells and oocytes are probably controlled by signal transduction pathways, efforts have been made to identify and characterize the participating molecules.
In the last decade, several genes encoding for signaling molecules from S. mansoni have been identified, some of which were found to be specifically or predominantly expressed in reproductive organs [reviewed in 16,17]. In contrast to the vitellarium, however, less is known about signaling molecules in the ovary. Among the molecules shown to be predominantly expressed in this organ is SmTK4, a member of the Syk (spleen tyrosine kinase) tyrosine-kinase family [18]. Syk kinases are characterized by a tandem Src-homology 2 (SH2) domain and a catalytic tyrosine kinase (TK) domain. Genome-project data have indicated that Syk kinase genes are absent in Caenorhabditis elegans, and in Drosophila melanogaster only the related kinase Shark (SH2 domain ankyrin repeat kinase; [19]) is present, which had suggested a recent evolutionary origin of kinases from the Syk family. However, Syk kinases were found in Hydra vulgaris as well as in sponge [20], and with SmTK4 also in the parasitic helminth S. mansoni.
In mammals, Syk kinases are expressed in hematopoietic cells playing well-characterized roles in inflammatory processes operating as downstream signaling molecules of immunoreceptors [21]. In the last years, evidence has accumulated for functions of Syk kinases in different signal transduction pathways also in non-hematopoietic cells [22]. Syk kinases regulate proliferation, differentiation, morphogenesis, and survival of epithelial [23],[24], endothelial [25], and neuronal cells [26]. In the hematopoietic system, Syk kinases interact with immune and antigen receptors lacking intrinsic catalytic activity [27]. The tandem-like structure of the SH2 domains confers higher binding specificity of Syk kinases to phosphorylated tyrosine residues of upstream interaction partners compared to individual SH2 domains [28]. Following receptor activation, each SH2 domain interacts with one immunoreceptor tyrosine-based activation motif (ITAM) in the intracellular part of the receptor leading to a conformational change in Syk accompanied by an increase in its enzymatic activity [29]. In SmTK4 the conserved sequence within the SH2 domains responsible for this binding is absent, suggesting that this Syk kinase interacts with molecules without ITAMs. Binding of upstream partners stimulates autophosphorylation of Syk on tyrosines within the activation loop, which influences kinase activity or creates docking sites for SH2-containing proteins [30]. The phosphorylation of Syk can be enhanced by interacting Src (Rous sarcoma virus kinase) tyrosine kinases [27]. In addition, a variety of other signaling and adaptor molecules have been reported to associate with Syk kinases, but the relevance of these interactions have not been elucidated yet [27].
With respect to the very specialized function of Syk kinases in the hematopoietic system of mammals, the existence of a schistosome homolog was unexpected. SmTK4 was found to be transcribed in the larval stages as well as adults, independently from the pairing-status. Localization studies had demonstrated its predominant expression in the testes of the male and ovary of the female, but not in the vitellarium [18]. This narrow expression profile contrasts with other identified cellular kinases in adult schistosomes, such as the Src kinases SmTK3 and SmTK5, whose expressions were demonstrated in all reproductive organs and other tissues [31],[32].
Recently, an interaction of the Src kinase SmTK3 with SmDia (Diaphanous-related formin) and SmRho1 (Ras-homologous GPTase) was shown by yeast two-hybrid (Y2H) analyses and localization studies, indicating a role of SmTK3 in cytoskeleton organization processes in the gonads of adult S. mansoni [33]. To elucidate the potential function of SmTK4 in adult schistosomes, two different strategies were followed in this study. First, by isolating signaling molecules acting up- and downstream of SmTK4 we expected to discover interacting proteins, whose identity could provide evidence for functionally conserved signaling pathways. Using a recently established S. mansoni Y2H cDNA-library of females and males [33] and different constructs of SmTK4 as probes, several interacting proteins were identified, such as SmTK3 and a novel schistosome Src kinase (SmTK6), a MAPK (mitogen-activated protein kinase)-activating protein, and mapmodulin. Second, inhibitor and RNA interference (RNAi) experiments were performed to functionally knock-down SmTK4 activity. With the Syk kinase-specific inhibitor Piceatannol [34],[35] or SmTK4-specific dsRNAs for RNAi, significant morphological changes in the testes and ovary of treated schistosomes were observed using carmine red-staining and confocal laser scanning microscopy (CLSM). In Xenopus oocytes it was finally shown that Piceatannol is able to suppress the catalytic activity of the TK domain of SmTK4. Taken together, our results in S. mansoni substantiate a pivotal role for the Syk kinase SmTK4 in gametogenesis, a function which has not yet been shown for a Syk kinase in other eukaryotes.
To identify upstream binding partners of the schistosome Syk kinase SmTK4, an adult stage Y2H cDNA-library [33] was screened with the bait construct encoding the tandem SH2-domain of SmTK4 as a fusion protein with the Gal4-BD (GAL4 DNA-binding domain) and the TK domain of SmTK3 (SmTK4-SH2SH2 + SmTK3-TK pBridge; Figure 1). The schistosome TK domain was co-expressed in a kind of yeast three-hybrid (Y3H) approach to ensure the phosphorylation of tyrosine residues of potential interaction partners, since yeast does not possess specific, endogenous tyrosine-kinase activity [36]. Phosphorylation, however, is decisive for Syk-kinase interactions, because only phosphorylated tyrosine residues of binding partners acting upstream in a signaling hierarchy are favored binding sites for the pocket-like structure of the tandem SH2-domain of Syk kinases [28],[37]. This principle has been successfully applied before to investigate the interaction between SH2 domain-containing proteins and tyrosine-containing substrates [38],[39]. Without additional protein-protein interacting domains such as SH2 or SH3, TK domains function promiscuously [40]. Therefore, we expected that tyrosine residues of yeast and library proteins were phosphorylated as a consequence of the expression of an individual TK domain. Indeed, Western-blot experiments with anti-phosphotyrosine antibodies have shown that tyrosine phosphorylation of yeast proteins was enhanced when the SmTK3 TK-domain was expressed (Phillip, unpublished). The expression of both baits (SmTK4-SH2SH2 and SmTK3-TK) was confirmed at the transcriptional level by RT-PCR analyses using total RNA extracts from transformed yeast cells (results not shown). Screening of the Y2H cDNA-library resulted in the identification of 77 initial prey clones, which underwent growth selection and β-galactosidase (β-Gal) filter assays. After isolation of the prey plasmids, sequence analyses of 14 remaining clones were performed by BlastX to unravel their identity (Table 1). Seven clones represented partial sequences from a schistosome homolog of a nonsense mRNA reducing factor (NORF1) from Mus musculus (accession number AAK08652). Three clones encoded proteins with homology to a dipeptidyl peptidase III from Mus musculus (accession number NP_598564). Two clones showed similarity to the schistosome Src/Fyn kinase SmTK5 ([32]; accession number AAF64151) and encoded a novel schistosome cellular tyrosine kinase (CTK), named SmTK6 (accession number FN397679). Finally, two clones were identical to the Src kinase SmTK3 from S. mansoni, identified and characterized in a previous study ([31]; accession number CAE51198).
To confirm their interactions and to determine their relative binding strength in a comparative approach, yeast cells (strain AH109) were transformed with individual prey plasmids together with the original bait construct SmTK4-SH2SH2 + SmTK3-TK pBridge. Additionally, yeast cells were transformed with the prey plasmids and the bait construct SmTK4-SH2SH2 pBridge to analyze the dependence of the observed interactions on the additional tyrosine phosphorylation. Finally, the prey plasmids were used for transformation together with bait constructs that contained only one SH2 domain of SmTK4 (SmTK4-SH2(1) or SmTK4-SH2(2) pBridge) to test if the interactions depended on the presence of the tandem SH2-domain structure. After re-transformation of the prey plasmids with the bait constructs containing the tandem SH2-domain of SmTK4, with or without the additional expression of the SmTK3 TK-domain, all yeast clones survived growth- (Trp−/Leu−/Ade−/His−) and color selection (β-Gal filter assays) thus confirming the observed interactions. However, when transformation was done with the bait plasmids containing only single SH2 domains of SmTK4, the yeast clones containing the prey plasmids encoding the Src kinases SmTK6 or SmTK3 were no longer able to grow. This indicated that Src/Syk-kinase interaction depended on the presence of the tandem SH2-domain structure. The other clones survived the selection procedure, which is a hint for unspecific interactions (Table S1).
To quantify the relative strengths of interaction β-Gal liquid assays were performed. To this end yeast cells were used, which were transformed with representative prey plasmids encoding for each potential binding partner and the bait plasmid SmTK4-SH2SH2 pBridge, or SmTK4-SH2SH2 + SmTK3-TK pBridge. These experiments again confirmed the observed interactions with SmTK4. Furthermore, the results demonstrated the strongest interaction between the SmTK4 tandem SH2-domain and the novel Src kinase SmTK6. This interaction was increased by a factor of five when the SmTK3 TK-domain was present indicating the significant influence of additional phosphorylation (Figure 2). The interaction between the SmTK4 tandem SH2-domain and the Src kinase SmTK3 was considerably weaker, but also enhanced by additional phosphorylation, although to a minor degree. The strength of interactions between the homologs of the nonsense mRNA reducing factor (NORF1) and the dipeptidyl peptidase III were also weak and not, or only slightly, influenced by the state of phosphorylation (Figure 2).
Since the novel Src kinase SmTK6 was the strongest upstream interaction partner of SmTK4 found in this screening, we carried out a preliminary characterization. The inserts of both SmTK6 prey clones were equal in size and about 1360 bp long. Comparisons by multiple alignment analysis revealed homology to CTKs and indicated that part of the 5′-region of the cDNA was missing in the isolated prey plasmids. Using schistosome sequencing data (www.sanger.ac.uk), the whole coding sequence was identified in silico. To verify its existence in our S. mansoni strain, we performed RT-PCR analyses using the primer pair TK6-fl-5′ (5′-CTCATTATGGGAATTTGTTTGTG-3′ containing the ATG codon) and TK6-fl-3′ (5′-AATTATCTAAATATTGAGCTTCTG-3′ containing the TAA stop codon), and total RNA from mixed-sex worms. Amplification products of the expected size were obtained and cloned. Sequence analysis showed that the complete cDNA sequence of SmTK6 is 1698 bp long encoding a protein of 566 amino acids (accession number FN397679). In silico analyses indicated the presence of one SH2, one SH3, and a TK domain that characterize CTKs of the Src family. Preliminary phylogenetic analyses showed that SmTK6 has also some similarity to the class of Abl kinases indicating that this kinase is a kind of Src/Abl intermediate (Beckmann, unpublished).
To additionally confirm the binding capacity of SmTK6 and SmTK4, co-immunoprecipitation (co-IP) experiments were performed. To this end the tandem SH2-domain of SmTK4 was cloned as a FLAG-tagged construct into one multiple cloning site of the pESC-His yeast expression vector. In the second multiple cloning site of the same vector a nearly complete version of SmTK6 was cloned as a cMyc-tagged fusion. Following yeast transformation and subsequent growth selection, protein expression was induced by galactose. The expression of the recombinant proteins was proven by Western-blot analyses (results not shown). Following co-IP, the presence of the FLAG- and cMyc-tagged schistosome proteins was finally confirmed by Western-blot analyses (Figure 3). Bands of the predicted sizes of 31 kDa (SmTK4-SH2SH2) and 55 kDa (SmTK6) were detected only in lysates precipitated by each of the antibodies, not in a control precipitated without antibody (Figure 3B). This result confirmed that the tandem SH2-domain of SmTK4 is able to bind to SmTK6 independent of additional yeast Gal4-AD/BD (GAL4 activating domain/DNA-binding domain) fusion protein partners.
By in situ hybridization, finally, SmTK6 was localized in the parenchyma of both genders, in the testes of the male and the ovary of the female (Figure 4, A–C). From the staining pattern obtained we cannot exclude that SmTK6 is also transcribed in the vitellarium. This testes- and ovary-preferential transcription pattern corresponded to that of SmTK4 [18], additionally supporting the conclusion that these kinases interact.
Downstream binding partners of Syk kinases are known to interact with the linker region or with the TK domain of a Syk kinase [41]. Furthermore, binding can be influenced by tyrosine residues within the TK domain, which are phosphorylated in trans by, e.g., Src kinases [27]. For these reasons we used SmTK4-linker+TK + SmTK3-TK pBridge as the bait construct for library screening to identify binding partners acting downstream of SmTK4 (Figure 1). This bait construct expressed the complete linker region together with the TK domain of SmTK4 fused to the Gal4-BD (multiple cloning site I, MCS I) as well as the TK domain of SmTK3 (multiple cloning site II, MCS II) to ensure tyrosine phosphorylations of bait proteins (Y3H).
The expression of the bait constructs SmTK4-linker+TK (+ SmTK3-TK) pBridge, SmTK4-linker pBridge, and SmTK4-TK pBridge was confirmed by RT-PCR analyses using total RNA extracts of the transformed yeast cells and appropriate primers (results not shown). The mating with the S. mansoni Y2H library resulted, after subsequent growth selection and β-Gal filter assays, in 19 clones as potential candidates for interaction. The appropriate prey plasmids were isolated and their inserts sequenced. By BlastX analyses the identities of these potential binding partners were unraveled due to their homology to proteins in the NCBI-database (Table 2). Three clones showed homology to small heat-shock proteins (HSPs) from Caenorhabditis elegans (accession number NP_001024376). One clone encoded a related but not identical protein, which was also homologs to small HSPs (accession number NP_001024376), as well as to the major egg antigen Smp40 from S. mansoni (accession number P12812). Two prey clones encoded a protein with homology to caspase 3 from Homo sapiens (accession number AAP36827). The inserts of five further clones encoded homologs of the MAPK-activating proteins PM20/PM21 from different organisms, such as Homo sapiens (accession number NP_001108072). One clone represented a leucine-rich protein from S. mansoni (accession number Q86QS6), which exhibited homology to mapmodulin from Drosophila melanogaster (accession number NP_001097361). Seven clones showed the same insert sequence encoding a protein with no significant homology to any known proteins from other organisms.
β-Gal liquid assays were performed to quantify the relative binding strengths of SmTK4 and its identified downstream interaction partners. To this end yeast cells were transformed with representative prey plasmids of the groups A–D together with the original bait construct SmTK4-linker+TK + SmTK3-TK pBridge to confirm the interactions and, additionally, with SmTK4-linker+TK pBridge to analyze the dependency of interactions on the supplementary tyrosine phosphorylation. Furthermore, to determine whether the linker region and/or TK domain of SmTK4 were responsible for binding, yeast cells were transformed with individual prey plasmids and further bait constructs containing only the linker region (SmTK4-linker pBridge) or only the TK domain of SmTK4 (SmTK4-TK pBridge). The performed assays again confirmed the observed interactions with SmTK4 and, additionally, demonstrated significant differences in the relative binding strengths. The interactions between the small HSP and caspase 3 homologs with the different fragments of SmTK4 were very weak. With respect to the different SmTK4 fragments, the small HSP showed the strongest binding to the isolated linker region of SmTK4, whereas the caspase 3 homolog showed no significant differences in the relative binding strengths (Figure 5). However, the interactions of the MAPK-activating protein (PM20/PM21) and the mapmodulin with SmTK4 were stronger and differed between the SmTK4 fragments (Figure 5). In the case of the MAPK-activating protein, the interaction with the combined linker region and TK domain of SmTK4 was slightly enhanced by additional phosphorylation. However, with the isolated linker region, the interaction was nearly seven times stronger. In contrast, the interaction with the TK domain was very weak (Figure 5). These results showed that the MAPK-activating protein bound presumably to the linker region of SmTK4, and that this binding was negatively influenced by the presence of the TK domain. Furthermore, binding to the linker region was strong even without additional phosphorylation. The interaction between mapmodulin and the combined linker and TK domain of SmTK4 was also slightly enhanced by the additional phosphorylation, as in the case with the MAPK-activating protein. Regarding the isolated linker region or TK domain, mapmodulin bound strongly to both with a slight bias towards the TK domain (Figure 5). Again, interaction was strong even without additional phosphorylation. As in the case with the MAPK-activating protein, the interactions of mapmodulin with the linker region or the TK domain were negatively influenced by the presence of both domains, indicating a potential inhibitory intramolecular conformation of these domains when expressed as a fusion protein.
For the two strongest downstream binding partners of SmTK4, MAPK-activating protein and mapmodulin, colocalization with SmTK4 in the ovary of the female and testes of the male was shown by in situ hybridizations (Figure 4, D–I) supporting the conclusion for interaction. Corresponding to the SmTK4 transcriptional profile [18], both binding partners were found to be transcribed also in parenchyma of both genders, but not in the vitellarium of the female.
Recently, we introduced an experimental approach to investigate the role of schistosome CTKs by specific inhibitors. Using the Src kinase-specific inhibitor Herbimycin A for treatment of schistosome couples in vitro, we demonstrated that mitogenic activity and egg production were significantly reduced in females [42]. This was correlated with the reduced stability of the Src kinase SmTK3 from S. mansoni. Here we have combined the inhibitor approach with a novel way of phenotypic analysis at the morphological level. Based on the procedure established by Machado-Silva et al. [43] and Neves et al. [44], we investigated inhibitor effects by treatment of schistosome couples in vitro with subsequent fixation and carmine red-staining. By CLSM finally, we looked for morphological effects with a focus on the testes of the male and the ovary of the female, because SmTK4 expression has been detected mainly in these organs [18]. To investigate whether the inhibitors influence egg production of paired females, the numbers of eggs were counted daily during the treatment period.
Towards this end schistosome couples were treated in vitro with the Syk kinase-specific inhibitor Piceatannol (3,4,3′,5′-tetrahydroxy-trans-stilbene). This phenolic stilbenoid inhibits the activity of Syk kinases in cell culture with an IC50 value of 10 µM for the human Syk kinase [45]. For tissues it is used at higher concentrations of 100–200 µM [34].
For S. mansoni couples maintained in vitro, we used Piceatannol at concentrations of 35 µM, 70 µM, and 100 µM over a period of six days to investigate dosage- and time-dependent effects. The medium was changed daily along with the inhibitor, and the viability of the worms as well as their pairing stability was examined. During this time period, no alterations in behavior, mortality rates, or worm pairing compared with DMSO (dimethyl sulfoxide)-treated controls could be observed (results not shown). Each day, an aliquot of treated worm couples was fixed in AFA, stained with carmine red and analyzed by CLSM. The control males and females, treated for the same time with DMSO, showed no morphological changes in the testes or ovaries (Figure 6, A–B) compared to completely untreated schistosomes ([43],[44], results not shown). The testes of untreated or DMSO-treated adult schistosome males are composed of several testicular lobes containing numerous spermatogonia and spermatocytes in different stages of maturation. Maturation of spermatocytes (spermatogenesis) begins in the dorsal part of the lobes with big round spermatogonia and ends in the ventral part with smaller elongated mature sperms (spermatozoa, Figure 6A). In the ventral part of the testicular lobes and in the vas deferens elongated mature sperms can be detected as well as in the anterior sperm vesicle, which is full of sperms (Figure 6A, arrows and asterisk). During treatment with 70 µM Piceatannol, however, this morphology changed considerably. After two days the size of the lobes was already slightly reduced, and the number of spermatocytes diminished (results not shown). After six days these effects were even more dramatic. The testicular lobes were shrunk accompanied by a significant decrease in the number of spermatocytes per lobe (Figure 6C). In the ventral part of the testicular lobes and in the anterior sperm vesicle, no elongated mature sperms were detected in most of the males (Figure 6C, asterisk). Instead of mature sperms, the sperm vesicle contained, in several cases, undifferentiated round spermatocytes. In addition to the effects within the testes, the inhibitor also caused morphological changes in the ovary. In untreated or DMSO-treated mature females the ovary is composed of small oogonia and immature oocytes in the anterior part and larger primary oocytes in the posterior part ([46], Figure 6B). After treatment with 70 µM Piceatannol for six days, in most of the females the number of large primary oocytes was clearly increased compared to the small, immature oocytes. Furthermore, the high number of large oocytes was distributed all over the ovary and no longer concentrated to the posterior part (Figure 6D). Using the lower concentration of 35 µM Piceatannol for worm treatment led to similar morphological changes in the testes and ovary, although with a delay of 1–2 days. A higher concentration (100 µM) led to the same phenotypes, but in a shorter time period indicating a time- and dosage-depending effect of this inhibitor on schistosomes (results not shown). Although at a comparatively low level, SmTK4 is also transcribed in subtegumental and parenchyma tissues in adults [18]. However, in these tissues we did not see an effect, which may be due to their low mitogenic activity compared to the gonads.
To analyze the influence of Piceatannol on the egg production of paired females, the number of eggs was determined for treated couples maintained in vitro. 70 µM Piceatannol reduced the number of eggs per couple within 7 days to 51% compared to the DMSO control (Figure 7).
Since obvious morphological changes after Piceatannol treatment were observed, the question arose, whether these effects could be ascribed to the inhibitor's effect on SmTK4. To answer this question, evidential and experimental data were assembled pinpointing SmTK4 as the target of Piceatannol. First, Southern-blot analysis had shown that SmTK4 is a single copy gene [18], and searches through the schistosome genome data ([6], assembly version 3.1; www.sanger.uk) provided no evidence for the existence of other kinases of the Syk class. Therefore, we assume that SmTK4 is the only Syk kinase present in S. mansoni. Since at the concentrations used Piceatannol is specific for Syk kinases [34],[35], we concluded that SmTK4 is the only molecule targeted in S. mansoni. Furthermore, localization studies had indicated the testes and ovary as the only reproductive organs transcribing SmTK4, which is not transcribed in the vitellarium. In this organ, no morphological changes following inhibitor treatment were observed (result not shown).
To specifically attribute the effect of Piceatannol to SmTK4, we post-transcriptionally inhibited SmTK4 using dsRNAs (RNAi). To this end, SmTK4-specific dsRNAs were generated spanning both SH2 domains, including the interdomain A (linker region between the SH2 domains, Figure 1; 813 bp long). To exclude nonspecific side-effects on other tyrosine kinases, SmTK3-specific dsRNAs were also generated and applied as control. Based on the protocol from Ndegwa et al. [47], ten S. mansoni couples were electroporated with a single square-wave impulse and 25 µg of SmTK4- or SmTK3-specific dsRNAs, respectively. As an additional control, worm couples were electroporated under the same conditions, but without dsRNA.
To investigate the RNAi effects at the molecular level, we analyzed SmTK4 transcript levels in the three independent groups of electroporated worms by semi-quantitative RT-PCR (Figure S1). As endogenous standard, the housekeeping gene SmPDI (protein disulfide isomerase; [48]) was used. Five days after electroporation, total RNA was extracted from one half of the worm couples and the SmTK4 transcript level determined by RT-PCR analysis. The products were analyzed by agarose gel electrophoresis, and the intensities of the amplification products were densitometrically quantified using RT-PCR results of SmPDI for normalization. In three independent experiments, the SmTK4-mRNA levels were significantly reduced, although at different levels. In worms electroporated with SmTK4-dsRNA, transcript levels decreased to 10–32% compared to control worms, which had been electroporated without dsRNA. Worm couples treated with SmTK3-specific dsRNAs showed no alterations in the SmTK4-transcript level, indicating the specific silencing effect of the SmTK4-dsRNAs (Figure S1).
Finally, by carmine red-staining and CLSM the morphology of those dsRNA-treated worms was investigated, which revealed the strongest silencing effect (transcript reduction to 10%) according to RT-PCR analysis. Control worms, electroporated without dsRNA, showed no alterations in the morphology of the testes or ovary (Figure 6E–F). Additional control worms electroporated with SmTK3-specific dsRNA also showed no morphological alterations in these tissues (results not shown). However, many of the SmTK4-dsRNA electroporated worm couples exhibited phenotypes that were qualitatively comparable to the phenotypes observed after Piceatannol treatment (Figure 6G–H). In males, the size of the testicular lobes and the number of spermatocytes were reduced, and nearly no mature elongated sperms were detected in the ventral part of the lobes or within the sperm vesicle, but some immature spermatogonia (Figure 6G, asterisk). In the ovary of the female, the number of mature oocytes was increased, but not to the same extent as observed after Piceatannol treatment (Figure 6H).
To finally test, whether SmTK4 exerts signal transduction activities which can be suppressed by Piceatannol, we made use of the Xenopus oocyte system [49]. Previous studies had shown that signal transduction proteins of schistosomes can be efficiently expressed in Xenopus oocytes. Moreover, parasite kinase activities were already studied in this heterologous system owing to their capacity to induce resumption of meiosis or germinal vesicle breakdown (GVBD) [50],[51]. To analyze the GVBD-inducing capacity of SmTK4 and inhibitor effects on this kinase, two constructs were cloned as templates for cRNA synthesis, a full-length construct and a shortened version of SmTK4 containing only the catalytic TK domain. Following cRNA injection, GVBD was studied by the appearance of a white spot at the center of the animal pole of the oocyte. The results showed that in the absence of progesterone, a steroid inducer of GVBD [49] used for positive control, the TK domain of SmTK4 was able to induce 100% GVBD. This confirmed its catalytic activity (Table S2). In non-injected oocytes, no GVBD was observed under the same conditions as well as in oocytes transfected with the full-length variant of SmTK4. This was probably due to a close conformation of the TK domain within the complete protein. When oocytes containing the catalytic TK domain of SmTK4 were incubated with increasing concentrations of Piceatannol, however, GVBD was negatively influenced in a concentration-dependent manner and completely suppressed at a concentration of 5 µM already. The inhibitor had no effect on non-injected oocytes, and did not inhibit progesterone-dependent maturation even when it was used at the concentration of 100 µM (Table S2). These results confirmed the specific action of Piceatannol on SmTK4.
SmTK4 was the first Syk kinase identified in a parasitic helminth and shown to be predominantly expressed in the testes and ovary of adults [18]. To elucidate signaling cascades in which SmTK4 participates, screenings of a S. mansoni adult stage Y2H library were performed. Using the tandem SH2-domain as bait, upstream binding partners of SmTK4 were identified such as homologs of a nonsense mRNA reducing factor, a dipeptidyl peptidase III, SmTK3, and SmTK6. Although most clones represented homologs of a nonsense mRNA-reducing protein factor (group A, [52]), binding tests with individual SH2 domains failed, indicating nonspecific interaction. Furthermore, interactions between Syk kinases and such factors are not known. Both arguments also apply to dipeptidyl peptidase III (group B, [53]), which represents another nonspecific interaction partner. The other two upstream interaction partners belong to the Src family of CTKs (groups C, D). One represented the already characterized SmTK3 [31], whilst the other was a novel Src kinase named SmTK6. In contrast to group A and B clones, the binding of both Src kinases occurred only with the tandem SH2-domain of SmTK4, and was enhanced by additional phosphorylation in yeast, in the case of SmTK6 by a factor of five. This confirmed the specificity of the detected Syk-Src interactions.
SmTK6 turned out to be the strongest upstream interaction partner, and in situ hybridizations demonstrated its transcription within the testes of the male and the ovary of the female. This corresponds to the transcription profiles of SmTK4, but also SmTK3, which among other tissues is also expressed in the testes and ovary [31]. In addition to the interaction analyses, the colocalization of SmTK6, SmTK3 and SmTK4 allows the conclusion that these kinases cooperate in the gonads. Src-Syk interactions are already known from other systems. In mammalian immune cells, Src kinases recruit downstream-acting Syk kinases to the plasma membrane to activate them by phosphorylation [54],[55]. Thus, high-affinity binding sites are created for molecules acting downstream of Syk [27] that bind to phosphotyrosine residues within the linker region, or the TK domain of Syk [27],[41],[56]. From these data we hypothesize that SmTK3 may recruit SmTK6 to the plasma membrane, where the latter one becomes phosphorylated. This may be a prerequisite for binding of SmTK4 by its SH2 domain. By co-IP we found further evidence for SmTK4-SmTK6 binding, which supports the sketched scenario. Future studies will aim to analyze these interactions in more detail.
As downstream interaction partners of SmTK4 homologs of caspase 3, small HSP, a PM20/PM21 type MAPK-activating protein, mapmodulin, and a protein with non-significant similarity were found. The caspase 3 homolog (group B, [57]) showed the weakest relative binding strength with all SmTK4 subfragments. Although for ZAP-70, the second member of the Syk-kinase family in mammals, an influence on caspase pathways has been described [58], this influence is indirect [59], and no direct interaction between Syk kinases and caspases has been described yet. In light of these facts, a nonspecific binding may have occurred. The relative binding strength of the small HSP homolog (group A) was slightly stronger as for caspase 3. Besides other functions [60], HSPs fulfill chaperone functions, and a protective role of HSP90 for ZAP-70 has been demonstrated [61]. Therefore, we assume that the schistosome HSP homolog may fulfill a chaperone-like function for SmTK4. Mapmodulin (group D), a microtubule-associated protein [62],[63], showed high relative binding strengths to both the linker region and the TK domain indicating that it is able to bind to both fragments of SmTK4. This suggests that SmTK4 may influence the reorganization of the cytoskeleton in spermatocytes or oocytes, which is supported by the in situ-hybridization data showing mapmodulin transcripts in the testes and ovary. Indeed, Syk kinases are known to phosphorylate substrates involved in cytoskeleton organization, or components of the cytoskeleton such as microtubules directly [23],[64],[65]. Finally, the MAPK-activating protein of the PM20/PM21 type (group C) showed the strongest relative binding activity to SmTK4. Since this group of proteins lacks defined structural or functional domains, the molecular basis for interaction as well as the function of the schistosome MAPK-activating protein remains uncertain [66],[67]. Results of our study provide first evidence for its binding to the linker region of SmTK4. MAPK signaling is acknowledged to be initiated by upstream molecules, such as growth factors and RTKs, which are known to be components of signaling cascades controlling proliferation, differentiation, and survival of cells. A critical role of Syk for MAPK activation has been postulated [68], although no direct interactions between Syk kinases and a MAPK-activating protein were shown yet. In situ-hybridization experiments colocalized transcripts of the schistosome MAPK-activating protein in the testes and ovary providing further evidence for interaction. This allows the speculation that SmTK4 may trigger a MAPK signaling-pathway in the gonads probably influencing cell proliferation. We tried to find evidence for MAPK activation using Western-blot analyses with antibodies directed against phosphorylated or non-phosphorylated forms of human MAP kinases and protein homogenates from mixed sex schistosomes. In another independent study, these antibodies showed a high specificity also for Echinococcus MAP kinases [69]. Although a highly conserved MAPK homolog exists in the genome of S. mansoni ([6], accession number XP_002575049), its structure may be different since the antibodies were not able to detect a band of the expected size (results not shown). However, the capacity of the TK domain of SmTK4 to induce GVBD in Xenopus oocytes provided at least indirect evidence for the potential of the schistosome Syk kinase to activate a MAP kinase cascade since GVBD is only induced when this signaling pathway has been activated [70].
The differentiation of germ cells in vertebrates and invertebrates depends on the re-organization of the actin and tubulin cytoskeletons [71]–[73]. The participation of Syk kinases in these processes, either directly by phosphorylation of cytoskeleton components [64], or indirectly by the activation of further signaling molecules has already been described [74],[75]. The identity of the isolated up- and downstream interaction partners of SmTK4 allowed first speculations for a role of SmTK4 in signaling pathways, regulating the proliferation and/or differentiation of germinal cells by activation of a MAPK pathway and/or by influencing cytoskeleton rearrangements. To investigate a cytoskeletal role, the Syk kinase-specific inhibitor Piceatannol was used for functional inhibition of the protein and second, SmTK4-specific dsRNAs were applied for post-transcriptional gene silencing. Since Southern-blot [18] and in silico analyses confirmed that SmTK4 is the only Syk kinase in S. mansoni, and since at the concentrations used Piceatannol only inhibits Syk kinases [34],[45], the morphological effects observed by CLSM could be specifically attributed to an inhibition of SmTK4. This conclusion was supported by GVBD assays in Xenopus oocytes confirming that Piceatannol is able to specifically inhibit the catalytic activity of the TK domain of SmTK4 in a concentration-dependent manner. Inhibitor and dsRNA treatment seemed to disrupt sperm development in schistosome males at an early stage, since the number of mature sperms as well as of spermatocytes was reduced within the testicular lobes. It seems likely that the inhibition influenced the proliferation of the spermatogonial cells in the dorsal part of the testes. The proliferation of these cells is necessary for the initiation of spermatogenesis and the continuous production of mature sperms [76]. Thus, a dysfunction of spermatogenesis at this early stage would lead to the absence of mature sperms. This coincides with studies from other organisms (mouse, human) showing that the proliferation of spermatogonial cells is activated by CTKs. Accordingly, Src-specific inhibitors or RNAi led to dysfunctions in spermatogenesis [76]. For CTKs of the Src family, an involvement in spermatogenesis is well-known, and this has also been hypothesized for the Lyn- and Fyn-family kinases [76],[77]. Indeed, the CLSM analysis of S. mansoni couples treated with the Src kinase-specific inhibitor Herbimycin A also showed significantly reduced numbers of sperms in the ventral part of the testicular lobes and within the sperm vesicle [78]. It was shown before that Herbimycin A affects the biochemical activity of SmTK3 [31], but we cannot exclude that it also affects SmTK6 activity. In contrast to the Src-family kinases, however, the involvement of a Syk kinase in spermatogenesis has not been confirmed previously in eukaryotes.
Harayama et al. [79] first reported the existence of a Syk kinase in spermatozoa of mammals. Here Syk is phosphorylated and activated by a cAMP-activated PKA indicating the involvement of a signaling pathway, which differs from that in hematopoietic cells. It was speculated that the PKA-Syk pathway is associated with the fertilization capacity of sperms [79], but functional data have not been provided yet. In untreated schistosome females the anterior part of the ovary contains the immature developing oogonia which have stem-cell character, whereas the posterior part contains mature primary oocytes with enlarged cytoplasms. As in other trematodes, primary oocytes leave the ovary to become fertilized within the oviduct before entering meiosis [46],[80]. Upon Piceatannol treatment, an increased number of primary oocytes and fewer immature oocytes were observed in females. One explanation for this structural change is an inhibition of processes occurring early in oocyte development such as oogonial divisions, leading to a reduced supply of primary oocytes. This is similar to the phenotype observed in the testes and supported by the result of the egg-count reduction test, which showed a significant reduction of egg production within one week in paired females treated with Piceatannol. We expected a complete fading of egg production with elongated time of treatment. However, a long-term effect of Piceatannol on egg production could not be studied because worms died nearly 10 days after treatment in culture before egg production had terminated. Furthermore, a hatching assay showed no morphological differences between developing eggs of inhibitor-treated worms or controls, indicating that SmTK4 may not have an essential function for egg maturation or developing miracidia (results not shown). RNAi approaches performed to post-transcriptionally silence SmTK4 resulted in similar phenotypes in the testes and ovary of adult S. mansoni. This indicates that electroporation combined with soaking was suitable and efficient to induce RNA silencing in the gonads. In previous studies, RNAi effects were observed for genes expressed in the tegument or in the gastrodermis [47],[81], tissues more easily accessible to dsRNA. At this time-point, the RNAi protocol for adult schistosomes is not fully reproducible in our hands, as indicated by variable rates of SmTK4 knock-down in the three independent experiments, and by the absence of a phenotype in the SmTK3-dsRNA control. Since the SmTK4-dsRNAs reduced the transcript levels incompletely, phenotypes had been obtained that were qualitatively but not quantitatively comparable to the Piceatannol-induced phenotypes. The efficiency of this inhibitor to suppress SmTK4 function was higher resulting in more dramatic phenotypes in the gonads of adult schistosomes.
In oocytes of marine nemertean worms (Cerebratulus sp.), inhibitor studies also showed evidence for a role of Syk kinases, as well as of Src-like CTKs during oocyte maturation [82]. Using the nonspecific tyrosine kinase inhibitor genestein (50–100 µM), or the Syk-kinase specific inhibitor Piceatannol (50–100 µM) a decrease in the maturation of oocytes was observed, indicated by reduced GVBD occurring in early meiosis, and also by reduced MAPK activity. With Src-kinase inhibitors (100 µM PP2, 20 µM SU6656) no reduction in GVBD was observed, but in that of MAPK activity. Based on their data Stricker & Smythe [82] propose that in oocytes of nemertean worms the signal is transduced from RTKs to Syk kinases, followed by Src kinases and ending in a MAPK-signaling pathway that leads to the activation of MPF (maturation-promoting factor) and GVBD. This model is basically, but not completely supported by the findings of our study. Besides the indirect evidence for the MAP kinase cascade-inducing capacity of SmTK4 in Xenopus oocytes, Y2/3H analyses identified its interactions with the Src kinases SmTK6 and SmTK3 as upstream partners as well as with a MAPK-activating protein as one of two candidate downstream partner. Since Src kinases are activated by RTKs, it seems feasible to us that in S. mansoni oocytes the signal is transduced from a RTK to a Src kinase leading to the activation of the Syk kinase SmTK4, which in turn activates a MAPK cascade.
Its essential function for gonad development and the long-term killing effect of worms by Piceatannol in culture suggest that SmTK4 may be a candidate target for blocking transmission and disease progression. Although CTKs such as Syk kinases are conserved in the animal kingdom, there may be possibilities to target specific homologs. A comparison of human Syk (accession number AAH11399) and SmTK4 (accession number CAD13249) revealed differences at the amino acid level between 17% (within the catalytic domain) and 63% (whole protein). This and the remarkable elongated linker-domain region of SmTK4 may provide a basis for selective drug design.
All experiments involving hamsters within this study have been performed in accordance with the European Convention for the Protection of Vertebrate Animals used for Experimental and other Scientific Purposes (ETS No 123; revised Appendix A) and have been approved by the Regional Council (Regierungspraesidium) Giessen (V54-19 c 20/15 c GI 18/10).
A Y2H cDNA-library based on RNA of adult male and female S. mansoni was constructed [33]. To this end, cDNAs were cloned into the prey vector pGADT7-Rec (leucine nutritional marker LEU2, Clontech) in frame with the GAL4 activation domain (Gal4-AD). For subsequent library screening, mating was performed according to the user manual (Yeast protocols handbook, Clontech). Two yeast strains were used for the mating; the library-containing strain AH109 (Mat a; reporter genes ADE2, HIS3, and LacZ) and the bait-containing strain Y187 (Mat α; reporter genes HIS3, LacZ). As bait vector for library screening, the plasmid pBridge (tryptophan nutritional marker TRP1, Clontech) was used, which contains two multiple cloning sites, MCS I and MCS II. MCS I allows the cloning of protein-coding gene sequences as fusion constructs with the GAL4 DNA binding domain (Gal4-BD). MCS II permits the cloning of a second gene sequence for the expression of an additional protein and, therefore, allows the establishment of a Y3H system.
Screening for upstream interaction partners of SmTK4 was performed with a bait vector containing the tandem SH2-domain of SmTK4 cloned into the MCS I in frame with the GAL4 DNA-binding domain (Gal4-BD). The encoding sequence was amplified by PCR using the SmTK4-specific primer pair SmTK4-SH2SH2-5′ (5′-GGATCCGTGGAGCTATTCCAC-3′; containing a BamHI restriction site) and SmTK4-SH2SH2-3′ (5′-CTGCAGTGATATACCACCGGA-3′; containing a PstI restriction site), and a full-length cDNA clone of SmTK4 as template. Amplification products of the expected size were cloned via BamHI/PstI into the MCS I of the vector pBridge. After cloning, the resulting construct SmTK4-SH2SH2 pBridge was sequenced to confirm the correct open reading frame of the Gal4-BD/SmTK4-SH2SH2 fusion. To perform a Y3H screening for upstream binding partners of SmTK4 a second bait vector was constructed, containing SmTK4-SH2SH2 in the MCS I and, additionally, the coding sequence for the TK domain of the schistosome Src kinase SmTK3 in the MCS II. To this end, the sequence for the SmTK3 TK-domain was amplified by PCR using the SmTK3-specific primers SmTK3-TK-5′ (5′-GCGGCCGCATCATCCAGAACCTGTGGG-3′; containing a BglII restriction site) and SmTK3-TK-3′ (5′-AGATCTGCTGGTTGCTCATCTTCAC-3′; containing a NotI restriction site), and a full-length cDNA clone of SmTK3 as template. The PCR product was cloned via BglII/NotI into the MCS II of SmTK4-SH2SH2 pBridge resulting in the construct SmTK4-SH2SH2 + SmTK3-TK pBridge. The success of this cloning approach was confirmed by sequencing. For control studies to test binding specificities, modified versions of this construct were additionally cloned by deletion of either the N-terminal or the C-terminal SH2-domain of SmTK4 (SmTK4-SH2(1) + SmTK3-TK pBridge, SmTK4-SH2(2) + SmTK3-TK pBridge).
To screen for downstream interaction partners of SmTK4 three different bait vectors were constructed, containing either the linker-region of SmTK4, the linker-region and the TK domain, or only the TK domain of SmTK4 in the MCS I. The coding sequences of these regions of SmTK4 were amplified by PCRs with the primer pairs TK4-bait1-5′ (5′-GGATCCTACAGAAACCAATACCAGTATC-3′) + TK4-bait1-3′ (5′-CTGCAGGGTATGCAAATATTTGTTTGT-3′), TK4-bait1-5′ + TK4-bait2-3′ (5′-CTGCAGTTCAACAAGAAATTCGATG-3′), or TK4-bait2-5′ (5′-GGATCCAAATTTATGATGAATTACCACC-3′) + TK4-bait2-3′, respectively, and a full-length cDNA clone of SmTK4 as template. The 5′-primers contained a BamHI restriction site and the 3′-primers a PstI restriction site, respectively. Amplification products of the expected sizes were cloned via BamHI/PstI in the MCS I of pBridge in frame with the Gal4-BD, resulting in the constructs SmTK4-linker+TK pBridge, SmTK4-linker pBridge, and SmTK4-TK pBridge. For Y3H analysis, an additional vector was constructed, containing the SmTK4 linker region and TK domain in the MCS I, and the SmTK3 TK-domain in the MCS II (SmTK4-linker+TK + SmTK3-TK pBridge). The integrity of the open reading frames with the Gal4-BD was confirmed by sequencing.
The screening for upstream or downstream interaction partners of SmTK4 was performed with either the bait construct SmTK4-SH2SH2 + SmTK3-TK pBridge or SmTK4-linker+TK + SmTK3-TK pBridge, respectively. Yeast cells (strain Y187) were individually transformed with both plasmids by the lithium acetate method (Yeast protocols handbook, Clontech). For the screening of the S. mansoni library, bait-expressing Y187 cells were mated with library-containing AH109 cells. Mating efficiencies of 75% or 28% were obtained, respectively, which exceeded the required minimum of 2% in both cases (Clontech). The first selection of diploid yeast cells containing interacting proteins was carried out on synthetic dropout medium lacking the amino acids tryptophan, leucine, and histidine (Trp−/Leu−/His−). To enhance the selection pressure on clones with interacting proteins, grown colonies were plated onto synthetic dropout medium lacking the amino acids tryptophan, leucine, histidine, and adenine (Trp−/Leu−/His−/Ade−). For further selection, β-Gal colony filter assays were performed using X-Gal as substrate according to the manufacturer's instructions (Yeast protocols handbook, Clontech). From positively tested yeast clones, plasmid DNA was isolated using cell disruption by vortexing with glass beads (Sigma) followed by plasmid preparation (PeqLab). Isolated plasmid DNA was transformed into heat shock-competent Escherichia coli cells (DH5α), and the bacteria selected on LB-plates containing ampicillin (100 µg/µl). To differentiate bacterial colonies containing the pBridge bait-plasmid from those containing a pGADT7 prey-plasmid, colony PCRs with pGADT7-specific primers were performed. Prey plasmids from PCR-positive bacterial clones were isolated and sequenced commercially (AGOWA, Berlin). For further binding analyses, the yeast strain AH109 was transformed with appropriate prey plasmids together with different bait plasmids. To confirm protein-protein interactions, the selection procedures were repeated. For quantification of relative interaction strengths, β-Gal liquid assays with ONPG (o-nitrophenol-galactopyranoside, SIGMA) as substrate were performed according to the Yeast protocols handbook from Clontech.
For isolation of yeast total RNA, a 5 ml overnight culture of the appropriate yeast clone was centrifuged to harvest cells. The pellet was washed two times with PBS and afterwards frozen in liquid nitrogen. Cells were disrupted by three freeze/thaw cycles (liquid nitrogen, 37°C water bath), 1 ml TriFast (PeqLab) was then added to the cell lysate, and total RNA was extracted according to the manufacturer's instructions.
Total protein extracts from yeast cells were obtained using the urea/SDS method as described in the Yeast protocols handbook (Clontech). For the inhibition of endogenous proteases and phosphatases, the buffer for cell disruption was supplemented with PMSF (phenylmethylsulfonylfluoride, SIGMA), protease inhibitor cocktail (Complete Mini, Roche), NaF (sodium fluoride, 50 mM), and Na3VO4 (sodium orthovanadate, 2 mM). Protein extracts were analyzed by standard SDS-PAGE and Western blotting with an anti-phosphotyrosine-specific antibody (Santa Cruz Biotechnology).
To confirm the transcription of the different bait-constructs inside the yeasts, RT-PCRs were performed with total RNA from bait-containing yeast clones as template. RT-PCRs were carried out stepwise in two separate reactions. First, 90 ng of total RNA was converted into cDNA using the SensiScript reverse transcriptase (Qiagen) and oligo-d(T), or bait sequence-specific primers. One fourth of the RT reaction volume was used for PCR amplification using appropriate bait sequence-specific primers. Amplification products were analyzed by agarose gel electrophoresis.
The pESC-His yeast expression system (Novagen) was used for immunoprecipitation experiments. The pESC-His vector contains two multiple cloning sites (MCS I, contains a FLAG-tag sequence; MCS II contains a cMyc-tag sequence) under the control of galactose-inducible GAL10 or GAL1 promoters. The tandem SH2-domain of SmTK4 was cloned into MCS I (FLAG-tag at the C-terminus), and the nearly complete version of SmTK6 (except 250 bp of the N-terminus) was cloned into MCS II (cMyc-tag at the N-terminus). For cloning, the tandem SH2-domain of SmTK4 was amplified by PCR using gene-specific primers containing appropriate restriction sites (TK4-SH2SH2-pESC-His-5′(NotI): 5′-GCGGCCGCAATGGGAGCTATTCCACCG-3′; TK4-SH2SH2-pESC-His-3′(ClaI): 5′-ATCGATGATATACCACCGGAACCTGA-3′). Before cloning into MCS II, the SmTK6 sequence was also PCR-amplified using gene-specific primers with restriction sites (TK6-Voll-pESC-His-5′(XhoI): 5′-CTCGAGAATGTTGTGACTGATGTGCAT-3′; TK6-Voll-pESC-His-3′(SacII): 5′-CCGCGGTTATCTAAATATTGAGCTTCTGTGTGC-3′). The integrity of the cloned constructs was confirmed by sequencing (AGOWA, Germany).
Following transformation, yeast cells (strain YPH501) were grown for 5 days at 30°C on selection media (synthetic drop-out medium [SD], His−, + glucose) to ensure the presence of the vector. Selected clones were grown over night in SD (His−, + glucose) liquid medium. The cells were centrifuged the next day and resuspendend in SG medium (His−, + galactose) and incubated for 5 h to induce transgene expression. As control, transformed yeast cells were alternatively resuspendend in SD (His−, + glucose). Protein was isolated from 10 ml culture volume for electrophoresis (10 µg protein each; 10% SDS-PAGE). After blotting on nitrocellulose (Schleicher & Schüll) the induction of the expression of the recombinant schistosome proteins was confirmed by Western-blot analysis using an anti-FLAG-tag (2.5 µg/µl; Novagen) or an anti-cMyc-tag antibodies (1 µg/µl; Novagen). A goat anti-rabbit-HRP (horse raddish peroxidase) secondary antibody was used for detection (1∶70.000; Novagen). Individual bands of the expected sizes of 31 kDa (SmTK4-SH2SH2) or 55 kDa (SmTK6) were observed in proteins obtained from the galactose-induced yeast culture but not in proteins obtained from the glucose-induced culture (results not shown).
For co-IP 70 µl of the protein lysate of the galactose-induced culture were pre-incubated with 30 µl protein A sepharose (Sigma) to remove proteins binding unspecifically. The remaining lysate was splitted in two parts, and each half incubated (over night at 4°C) with 4 µg anti-FLAG-tag antibody or 1 µg anti-cMyc-tag antibody, respectively. Antibody complexes were precipitated by protein A sepharose (2 h at 4°C). Following washing steps, the protein complexes were eluated from the protein A sepharose, and a Western-blot analysis was done with the recovered yeast protein as described above. After nitrocellulose transfer, the proteins were visualized by INDIA ink staining (Pelikan, Germany; 1% acetic acid, 0.04% Tween 20, 0.1% Fount India).
To monitor the transcription of SmTK4 in control and treated S. mansoni, total RNA was extracted using TriFast (PeqLab) following the manufacturer's instructions. Residual DNA remaining in the RNA preparations was removed by DNase digestion using RNase-free DNaseI (Fermentas). cDNA was synthesized using 1 µg total RNA, 1 µl oligo-d(T) primer (dT24VN; 20 µM), 1 µl nonamer primer (dN9; 20 µM) and Superscript II reverse transcriptase (Invitrogen). Subsequent PCRs were performed with 1/10 of the cDNA as template, FIREPol Taq polymerase (Solis BioDyne), and the following primer combination: SmTK4-Sub3-5′ (5′-ATGACGTAAAAGATTCACGTG-3′) and SmTK4-Sub3-3′ (5′-TGCATGTTCTTCACTACAATC-3′), which flank the region used as target for the dsRNAs. For normalization, the transcription of the housekeeping gene SmPDI [48] was monitored using the same cDNAs as template, but using the following primer combination: SmPDI-5′ (5′-GGGATTTATCAAGGATACGGACTC-3′) and SmPDI-3′ (5′-CACCAAGGAGCATACAGTTTGAC-3′). All PCRs were performed in a final volume of 25 µl. PCR products were separated on 1.5% agarose gels stained with ethidium bromide. The relative intensities of the amplification products were determined densitometrically using the program ImageJ (version 1.4.1; http://rsbweb.nih.gov/ij/index.html). For relative quantification of the SmTK4 products, the SmPDI products were used as endogenous standard.
In situ hybridizations were done as described elsewhere in detail [31],[33]. In short, adult worm pairs were fixed in Bouin's solution (picric acid/acetic acid/formaldehyde; 15/1/5) before embedding in paraplast (Histowax, Reichert-Jung). Sections of 5 µm were generated and incubated in xylol to remove the paraplast. Following re-hydration, proteins were removed by proteinase K treatment (final concentration 1 µg/ml), and the sections were dehydrated. For hybridization, in vitro transcripts were labeled with digoxigenin following the manufacturers' instructions (Roche). Labeled sense and antisense transcripts of SmTK6 (unique site, length 354 bp), the MAPK activating protein PM20/21 (length 452 bp), or mapmodulin (length 420 bp), were size-controlled by gel electrophoresis. To prove their quality, transcript blots were made to confirm digoxigenin incorporation by alkaline phosphatase-conjugated anti-digoxigenin antibodies, naphthol-AS-phosphate, and Fast Red TR (Sigma). All in situ hybridization were performed for 16 h at 42°C. Sections were stringently washed up to 0.5×SSC, and detection was achieved as described for transcript blots.
A Liberian isolate of S. mansoni was maintained in Biomphalaria glabrata as intermediate host and in Syrian hamsters (Mesocricetus auratus) as definitive host [83]. Adult worms were obtained by hepatoportal perfusion at 42–49 days post-infection.
After perfusion, adult schistosomes were washed three times with M199 medium before being cultured in vitro in M199 (Gibco; including glucose, sodium bicarbonate, 4-(2-hydroxyethyl)-1-piperazineethane sulfonic acid) supplemented with an antibiotic/antimycotic mixture (1.25%, Sigma) and FCS (10%, Gibco) at 37°C and 5% CO2 [42]. For each experiment, 10–30 pairs of S. mansoni were kept in 60 mm diameter culture dishes in 3 ml culture medium. The medium was changed every 24 hours. If needed, schistosome pairs were carefully separated with fine tweezers.
Piceatannol (3,4,3′,5′-Tetrahydroxy-trans-stilbene; Alexis Biochemicals) was dissolved in dimethyl sulfoxide (DMSO) (5 µg/µl). For each inhibitor treatment experiment 20 adult couples of S. mansoni were maintained in 10 ml culture medium [42], supplemented with various concentrations of Piceatannol (35 µM, 70 µM, 100 µM). Medium and inhibitor were refreshed every 24 hours during the treatment periods in vitro. For morphological analysis, adult worms were fixed for at least 24 hours in AFA (alcohol 95%, formalin 3%, and glacial acetic acid 2%), stained for 30 minutes with 2.5% hydrochloric carmine (Certistain®, Merck), and destained in acidic 70% ethanol. After dehydration for 5 minutes in 70%, 90%, and 100% ethanol, respectively, worms were preserved as whole-mounts in Canada balsam (Merck) on glass slides [43],[44]. CLSM images were made on a Leica TSC SP2 microscope using a 488 nm He/Ne laser and a 470 nm long-pass filter in reflection mode.
As basis for double-stranded RNA (dsRNA) synthesis, a 813 bp fragment of the SmTK4-coding DNA was amplified by PCR using the gene-specific primers SmTK4-5′ (5′-ATGCCTGGAGCTATTCCA-3′) and SmTK4-3′ (5′-TGATATACCACCGGA-3′) and a full-length cDNA clone of SmTK4 as template. The amplification product of expected size was cloned into the pDrive cloning vector (Qiagen). The resulting construct, containing T7 and SP6 RNA polymerase promoters flanking the SmTK4 sequence was used to generate single-stranded RNA by in vitro transcription with T7 and SP6 RNA polymerases (MEGAscript RNA transcription kit, Ambion). The single-stranded RNAs were purified by LiCl-precipitation, resuspended in dH2O, and quantified by spectrophotometry. Equal amounts of the single-stranded RNAs were mixed in annealing buffer (500 mM potassium acetate, 150 mM HEPES-KOH, pH 7.4, 19 mM magnesium acetate, sterile dH2O) and incubated at 68°C for 15 min. Annealing and integrity of the dsRNAs were confirmed by agarose gel electrophoresis.
The dsRNA was delivered to adult worms according to the electroporation protocol of Correnti et al. [84] and Ndegwa et al. [47]. Briefly, electroporations were performed in 4 mm cuvettes with 10 couples in 50 µl electroporation buffer (Ambion) containing 25 µg dsRNA. A square-wave protocol was applied with a single 20 ms impulse at 125 V and at room temperature (Gene Pulser XCell™, Biorad). After electroporation, the worms were transferred to complete M199 medium and incubated for 5 days; 48 hours after electroporation the medium was refreshed.
Capped messenger RNA (cRNA) encoding the full-length cDNA of SmTK4 or a shortened variant containing the catalytic TK domain of SmTK4 was synthesized in vitro using the T7 mMessage mMachine Kit (Ambion, USA). Following synthesis, cRNA was injected into stage VI oocytes according to previously published protocols [50],[85]. To investigate Piceatannol effects, transfected oocytes were incubated with increasing concentrations (1, 2, 5, 10, 20, 50, or 100 µM) of Piceatannol. Non-injected control oocytes were incubated with the same inhibitor concentrations. As positive control progesterone was used, a steroid known to induce germinal vesicle break down (GVBD) in Xenopus oocytes [49]. This oocyte-specific physiological activity was detected by the appearance of a white spot at the center of the animal pole.
The following public domain tools were used for sequence analyses: NCBI-BLAST (http://blast.ncbi.nlm.nih.gov/Blast.cgi), the Wellcome Trust Sanger Institute S. mansoni OmniBlast server (http://www.sanger.ac.uk/cgi-bin/blast/submitblast/s_mansoni/omni), and GeneDB (http://www.genedb.org). For BLAST analyses to identify Syk kinases in the schistosome genome data set [6] we used the kinase domain sequence as template as the most conserved part of tyrosine kinases.
|
10.1371/journal.pgen.1004278 | The Caenorhabditis elegans Myc-Mondo/Mad Complexes Integrate Diverse Longevity Signals | The Myc family of transcription factors regulates a variety of biological processes, including the cell cycle, growth, proliferation, metabolism, and apoptosis. In Caenorhabditis elegans, the “Myc interaction network” consists of two opposing heterodimeric complexes with antagonistic functions in transcriptional control: the Myc-Mondo:Mlx transcriptional activation complex and the Mad:Max transcriptional repression complex. In C. elegans, Mondo, Mlx, Mad, and Max are encoded by mml-1, mxl-2, mdl-1, and mxl-1, respectively. Here we show a similar antagonistic role for the C. elegans Myc-Mondo and Mad complexes in longevity control. Loss of mml-1 or mxl-2 shortens C. elegans lifespan. In contrast, loss of mdl-1 or mxl-1 increases longevity, dependent upon MML-1:MXL-2. The MML-1:MXL-2 and MDL-1:MXL-1 complexes function in both the insulin signaling and dietary restriction pathways. Furthermore, decreased insulin-like/IGF-1 signaling (ILS) or conditions of dietary restriction increase the accumulation of MML-1, consistent with the notion that the Myc family members function as sensors of metabolic status. Additionally, we find that Myc family members are regulated by distinct mechanisms, which would allow for integrated control of gene expression from diverse signals of metabolic status. We compared putative target genes based on ChIP-sequencing data in the modENCODE project and found significant overlap in genomic DNA binding between the major effectors of ILS (DAF-16/FoxO), DR (PHA-4/FoxA), and Myc family (MDL-1/Mad/Mxd) at common target genes, which suggests that diverse signals of metabolic status converge on overlapping transcriptional programs that influence aging. Consistent with this, there is over-enrichment at these common targets for genes that function in lifespan, stress response, and carbohydrate metabolism. Additionally, we find that Myc family members are also involved in stress response and the maintenance of protein homeostasis. Collectively, these findings indicate that Myc family members integrate diverse signals of metabolic status, to coordinate overlapping metabolic and cytoprotective transcriptional programs that determine the progression of aging.
| Transcription factors are essential proteins that regulate the expression of genes and play an important role in most biological processes. The results of our study presented here demonstrate for the first time a role in aging for a small family of transcription factors in the nematode worm Caenorhabditis elegans. Importantly, these proteins have close relatives in higher organisms, including humans that influence metabolism, cell replication, and have been implicated in the development of cancer. Moreover, the loss of one homologue has also been implicated in Williams-Beuren syndrome, a disease characterized in part by signs of premature aging. Our data demonstrate that these transcription factors function within insulin/IGF-1 signaling and dietary restriction, two highly conserved pathways that link nutrient sensing to longevity. Taken together, our findings provide exciting new insight into a family of proteins that may be essential for linking nutrient sensing to longevity and have implications for the improvement of human healthspan.
| A large body of evidence indicates that sensors and regulators of cell metabolism modulate organismal aging through evolutionarily conserved pathways. Research on the nematode C. elegans has been instrumental in the characterization of many of these pathways, including the insulin/IGF (ILS), Target of Rapamycin (TOR), AMP protein kinase (AMPK), sirtuin (Sir2), and dietary restriction (DR) pathways [1]–[5]. For example, the ILS and DR signaling pathways target the forkhead transcription factors DAF-16 (FoxO) and PHA-4 (FoxA), respectively, suggesting that these two regulatory inputs function independently [6]–[8]. Although much is known about the structure and function of these pathways, many questions remain about how those signals are integrated, the extent of their functional overlap, and how these pathways determine the progression of aging.
The large Myc super-family is comprised of basic helix-loop-helix leucine zipper (bHLHZip) containing transcription factors that are important regulators of cell growth, proliferation, and energy metabolism. In mammals, there are over 100 bHLH transcription factors [9]. Within this larger super-family, the core mammalian “Myc interaction network” consists of 11 proteins: 3 Myc; 4 Mad/Mxi; 2 Mnt/Mga; and 2 Mondo proteins. These proteins can heterodimerize in a complex manner with either Max or Mlx to regulate transcription by binding to a range of enhancer box (E-box) DNA sequences. C. elegans possesses 42 bHLH transcription factors, including a simplified Myc interaction network [10] comprised of four genes: mml-1 (T20B12.6, Myc and Mondo like), mxl-2 (F40G9.11, Mlx), mxl-1 (T19B10.11, Max), and mdl-1 (R03E9.1, Mad) (Figure 1A). While C. elegans lacks a true Myc orthologue, mml-1 contains a region that is highly similar to the N-terminal region of c-Myc, and presumably incorporates both Myc and Mondo functions [11]. Similar to their mammalian counterparts, members of the C. elegans Myc interaction network form obligate heterodimers that bind to a range of E-box sequences and either activate or repress transcription (respectively: MML-1:MXL-2, hereafter referred to as the Myc-Mondo complex and MDL-1:MXL-1, referred to as the Mad complex) (Figure 1A). Unlike mammalian Mlx and Max, these protein-protein interactions are specific, and no evidence exists for cross-heterodimerization. In C. elegans, null mutations in any of the four genes produces no overt developmental phenotype, although a subtle cell migration defect has been reported in mxl-2(tm1516) mutant animals [11]. From a comprehensive genome-wide RNAi screen, we had previously discovered that mxl-2 is essential for the extended longevity of the daf-2(e1370) insulin/IGF1 receptor loss of function mutants [12]. Additionally, mdl-1 has been putatively identified as a target of ILS based on microarray analysis, and loss of mdl-1 has been reported to produce a subtle negative effect on the extension of lifespan in ILS mutants [13].
In mammals, two Mondo paralogs (MondoA and ChREBP) function as intracellular sensors of carbohydrate availability and regulators of glycolytic and lipogenic gene expression. MondoA (MLXIP) is predominately expressed in skeletal muscle. When activated by glucose-6-phosphate (G6P), it translocates from the mitochondrial outer membrane to the nucleus where it upregulates the expression of glycolytic genes, by DNA binding to E-box sequences. ChREBP (Carbohydrate Response Element Binding Protein/MLXIPL/MondoB/WBSCR14) is expressed in the liver and is also activated by carbohydrate intermediates, most likely G6P (reviewed in [14]). Activated ChREBP exhibits cytoplasmic-to-nuclear shuttling similar to MondoA and upregulates the expression of both glycolytic and lipogenic genes (triglyceride synthesis) through binding of carbohydrate response elements (ChORE sequences) within the genome [15]–[17]. In mammals, the absence of ChREBP ablates glucose-induced target gene activation [18], [19]. ChREBP –/– mice display marked physiological changes: large glycogen-laden livers; smaller fat deposits; decreased plasma free fatty acid levels and lipogenic enzyme expression; signs of insulin resistance; reduced glycolytic flux; and enhanced glycogen accumulation [20]. While the specific roles of ChREBP in humans have not been elucidated, the gene encoding ChREBP is one of 27 genes deleted in Williams-Beuren Syndrome (WBS), a complex developmental disorder. Interestingly, WBS patients present signs of premature aging including: premature graying of hair; glucose intolerance; diabetes; and hearing loss that commonly develops during adolescence or young adulthood, concomitant with declining memory skills or dementia [21].
Prompted by our previous identification of mxl-2 as an effector of daf-2 mutants in the extension of C. elegans lifespan [12], we asked whether Myc-Mondo and Mad complexes might affect longevity. Thus we obtained null mutant animals for mml-1, mxl-2, mdl-1, and mxl-1, then assessed their lifespan. Loss of either component of the transcriptional repressor complex resulted in a robust extension of lifespan: mdl-1(tm311) and mxl-1(tm1530) single null mutant animals had a 44.5% and 33.2% (p<0.0001) increase in median lifespan compared to wild-type (N2) animals, which had a median lifespan of 18.4+/–0.4 days (Figure 1B, Dataset S1). Using traditional and replica set lifespan approaches, similar results were obtained when either mdl-1 or mxl-1 were inactivated by RNAi (Figure S6B, S6C, Dataset S1). Furthermore, simultaneous inactivation of mxl-1 and mdl-1 resulted in lifespans comparable to the single mutant animals (Figure S1A, S1B, Dataset S1), consistent with published observations in C. elegans that these two proteins only function as an obligate heterodimer [10], [11]. In contrast, loss of either component of the transcriptional activation complex shortened lifespan; mml-1(ok849) and mxl-2(tm1516) single null mutant animals had a 14.1% and 22.8% reduction in median lifespan compared to wild-type animals (Figure 1C, Dataset S1). Inactivation of mml-1 or mxl-2 by RNAi produced a similar result using both traditional and replica set lifespan analyses (Figure S6, Dataset S1). Analogous to what was observed with mdl-1 and mxl-1, simultaneous inactivation of mml-1 and mxl-2 resulted in lifespans comparable to the single mutant animals (Figure S1C, S1D, Dataset S1). Thus loss of the Mad (MDL-1:MXL-1) transcriptional repressor complex extends lifespan, while loss of the Myc-Mondo (MML-1:MXL-2) activator complex shortens lifespan.
We conducted an epistasis analysis to determine whether the lifespan extension conferred by loss of the Mad complex (MDL-1:MXL-1) was dependent on the function of the Myc-Mondo complex (MML-1:MXL-2). Inactivation of either mml-1 or mxl-2 via RNAi fully suppressed the extended longevity of both mdl-1(tm311) and mxl-1(tm1530) mutant animals (Figure 1D, Dataset S1). Therefore, the extended longevity conferred by the loss of the Mad transcriptional repression complex is dependent upon an intact Myc-Mondo complex. Interestingly, C. elegans possess a second Max like protein, MXL-3 (F46G10.6) [22]. Both MXL-2 and MXL-3 share similar sequence homology to mammalian Max; however, MXL-3 likely functions as a homodimer unlike MXL-2 and mammalian Max [10]. Moreover, mxl-3 links lipolysis and autophagy to nutrient availability [10], [23], and mxl-3(ok1947) null mutant animals are also long-lived (Figures S1E, S1F, Dataset S1 and [23]). In contrast to mdl-1 or mxl-1 mutants, inactivation of either mml-1 or mxl-2 had a negligible effect on mxl-3(ok1947) longevity (Figures S1E, S1F, Dataset S1). Thus the Myc-Mondo transactivation complex is required for the extension of lifespan in the absence of the Mad complex. In contrast, Myc-Mondo complex function is dispensable for the extension of lifespan conferred by loss of the Max paralog mxl-3. Collectively, this implies that the Myc-Mondo and Mad complexes have similar functions that influence longevity, which are separate from those of the Max paralog mxl-3.
We next determined whether the Myc-Mondo and Mad complexes functioned in insulin/IGF1 signaling, which is known to depend on the FoxO homologue daf-16. First we asked whether daf-16 was necessary for loss of the Mad complex (MDL-1:MXL-1) to extend longevity; second, we asked whether inactivation of daf-16 further shortened longevity in the absence of the Myc-Mondo complex (MML-1:MXL-2). Inactivation of daf-16 by RNAi completely suppressed the extended lifespan of mxl-1(tm1530) mutant animals (Figure 2A, Dataset S1). Similarly, inactivation of either mdl-1 or mxl-1 failed to confer longevity in daf-16(mgDf47) null mutant animals (Dataset S1). Conversely, daf-16(RNAi) did not further shorten the lifespan of mxl-2(tm1516) null mutant animals (Figure 2A, Dataset S1) and neither mml-1(RNAi) nor mxl-2(RNAi) further shortened the longevity of daf-16(mgDf47) null mutant animals (Dataset S1). Thus daf-16 (FoxO) is essential for loss of the Mad complex to extend longevity, and loss of the Myc-Mondo complex has no further negative effect on longevity in the absence of daf-16.
To further study the role of the Myc proteins in the longevity effects of ILS, we asked whether the transcription-activating Myc-Mondo complex (MML-1:MXL-2) was required for daf-2(e1370) extension of lifespan. daf-2(e1370) mutant animals fed control RNAi had a median and maximum lifespan of 46.0 +/– 1.2 days and 60.5 +/–2.4 at 20°C, respectively (Figure 2B, Dataset S1). Inactivation of either mxl-2 or mml-1 by RNAi reduced the median lifespan of daf-2(e1370) mutant animals, to 27.6+/–0.9 days and 27.2+/–1.3 days, respectively; a result consistent with our previous findings (Figures 2B, S2A, Dataset S1, and [12]). Similarly, daf-2(e1370);mxl-2(tm1516) double mutant animals had a median lifespan of 28.0+/–1.4 days (Figure 2C, Dataset S1). Thus loss of Myc-Mondo complex function by null mutation reduces daf-2(e1370) median lifespan by 18.0+/–1.4 days and wild-type lifespan by 4.2+/–0.5 days. The reduction caused by loss of mxl-2 is significantly greater (p = 0.001) in the daf-2(e1370) background (39.1%) than in the wild-type background (22.8%). This implies that the effect of the loss of the Myc-Mondo complex on lifespan has specificity to insulin/IGF-1 signaling. This is consistent with the results of a global analysis of synthetic genetic interactions [24], which predicted a relationship between mml-1 and daf-2.
If, as suggested by the above experiments, daf-2 and the Myc proteins are part of a common lifespan regulating system, they would not be expected to act independently. We sought to confirm this hypothesis by testing whether there is additivity between loss of the Mad complex and decreased ILS signaling. However, interpreting an epistasis analysis with ILS mutants is complicated, as even strong temperature-sensitive (ts) alleles of daf-2 [25] retain some level of ILS. For example, placing daf-2(e1370) mutant animals onto daf-2(RNAi) results in a 21.8% increase in median longevity from 46.0+/–1.2 to 56.0+/–1.6 days at 20°C (p = 0.0032) (Figure S2B). In order to perform daf-2 pathway analysis with the lowest sub-lethal amount of insulin-like signaling we examined the lifespan of daf-2(e1370) and daf-2(e1370);mxl-1(tm1530) mutants in the presence of daf-2 RNAi. Unlike daf-2(e1370) single mutant animals, daf-2(e1370);mxl-1(tm1530) double mutant animals show no significant further increase in lifespan on daf-2(RNAi) (p = 0.1164) (Figure 2D), implying that the C. elegans Mad complex and decreased ILS have similar functions in longevity.
As mammalian MondoA and ChREBP are known sensors of carbohydrate availability [14], we hypothesized that the Myc-Mondo (MML-1:MXL-2) and Mad (MDL-1:MXL-1) complexes may function in DR signaling. The pha-4 (FoxA) transcription factor is a major transcriptional output of DR signaling [8]. Analogous to our analysis with daf-16(RNAi) and ILS, we conducted epistasis experiments with pha-4(RNAi) to determine whether the Myc-Mondo and Mad complexes function in DR signaling. Inactivation of pha-4 completely suppressed the extended lifespan of both mxl-1(tm1530) (p = 0.0001) and mdl-1(tm311) (p = 0.0002) mutant animals, similar to our results with daf-16(RNAi) (Figure 3A, Dataset S1). Additionally, pha-4(RNAi) did not further shorten the lifespan of mxl-2(tm1516) mutant animals (p = 0.5835) (Figure 3A, Dataset S1). Thus pha-4 is required for the extension of lifespan conferred by loss of the Mad complex, and has similar pro-longevity functions as the Myc-Mondo complex.
Next, we tested whether the Myc-Mondo (MML-1:MXL-2) complex is required for lifespan extension by DR. To do this, we conducted a lifespan analysis in eat-2(ad465) mutant animals, which harbor a mutation in a nicotinic acetylcholine receptor, resulting in drastically decreased pharyngeal pumping and increased lifespan [26], [27]. Notably, eat-2(ad465);mxl-2(tm1516) and eat-2(ad465);mxl-1(tm1530) mutant animals have pumping rates identical to those seen in eat-2(ad465) mutants (Figure S7). Loss of the Myc-Mondo complex by mxl-2(RNAi), or eat-2(ad465);mxl-2(tm1516) double mutant partially suppressed the DR-mediated lifespan extension (Figure 3B, 3C, Dataset S1) by 11.8+/–1.3 days (40.5%), and 10.3+/–1.3 days (35.4%), respectively. This reduction is significantly greater than the 4.1+/–0.5 day decrease (22.7% reduction) in lifespan that results from the loss of mxl-2 in an otherwise wild-type background (p = 0.00055). Thus the Myc-Mondo complex is required for the increased lifespan conferred by DR, similar to what we found with ILS.
Lastly, we tested whether there would be additivity to overall lifespan in the absence of the Mad complex (MDL-1:MXL-1) under conditions of DR. Inactivation of either mxl-1 or mdl-1 did not significantly extend eat-2(ad465) lifespan, with median lifespans of 29.1+/–1.1; 29.3+/–1.1 (p = 0.398); 32.0+/–2.2 days (p = 0.28), respectively (Dataset S1). Additionally, eat-2(ad465) and eat-2(ad465);mxl-1(tm1530) mutant animals have similar lifespans (median lifespans of 29.1+/–1.1, and 30.1+/–1.3 (p = 0.565), Figure 3D, Dataset S1). Collectively, these results indicate that, in addition to ILS, C. elegans Myc-Mondo and Mad complexes function in DR longevity signaling.
The TGF-β pathway functions in parallel with ILS to regulate energy-balance, thereby affecting dauer formation, fat metabolism, egg laying, feeding behavior, and lifespan [13],[28]–[31]. daf-7 encodes a member of the TGF-β superfamily, which inactivates the co-SMAD DAF-3 via the TGF-β Type I and II receptors DAF-1 and DAF-4. Loss of daf-7 extends longevity dependent upon daf-3 [32]. Interestingly, DAF-3 binds the mdl-1 promoter, and mdl-1 expression in the pharynx is increased in daf-3 RNAi-treated animals, suggesting that DAF-3 negatively regulates the expression of mdl-1 [33]. Thus the decreased longevity of daf-3(tm4940) null mutant animals might, at least in part, be explained by upregulation of the Mad complex, and we predicted that loss of the Mad complex might rescue the shortened longevity of daf-3(tm4940) animals. To test this hypothesis, wild-type and daf-3(tm4940) null mutant animals were treated with control and mxl-1 RNAi, and lifespan was assessed. mxl-1 RNAi robustly increased the lifespan of wild-type animals by 5.0+/–0.9 days (27.3% increase, p<0.0001, Figure S3, Dataset S1), which is comparable to the effect of the mxl-1 null mutation (Figure 1B, Dataset S1). In contrast, inactivation of mxl-1 failed to extend the lifespan of daf-3(tm4940) null mutant animals (Figure S3, p = 0.24). Therefore, we conclude that TGF-β signaling does not influence lifespan through regulation of the Mad complex.
In mammals, the MondoA/ChREBP:Mlx complexes exhibit translocation from cytoplasm to nucleus in response to changes in carbohydrate availability [34], [35]. We hypothesized that changes in metabolic status, as relayed through decreased ILS or conditions of DR, might alter MML-1 subcellular localization. To determine whether the C. elegans Myc-Mondo complex (MML-1:MXL-2) might be regulated by a similar translocation, we generated an MML-1::GFP translational reporter under control of the endogenous mml-1 gene promoter. As has been previously reported, MML-1::GFP is broadly expressed (Figure 4A and [11]). However, contrary to previous reports, MML-1::GFP was observed to reside in both the nucleus and the cytoplasm under basal conditions (Figure 4A). The MML-1::GFP fusion was functional, as it rescued the thermotolerance defect of mml-1(ok849) (data not shown). In control animal populations, the relative distribution of MML-1::GFP varied between animals, ranging from entirely nuclear to entirely cytoplasmic; however, subpopulations showed predominately either nuclear with some cytoplasmic (56.1%), or approximately equal distribution between the nucleus and cytoplasm (28.3%) (Figure 4A, File S1). In contrast, decreased ILS (daf-2(e1370)) or conditions of DR (eat-2(ad465)) significantly increased the prevalence of the subpopulation with entirely nuclear MML-1::GFP from 8.3% to 36.1% and 40.0%, respectively (Figure 4A, compare columns 1 to 4 and 1 to 7). The increased nuclear accumulation of MML-1::GFP in response to reduced ILS and conditions of DR supports the idea that this aspect of MML-1/MondoA/ChREBP biology is conserved in nematodes, and provides a potential mechanistic explanation for the requirement of the Myc-Mondo complex in ILS and DR-mediated longevity; specifically that Myc-Mondo complex function is activated by decreased ILS or by conditions of DR to increase the expression of genes that extend longevity.
We hypothesized that the nuclear translocation of MML-1 in conditions of decreased ILS might be dependent upon daf-16, while increased nuclear accumulation in conditions of DR would require pha-4. Inactivation of either daf-16 or pha-4 failed to significantly alter the basal spectrum of MML-1::GFP nuclear accumulation in control animals (Figure 4A compare columns 2 and 3 to 1). To our surprise, inactivation of daf-16 failed to alter the increased nuclear accumulation of MML-1::GFP conferred by decreased ILS or conditions of DR (Figure 4A compare column 4 to 5 and column 7 to 8). In contrast, inactivation of pha-4 completely abolished the increased nuclear accumulation of MML-1::GFP conferred by decreased ILS or conditions of DR back to the distributions found in control populations (Figure 4A, compare column 4 to 6 and column 7 to 9). Thus we conclude that pha-4, but not daf-16, is essential for the increased nuclear accumulation of MML-1 under conditions of DR or decreased ILS.
We next asked whether ILS or DR influenced the expression of Myc family members. Interestingly, according to the modENCODE database (www.modENCODE.org) only the mml-1 promoter is directly bound by DAF-16, PHA-4, and MDL-1, making it the most likely target for transcriptional regulation in response to aging-relevant signals, such as ILS and DR [36]. Thus we first assessed whether decreased ILS, DR, or loss of mdl-1 altered expression of mml-1 by qRT-PCR. In daf-2(e1370) mutant animals, mml-1 mRNA levels were significantly increased by 2-fold (Figure 4B, p = 0.0027). In contrast, a significant change in mml-1 mRNA levels was not detected in mdl-1(tm311), or eat-2(ad465), mutant animals (Figure 4B, p = 0.3817 and 0.6033). To test whether the increased mRNA expression under conditions of decreased ILS was unique to mml-1, we also used qRT-PCR to measure the relative mRNA levels of mxl-2, mxl-1, and mdl-1 in daf-2(e1370) compared to wild-type animals. Only mml-1 was significantly upregulated under conditions of low ILS (Figure 4C, p = 0.0024); this implies that insulin-like signaling only transcriptionally regulates Myc-Mondo complex function at the level of mml-1 expression. Consistent with that notion, mxl-2 and mxl-1 are not direct targets of DAF-16 as measured by ChIP-seq (www.modENCODE.org), although mxl-1 is the last gene in an operon that is a target of DAF-16. Therefore, it is possible that ILS regulates mxl-1 expression within the context of the operon at a level that cannot be distinguished by qRT-PCR.
Since decreased ILS, but not conditions of DR, regulated the Myc family at the mRNA level, we explored the role of DAF-16 in this regulation. We first determined in qRT-PCR experiments whether the upregulation of mml-1 by decreased ILS was daf-16 dependent. The increased mml-1 mRNA levels observed in daf-2(e1370) mutant animals is daf-16 dependent, as daf-2(e1370);daf-16(mgDf47) double mutant animals expressed mml-1 at levels comparable to wild-type control animals (Figure 4C). Thus we conclude that decreased ILS activates DAF-16 to induce expression of mml-1, perhaps as a feed-forward mechanism to amplify DAF-16 signaling. Since MML-1 (Mondo) and MDL-1 (Mad) have opposing roles in transcription and longevity control, we hypothesized that loss of daf-16 might induce expression of mdl-1. Indeed, daf-2(e1370);daf-16(mgDf47) double mutant animals have a 3-fold increase in mdl-1 mRNA levels (Figure 4D, p = 0.0013). DAF-16 binds in the promoter region of mdl-1 (http://www.modENCODE.org and [37]). Thus we conclude that DAF-16 activity regulates expression of the key components of the Myc-Mondo and Mad complexes in a manner that correlates with their role in longevity. Specifically, decreased ILS activates DAF-16, which in turn results in increased mml-1 expression and extension of longevity; in contrast, loss of daf-16 results in increased mdl-1 expression and shortened lifespan.
We found that the Myc family members are regulated by DR and ILS signaling and have overlapping functions in longevity control, implying that the C. elegans Myc-Mondo/Mad complexes may cooperate with PHA-4 and DAF-16 at shared target gene promoters. Through an informatics analysis of available ChIP-seq data provided by the modENCODE project (www.modENCODE.org) [38], we sought to identify the overlap in genomic binding sites for DAF-16, PHA-4, and MDL-1. We compared the global frequency of the MDL-1 and PHA-4 binding midpoints to the position of the transcription start site (TSS), as has been done for DAF-16 [37]. MDL-1 and PHA-4 had enriched binding frequency within a region almost identical to what has been reported for DAF-16 (between –700 to +100 of the TSS, Figure 5A). We used this metric to distinguish genes more likely to be regulated by MDL-1, and identified 4605 possible target genes. The Myc family of bHLH transcription factors bind to E box sequences, therefore we asked whether the MDL-1 binding regions identified by ChIP-seq also contained E boxes, and found this to be true in 3555 cases (77.2%, Dataset S2). Next, we found that there is significant overlap of 2879 target genes that are bound by DAF-16, PHA-4, and MDL-1 (Figure 5B, p = 2.2e−16). Of 963 genes that have been identified as influencing longevity (based on published studies, gene ontology and description), 299 are bound by DAF-16, PHA-4, and MDL-1. In contrast, only 13 previously identified longevity genes are bound by MDL-1 and not DAF-16 or PHA-4, a significant enrichment (p = 2.2e−16). We next asked whether the actual binding sites of the three transcription factors overlapped. For 234 of 299 putative longevity genes the MDL-1 binding peak is 75% or more overlapped by DAF-16 and PHA-4, suggesting that binding occurs in common regulatory regions. We conclude that DAF-16, PHA-4, and MDL-1 bind at many common target genes, in common regions.
To gain mechanistic insight into the potential transcriptional programs co-regulated by DAF-16, PHA-4, and MDL-1, we determined whether there was over-representation of aging related gene ontology terms at putative common target genes. Comparing gene ontology of genes bound by DAF-16, PHA-4, and MDL-1 versus all genes revealed that determination of adult lifespan, carbohydrate metabolism, autophagy, unfolded protein binding, and stress response were significantly over-represented (Figure 5C and Dataset S2). An analysis of all potential MDL-1 targets (an additional 1726 genes) revealed similar classes of over-represented gene functions. Further, comparing MDL-1 target genes to previously identified putative targets of mammalian Mad1 [39], [40] reveals a large degree of overlap, which implies evolutionary conservation (Dataset S2). This suggests that DAF-16, PHA-4, and the Myc family converge on a significant subset of common target genes that have functions relevant to aging, carbohydrate metabolism, proteostasis, and stress response.
An examination of the specific carbohydrate metabolic pathways putatively regulated by MDL-1, DAF-16, and PHA-4 revealed a number of genes involved in glycolysis, gluconeogenesis, and glycogen metabolism (Figure 5D, Dataset S2). This is consistent with the reported function of the mammalian Mondo/ChREBP complexes, which are essential for the expression of genes in the same metabolic pathways [18]–[20], [41], [42]. Additionally, our analysis identified a number of stress response pathways that may be influenced by MDL-1 in conjunction with DAF-16 and PHA-4, including DNA damage response, proteostasis, unfolded protein response (UPR), oxidative stress, and thermotolerance (Figure 5E, Dataset S2).
If DAF-16 and Myc family members converge on common target genes, then they should have similar functions in longevity control. We tested this hypothesis by assessing whether overexpression of a DAF-16::GFP translational fusion could rescue the shortened lifespan that occurs in the absence of the Myc-Mondo activation complex. Compared to wild-type control, daf-16 overexpression increased lifespan by 6 days (median lifespan of 17 and 23 days, respectively a 35.3% increase, p<0.0001) consistent with published findings [43]. Interestingly, overexpression of daf-16 completely rescued the shortened lifespan of the mxl-2(tm1516) null mutation (Figure S4). Additionally, the mxl-1(tm1530)) null mutation did not further extend lifespan when daf-16 was over-expressed (Figure S4). Thus over-expression of daf-16 is sufficient to compensate for the lack of Myc-Mondo transactivation, and loss of the Mad transcriptional repression complex does not further extend lifespan when daf-16 is overexpressed, which is consistent with the notion that these transcription factors may act at common target genes.
We next asked whether Myc-Mondo function was necessary for the activation of DAF-16 target genes under conditions of decreased ILS. To this end, we measured mRNA expression levels of several candidates by qRT-PCR, which were chosen based on the criteria that they:are bound by DAF-16, PHA-4, and MDL-1 (modENCODE and [37]); are upregulated under conditions of decreased ILS in a daf-16 dependent manner (i.e. are “Class 1” DAF-16 target genes, [37]); and have been implicated in the determination of adult longevity (GO term: 0008340). Based on these criteria five candidate genes were chosen: dod-3, an apparently nematode-specific gene previously shown to be “downstream of daf-16” [13]; icl-1, encoding isocitrate lyase and malate synthetase, which together form the glyoxylate shunt, an alternative metabolic pathway known to be favored in daf-2 mutants, dietarily restricted worms, and Mit mutants [13], [44], [45]; sodh-1 which encodes a sorbitol dehydrogenase; hsp-12.6, a small heat-shock protein; and stdh-1, a sterol dehydrogenase. Consistent with published findings, all but sodh-1 were induced by decreased ILS in a daf-16 dependent manner [37]. Of these four genes, the expression of three (icl-1, hsp-12.6, and stdh-1) were also suppressed in daf-2(e1370);mxl-2(tm1516) double mutant animals (Figure 5F). Additionally, we discovered that Myc-Mondo complex was not necessary for the increased nuclear accumulation of DAF-16 under conditions of decreased ILS, as GFP was solely nuclear in daf-2(e1370);DAF-16::GFP and daf-2(e1370);mxl-2(tm1516);DAF-16::GFP mutant animals (data not shown). We conclude that the Myc-Mondo transcriptional activation complex is necessary for the activation of at least a subset of DAF-16 target genes under conditions of decreased ILS.
Stress resistance is intimately connected to aging. For instance, reactive oxygen species cause the accumulation of oxidative damage to proteins and other cellular constituents as an organism ages [46]. Thermal stress also challenges the molecular chaperone network, leading to the collapse of proteostasis and a shortened lifespan [47]. Indeed, many of the genes known to influence longevity also function in stress response; the most studied example being ILS and DAF-16 (reviewed in [48]). DAF-16 functions in numerous forms of stress response including: heat shock, oxidative damage, exposure to heavy metals, radiation, anoxia, osmotic, and pathogen exposure [49]–[54]. Thus sensors of metabolic status or environmental damage may also regulate the expression of cytoprotective genes that influence aging. Based on the functional overlap between DAF-16 and Myc family members, we hypothesized that the Myc family influences longevity through altered stress response.
To assess whether loss of the Myc-Mondo or Mad complex altered C. elegans oxidative stress resistance, N2, mxl-2(tm1516), and mxl-1(tm1530) mutants were grown until day 2 of adulthood and survival was measured following oxidative stress imposed by exposure to tert-butylhydroperoxide (tBOOH; an organic peroxide). While loss of mxl-1 failed to improve survival under any conditions (data not shown), loss of the Myc-Mondo complex via deletion of mxl-2 impaired survival to oxidative stress (Figure 6A, Dataset S3). Specifically, following exposure to 7.7mM tBOOH, median survival was reduced from 13.6+/–0.8 hours in N2 animals, to 7.5+/–0.5 hours in mxl-2(tm1516) animals, a reduction of 45% (Figure 6A, Dataset S3). As expected, daf-2(e1370) mutant animals were much more resistant to tBOOH treatment; median survival was 25.0+/–1.0 hours. However, the median survival of daf-2(e1370);mxl-2(tm1516) double mutant animals was only 15.5+/–0.6 hours, a reduction of 38% (Figure 6A, Dataset S3). Thus loss of mxl-2 does not reduce daf-2(e1370) tBOOH survival to a greater extent when compared to N2 (p = 0.986). To our surprise, eat-2(ad465) mutant animals were significantly more sensitive to tBOOH treatment than wild-type animals (data not shown), precluding a requirement to test the effects of mxl-2 deletion in this background. Thus the Myc-Mondo complex is essential for resistance to oxidative stress and loss of the Mad complex is insufficient to increase stress resistance. Our bioinformatic analysis of oxidative stress response genes found that 12 out of 50 genes annotated as responding to oxidative stress are bound by DAF-16 and MDL-1 (i.e. no overrepresentation vs. the genome, p = 0.088). In contrast, 20 out of 50 oxidative stress response genes are bound by DAF-16, a significant enrichment (p = 0.049). Taken together, these data suggest that the C. elegans Myc-Mondo complex is generally required for resistance to oxidative stress, rather than specifically required for the enhanced resistance of the daf-2(e1370) mutant.
We next assessed whether loss of the Myc-Mondo or Mad complexes altered C. elegans thermotolerance (i.e. intrinsic thermal tolerance, ITT). N2, mxl-2(tm1516), and mxl-1(tm1530) mutants were grown until day 2 of adulthood and survival at 35°C was measured. Loss of mxl-2 significantly shortened the median survival time of wild-type animals by 22.0%, from 12.3+/–0.4 hours to 9.6+/–0.3 hours at 35°C (Figure 6B, Dataset S3, p<0.0001). Similar results were obtained in the mml-1(ok849) mutant (data not shown). Nearly identical results were obtained in the eat-2(ad465) and eat-2(ad465);mxl-2(tm1516) animals (Figure S5) confirming previous reports that eat-2 mutations do not enhance thermal stress resistance [55]. Conversely, loss of mxl-1 failed to alter survival to thermal stress (Dataset S3) in all three genetic backgrounds.
As previously reported, daf-2(e1370) mutant animals had increased thermal stress resistance compared to N2 animals (median survival of 16.6+/–0.8 hours versus 12.3+/–0.4 hours, Figure 6B, 6C, Dataset S3), and daf-16 was essential for this increased resistance (Figure 6C, Dataset 3, and [56]). daf-2(e1370);mxl-2(tm1516) had thermal stress resistance comparable to wild-type control animals (median survival of 12.4+/–0.5 and 12.3+/–0.4 hours, respectively) (Figure 6B, 6C). Thus loss of Myc-Mondo function suppresses the enhanced survival of daf-2(e1370) mutant animals and wild-type survival to similar extents (25.3% and 22.2%, respectively).
We sought to better understand the relationship between ILS, DAF-16, the Myc family of transcription factors, and thermal stress resistance. To this end we first tested whether inactivating mxl-2 in the absence of daf-16 would further impair survival to thermal stress. Inactivating mxl-2 in daf-2(e1370);daf-16(mgDf47) resulted in no significant further reduction in thermal stress survival when compared to control (median survival 8.0+/–0.11 and 7.6+/–0.10, Figure 6C, p = 0.313). We next asked whether mxl-2 was required for the acquired thermotolerance (ATT) of daf-2(e1370) mutant animals. daf-2(e1370) mutant animals grown at 25°C become extremely tolerant to thermal stress. By day 3 of adulthood daf-2(e1370) animals subsequently exposed to 35°C have a median survival of 25.1+/–0.6 hours (Figure 6D), which is significantly greater than daf-2(e1370) intrinsic thermotolerance (ITT) (p<0.0001). Similar to ITT, the acquired thermotolerance of daf-2(e1370) is dependent upon daf-16, as median survival is reduced from 25.1+/–0.6 to 17.2+/–0.53 hours (Figure 6D). In contrast, mxl-2 is dispensable for the acquired thermotolerance of daf-2(e1370) as median survival is unchanged (Figure 6D, p = 0.17). A comparison of genes that have promoter regions bound by DAF-16, PHA-4, and MDL-1 against all genes in the genome revealed a 2.17-fold enrichment (p = 0.006) for genes associated with a “response to heat”; which includes chaperones such sti-1, hsp16.2, and hsp-16.41. Collectively, we conclude that the Myc-Mondo complex: is required for intrinsic but not acquired thermotolerance; has overlapping genetic requirements with daf-16; and that MDL-1 and DAF-16 bind at the promoters of molecular chaperones implicated in the heat shock response.
We determined whether loss of C. elegans Myc-Mondo or Mad complexes might alter the decline/progressive collapse of protein homeostasis (‘proteostasis’), a hallmark of normal aging [57]. One way to assess whether a genetic perturbation affects C. elegans proteostasis is by monitoring the solubility of a polyglutamine repeat fused to YFP (Punc-54::Q35::YFP). During normal aging transgenic animals expressing such poly Gln-YFP fusion proteins in body wall muscle cells accumulate protein aggregates that can be visualized as fluorescent foci [58]. Later in life, these toxic foci overwhelm the chaperone network that maintains proper protein folding, resulting in the collapse of proteostasis within the body wall muscle cells, thus causing paralysis in the animal [59].
To test whether the Myc-Mondo and Mad complexes might function in proteostasis we inactivated mml-1, mxl-2, mdl-1, or mxl-1 in Pmyo-3::Q35::YFP animals and quantified the accumulation of fluorescent foci (protein aggregates) over time. Loss of the Myc-Mondo complex (either mml-1 or mxl-2) resulted in a significantly premature accumulation of protein aggregates (mxl-2: day 1 p = 7.6×10−4, day 2 p = 4.21×10−6, day 3 p = 9.66×10−6; mml-1: day 1 p = 5.68×10−10, day 2 p = 2.89×10−9, day 3 p = 5.86×10−8). For instance, by day 3 of adulthood mml-1(RNAi) or mxl-2(RNAi) treated Q35::YFP animals showed on average 113+/–16.7 and 108+/–22.1 foci protein aggregates, respectively. In contrast, control RNAi animals had an average of 72+/–13.6 foci per animal (Figure 7A, Dataset 3). This is consistent with the notion that loss of the Myc-Mondo complex compromises the protein chaperone network. In contrast, inactivation of the Mad complex (mdl-1 or mxl-1) had no effect on polyglutamine foci formation (Figure 7A, Dataset 3). For example, at day 3 adulthood, Q35::YFP animals treated with mdl-1 or mxl-1 RNAi had on average 70+/–11.7 and 75+/–10.4 foci, which is not significantly different from animals fed control RNAi (mdl-1 p = 0.711, mxl-1 p = 0.458). Of note, inactivation of daf-2 also failed to appreciably delay the accumulation of foci (Figure 5A): the average number of foci at day 3 adulthood is 70+/–18.0 (p = 0.791). Thus loss of the Mad complex or decreased ILS does not appreciably alter the acute accumulation of protein aggregates in Q35::YFP animals, but loss of the Myc-Mondo complex hastens the collapse of proteostasis as measured by the accumulation of fluorescent aggregates.
We assessed the long term consequences of loss of Myc-Mondo or Mad complex function in Punc-54::Q35::YFP animals as measured by the onset of paralysis. daf-2(RNAi) served as a positive control and completely negated the onset of paralysis over the course of our analysis, consistent with published results [60]. In contrast, the average age of paralysis of Q35::YFP animals treated with control RNAi was day 7 of adulthood, and all animals were paralyzed by day 9. Inactivation of mml-1 or mxl-2 resulted in a significant premature onset of paralysis (mxl-2 p = 0.0000 and mml-1 p = 4.93×10−11, Figure 7B). Conversely, inactivation of mdl-1 or mxl-1 resulted in a significant (mxl-1 p = 0.0187; mdl-1 p = 0.0384) delay in the onset of paralysis (Figure 7B). Collectively, our findings indicate that the C. elegans Myc-Mondo and Mad complexes have opposing effects on proteostasis, which parallel their effect on longevity and known roles in transcriptional control [11].
The Myc family of transcription factors have well established functions in growth control and metabolic regulation; processes that are known to influence lifespan. Yet, this is the first study to identify a role for the Myc interaction network in longevity control and proteostasis. We have identified a role for the Myc-family of bHLH transcription factors: mml-1, mdl-1, mxl-1, and mxl-2 in the regulation of C. elegans lifespan. We show that the Myc-Mondo (MML-1:MXL-2) and Mad (MDL-1:MXL-1) transcriptional complexes have opposing roles in longevity control and proteostasis analogous to their known functions as transcriptional activators and repressors in C. elegans [11]. Specifically, loss of the Myc-Mondo complex leads to premature aging, while loss of the Mad complex delays aging. Lifespan extension by loss of the Mad complex is dependent upon the Myc-Mondo complex, daf-16, and pha-4, suggesting a common mechanism in longevity control. Conversely, Myc-Mondo is required for lifespan extension by decreased ILS or conditions of DR. Both decreased ILS and conditions of DR promote nuclear accumulation of MML-1, but do so through distinct mechanisms; altered DAF-16 activity regulates expression of Myc family members, while pha-4 is essential for the increased nuclear accumulation of MML-1. In contrast, the Myc-Mondo complex is dispensable for the nuclear accumulation of DAF-16 by decreased ILS, and overexpression of DAF-16::GFP rescues the shortened lifespan of mxl-2 null mutant animals, which suggests that DAF-16 and the Myc-Mondo complex may co-regulate transcription at overlapping target genes. From a candidate approach, we find that Myc-Mondo is required for the induction of longevity genes by decreased ILS. DAF-16, PHA-4, and MDL-1 bind within the genome at many overlapping target genes involved in unfolded protein binding, carbohydrate metabolism, autophagy, and stress response. Finally, loss of the Myc-Mondo complex impairs oxidative and thermal stress survival. Collectively, our results suggest that Myc family members are regulated by diverse signals of metabolic status and converge with DAF-16 and PHA-4 at metabolic and cytoprotective transcriptional programs that influence aging (Figure 8).
The Myc interaction network in mammals has been defined based on the combinations of known protein-protein interactions between Myc/Mnt/Mad/Mondo and either Max or Mlx, which distinguishes them from the larger super-family of bHLH transcription factors [61]–[64]. The discovery that the Myc-Mondo and Mad complexes have antagonistic functions in both transcription and longevity control sets them apart from other transcription factors that are relevant to aging, such as DAF-16, PHA-4, or SKN-1. The related bHLH protein, MXL-3, has recently been implicated in lifespan [23], [65]. mxl-3 encodes a paralog of mxl-1 (Max), which is unique to nematodes [22]. Furthermore, MXL-3 homodimerizes and does not interact with other C. elegans Myc family members [10]. This is in contrast to mammalian Max and C. elegans MXL-1 (Max), which function as heterodimers. Additionally, there is little overlap between predicted MXL-3 and MDL-1::MXL-1 target genes [10]. In C. elegans, MXL-3 functions in conjunction with HLH-30 to regulate the expression of a family of lipases, which mobilize lysosomal fat stores, and function independent of ILS and DR to influence longevity [23], [65]. We find no genetic interaction between the C. elegans Myc-Mondo or Mad complexes and mxl-3, implying that MXL-3 functions are distinct from the function of the Myc-Mondo/Mad interaction network.
MML-1, (for ““Myc and Mondo-like”), has a high degree of sequence identity to c-Myc, MondoA, and MondoB/ChREBP. The C. elegans genome does not encode a clear Myc ortholog [22]. Therefore, it is likely that C. elegans MML-1 incorporates both mammalian c-Myc and Mondo functions. Myc proteins have been implicated in ribosomal and mitochondrial biogenesis, energy metabolism, biosynthesis, cell growth, and cell cycle regulation. For example, many glucose metabolism genes are directly regulated by Myc [66]–[71]. In this way, Myc promotes glucose transport and catabolism. C. elegans MML-1 has an overall domain structure and size similar to mammalian Mondo/ChREBP. Work in mammals has identified MondoA:MLX and ChREBP:MLX complexes as non-hormonal sensors of glucose and key regulators of glycolytic metabolism, fat storage, and energy sensing [20], [72]. Recently, ChREBP has been shown in adipose tissue to regulate systemic glucose metabolism, fatty acid synthesis, and insulin sensitivity [73]. While the genomic binding patterns of MML-1 remain to be determined, we find an overrepresentation of genes involved in carbohydrate metabolism bound by MDL-1, DAF-16, and PHA-4. Thus the Myc-Mondo/Mad complex functions relevant to longevity may be at least partially linked to evolutionarily conserved alterations in carbohydrate metabolism.
Surprisingly, there have been few large-scale efforts to identify Mad targets in mammalian systems. Individual Mad target genes have largely been identified by the examination of genes previously identified as Myc targets [40]. However, one study overexpressed Mad1 in T-lymphocytes and identified reduced expression in 57 genes. [39]. Collectively, of 64 putative Mad1 target genes we found 53 clear C. elegans homologues, of which 41 are MDL-1 targets (Figure 5, Dataset S2). Many of these homologues are components of important cellular processes such as protein translation/ribosome function (23), metabolism (7), DNA transcription (2), and cell cycle control (2). Moreover, 11 have been implicated in C. elegans longevity: the eukaryotic initiation factors eif-1/eIF1, ifg-1/eIFGI, ife-2/eIF4E, and inf-1/eIF4AI; elongation factors eef-1a.2/EF1 and eef-2/EF2; the ribosomal protein rpl-18/Rpl18; the PTEN homologue daf-18; the glucose-6-phosphate isomerase gpi-1; the mitochondrial stress protein hsp-6; and the cytochrome b-c1 complex subunit ucr-1. Thus while the full extent of bona fide Mad target genes in mammals is relatively unknown, we find good overlap at conserved putative target genes, suggesting that the expression of Mad target genes relevant to longevity may be conserved.
We have found that decreased ILS and conditions of DR regulate C. elegans Myc family activity. Both decreased ILS and conditions of DR promote nuclear accumulation of MML-1, presumably to alter gene expression. In this respect MML-1 resembles its mammalian homologues, MondoA and ChREBP, which are also regulated by cytoplasmic-to-nuclear shuttling under changing conditions of nutrient availability. In contrast to mammalian Mondo complexes, Mad complexes are reported to be constitutively nuclear [74], [75]. Furthermore, we find that DAF-16 activity regulates Myc family expression. This is reminiscent of the reported function of mammalian MAD1, which is upregulated at the transcriptional level by several different signaling pathways, including activated PI3K/Akt downstream of the granulocyte-colony stimulating factor receptor (G-CSFR) [76]. This contrasts with findings that identified mdl-1 as an upregulated DAF-16 target gene by microarray analysis [13], [77]. Furthermore, a previous study found that mdl-1 inactivation slightly shortened daf-2 mutant longevity [13]. We find no effect of loss of mdl-1 on daf-2(e1370) longevity (data not shown). One possible explanation is that the published lifespan analysis was conducted at 25°C, which we find is a slightly stressed state that alters the genetic requirements for lifespan extension by decreased ILS (A.V.S. unpublished observations).
We found that the Myc-Mondo and Mad complexes function in both DR and ILS signaling, implying that the Myc-Mondo/Mad complexes may act as a molecular convergence point to integrate cues of metabolic status, and coordinate the transcriptional response to these diverse signals. A number of metabolic signals are widely accepted as being intimately linked to aging, yet how these signals are integrated to “fine-tune” the appropriate transcriptional response to maximize longevity under adverse environmental conditions is unknown. Interestingly, Myc has recently been shown to function as a “universal amplifier” of transcription; instead of having distinct target genes, Myc amplifies the output of the existing gene expression [78], . Thus the Myc-Mondo and Mad complexes’ antagonistic effect on transcription may function as a rheostat to modulate gene expression in response to multiple inputs of metabolic status, which is a possibility mutually exclusive from regulating transcription of distinct target genes.
Our results raise a number of questions of how Myc family members integrate diverse signals of metabolic status. The roles of ILS and DR in longevity control are genetically separable, yet there is evidence of overlap between the transcriptional programs that are activated by decreased ILS and conditions of DR (Reviewed in [1]). For instance, overexpression of pha-4 has the greater influence on lifespan in the absence of daf-16 [8], which suggests either an inherent competition between daf-16 and pha-4 in wild-type animals, or that the role of daf-16 and pha-4 may be partially redundant. However, pha-4 is dispensable for the increased lifespan of daf-2 mutants [8], yet wefind that pha-4 is necessary for the increased nuclear accumulation of MML-1::GFP by decreased ILS. One relevant finding is that loss of pha-4 has no detrimental effect on the distribution of MML-1::GFP localization in wild-type animals, which still have substantial amounts of nuclear MML-1. Thus the basal distribution of MML-1 may be sufficient to potentiate insulin/IGF1 signals either through altered patterns of DNA binding, recruitment of co-factors, or changes of stability at DNA. However, our results predict that there would be no additivity of lifespan extension by the combination of DR and decreased ILS in the absence of pha-4. The mechanism through which PHA-4 potentiates DR signals through increased nuclear accumulation of MML-1 remains to be determined. Nuclear localization of ChREBP, the mammalian homologue of MML-1, is regulated by changes in carbohydrate availability through post-translational modifications [80]. Thus identifying whether MML-1 is similarly regulated will be an important step to further understand how DR signaling alters Myc family function.
We find that overexpression of DAF-16 can rescue the shortened lifespan caused by the loss of Myc-Mondo, yet Myc-Mondo is required for the increased expression of DAF-16 target genes under conditions of decreased ILS. How might these two paradoxical findings be resolved? One possibility is that DAF-16 activity is not maximally induced by daf-2(e1370). Our discovery that the daf-16 dependent acquired thermotolerance of daf-2(e1370) is independent of mxl-2 is consistent with the possibility that mild heat stress in daf-2(e1370) maximally potentiates DAF-16 activity, thereby removing the need for the Myc-Mondo transcriptional activation complex. A second possibility is that DAF-16::GFP overexpression is not equivalent to the activation of DAF-16 in daf-2(e1370) mutant animals.
We sought to identify putative common target genes between DAF-16, PHA-4, and MDL-1 to begin to elucidate the molecular mechanisms through which the Myc family may influence longevity. In contrast to DAF-16/PHA-4, MDL-1 and the Mad mammalian homologues are only known to function as transcriptional repressors [11], [40]. Importantly, MDL-1 genomic binding patterns were identified in wild-type animals; thus under normal conditions the MDL-1 complex limits the expression of DAF-16/PHA-4 target genes relevant to aging. We favor a model where under conditions of reduced ILS or DR, increased nuclear accumulation of MML-1 results in the replacement of MDL-1 binding at target genes. This idea is consistent with the model of complex switching that has been proposed to describe the antagonistic relationship between Myc and Mad complexes in mammals [40]. We found significant overlap of genomic binding between these transcription factors at many target genes, consistent with the possibility of shared transcriptional programs. Assessing whether these putative common target genes were enriched over the genome for any specific biological function by gene ontology revealed a significant enrichment for genes that function in lifespan (299), stress response (117), carbohydrate metabolism (93), unfolded protein binding (27), and autophagy (9). Specific forms of stress response that were over-represented included response to heat (12), unfolded protein response (20), immunity/pathogen resistance (15), and DNA damage response (49). Surprisingly, response to oxidative stress was not over-represented. However, we found no evidence that heat or oxidative stress altered the expression of Myc family members, nor the nuclear accumulation of MML-1 (D.W.J., unpublished observations). Thus despite their requirement for stress response survival, and the finding that MDL-1 binds at many genes involved in survival to stress, Myc family members do not seem to be stress response genes, which is in contrast to their regulation by altered signals of metabolic status. Rather, these data support a model where the Myc family converges with DAF-16 and PHA-4 at overlapping metabolic and cytoprotective transcriptional programs that set the progression of aging.
Given their close ties to metabolism it is not surprising that DAF-16 and PHA-4 bind extensively in the promoter regions of key metabolic genes. Our analysis shows that MDL-1, DAF-16, and PHA-4 binding is significantly enriched at genes involved in carbohydrate metabolism (Dataset S2), including nine glycolytic genes (Y77E11A.1, H25P06.1, enol-1, pyk-2, ZK632.4, gpi-1, aldo-2, F57B10.3, pgk-1), and five gluconeogenesis genes (gpi-1, aldo-2, F57B10.3, pgk-1, pyc-1), which is strongly reminiscent of mammalian Mondo-ChREBP/Mlx complexes [18]-[20], [42]. Interestingly, it has been found that gluconeogenesis is upregulated in conditions of low ILS and DR [44], [81]; therefore, it is possible that one role for Myc-Mondo complexes under these conditions is to promote the expression of gluconeogenic genes, which results in the increase of intracellular glucose. Other metabolic genes with known roles in longevity included icl-1 and stdh-1. Both are metabolic enzymes that are upregulated by decreased ILS, in a daf-16 dependent manner [37]. We discovered their upregulation by decreased ILS is also dependent on the Myc-Mondo complex (Figure 5F). stdh-1 encodes a steroid dehydrogenase, which is involved in hormone biosynthesis and suggests that the Myc family may contribute to cell non-autonomous signaling. icl-1 encodes a dual enzyme isocitrate lyase/malate synthase essential for the glyoxylate cycle, a variation of the citric acid cycle that allows the production of sugar from fat. Like gluconeogenesis, the expression of icl-1 is also upregulated in conditions of low ILS and DR [13], [44], and increased glyoxylate cycle activity is required for the increased longevity of C. elegans insulin signaling mutants [13].
Proteostasis is the ability of the cell’s proteome to maintain a proper folding environment. The predominant theory is that during aging, post-mitotic cells accumulate protein damage (e.g. through oxidative damage and DNA mutations), which in turn produces misfolded proteins that rapidly form highly stable protein aggregates that ultimately overwhelm the ability of the chaperone network, protein degradation machinery, and sequestration strategies to maintain a proper folding environment. This collapse may be the underlying basis for age-associated proteotoxic diseases [82]. We find that the Myc-Mondo and Mad complexes have opposite effects on proteostasis, analogous to their effect on transcription and longevity. Since many putative target genes of MDL-1 are involved in either protein folding, unfolded protein response, or heat-shock response, we favor a model where changes in Myc family activity alter the expression of genes involved in protein homeostasis, which in turn influences the progression of aging.
Putative target genes of MDL-1 and DAF-16 were enriched over genes bound only by DAF-16 in two major gene ontology terms that are highly relevant for the maintenance of proteostasis: unfolded protein binding and the unfolded protein response. These somewhat mutually exclusive categories cover de novo protein folding and the response to proteotoxic stress, respectively. Twenty-eight of 52 genes annotated to function in unfolded protein binding were discovered and include the gene classes: cct (Chaperonin Containing TCP-1; 6 of 8), pfd (PreFolDin; 4 of 6), and dnj (DNaJ domain; 11 of 29) (Dataset 3). The cct genes encode the components of the TCP-1 ring complex (TRiC), which complete de novo protein folding and interacts with prefoldin co-chaperones. The TRiC complex has been implicated in longevity control and an ILS-mediated soma to germline transformation gene expression program [83], [84]. In addition, 20 of 40 genes annotated to function in the unfolded protein response are bound by DAF-16, PHA-4, and MDL-1. Genes involved in the mitochondrial and ER unfolded protein responses were represented (3 and 9 genes respectively, Dataset S2). Lifespan extension in daf-2 mutants is dependent on both ire-1 and xbp-1, suggesting that genes involved in ER proteostasis may contribute to daf-2 mutant lifespan [85]. We found several small heat shock proteins (hsp-16.2, hsp-16.41, and hsp-12.6) to be bound by DAF-16, PHA-4, and MDL-1. The latter is not stress responsive, but is upregulated in daf-2 mutant animals dependent on daf-16 [13], [86], [87], and we find this upregulation is also mxl-2 dependent (Figure 5D). Interestingly, hsp-12.6 is required for the increased lifespan of ILS mutants, and hsp-12.6(RNAi) results in the premature collapse of proteostasis [88].
This is the first study to identify a role of the Myc-Mondo/Mad interaction network in longevity. Significantly, a homologue of one of these transcription factors may also influence aging in mammals. The mml-1 mammalian homologue ChREBP is also known as William-Beuren Syndrome Chromosomal Region 14 (WBSCR14). Williams-Beuren Syndrome (WBS) is a multisystem disorder resulting from the deletion of 26 to 28 genes, including ChREBP, on human chromosome 7. WBS has been characterized as a disorder of mild accelerated aging. Patients display premature gray hair, diverticulosis, diabetes, and hearing loss that develops during adolescence or young adulthood, in combination with instances of declining memory skills or dementia [21]. Thus our findings collectively imply that the Myc-Mondo/Mad interaction network of basic-helix transcription factors function as a molecular convergence point: integrating diverse signals of metabolic status that converge at evolutionarily conserved metabolic and cytoprotective transcriptional programs to influence the progression of aging.
Standard culture techniques were used to maintain nematodes [89]. The wild-type strain is Bristol N2 and all other strains used in this work were backcrossed into Bristol N2 a minimum of six times before use. The mxl-2(tm1516), mxl-1(tm1530), and mdl-1(tm311) strains were created by and obtained from Dr. Shohei Mitani with the Japanese Bioresource Project; the mml-1(ok849) and mxl-3(ok1947) were generated by Dr. Robert Barstead with the Oklahoma Medical Research Foundation Knockout Group, and were generously provided by the C. elegans Genetics Center (University of Minnesota).
The Pmml-1::mml-1::gfp construct was generated through PCR sewing [90]. The mml-1 genomic fragment was PCR amplified from the WRM0615CE07 fosmid using the following primers: 5′-AGTCGACCTGCAGGCATGCAAGCTaaatggatttttgagttgttgcat-3′ and 5′-TAGAGGGTCGCGTTAGAGGT-3′. The GFP open-reading frame was PCR amplified from pPD95.75 (Fire Lab vector kit) using the following primers: 5′-AGCTTGCATGCCTGCAGGTCG-3′ and 5′-AAGGGCCCGTACGGCCGACTA-3′. The two fragments were then PCR sewn together using the following primers: 5′-TTCCACTCGTTTCTCCGTCC-3′ and 5′-GGAAACAGTTATGTTTGGTATA-3′. Eight sewing reactions were simultaneously performed. Each reaction generated a single PCR product, which were combined to form a single pool of DNA which was injected into mml-1(ok849) young adult hermaphrodites at a concentration of 5 ng/μL along with 2.5 ng/μL pCFJ90 as a co-injection marker, and 142.5 ng/μL 2-log DNA ladder (New England Biolabs). The transgene was maintained as a high-transmission rate extrachromosomal array. Three independent mml-1(ok849) lines expressing this array were subsequently mated into daf-2(e1370) and eat-2(ad465) mutant to generate independent lines for examination. Three independently generated lines were examined per background.
The strains used in this work are: N2 Bristol, eat-2(ad465), daf-2(e1370), mxl-2(tm1516), mxl-1(tm1530), mdl-1(tm311), mml-1(ok849), mxl-3(tm1947), mxl-2(tm1516);daf-2(e1370), mxl-1(tm1530); daf-2(e1370), mxl-2(tm1516);eat-2(ad465), mxl-1(tm1530);eat-2(ad465), AM140 rmIs132 [Punc-54::Q35::YFP] [58]; AVS317-319 mml-1(ok849);artEx6-8 [Pmml-1::mml-1::gfp; pCFJ90] (three independent lines), AVS329-331 daf-2(e1370);artEx6-8 [Pmml-1::mml-1::gfp; pCFJ90] (three independent lines), AVS332-334 eat-2(ad465);artEx6-8artEx3 [Pmml-1::mml-1::gfp; pCFJ90] (three independent lines). RNAi clones were grown overnight in Luria broth and seeded onto plates containing 5 mM isopropylthiogalatoside, to induce dsRNA expression overnight at room temperature.
Lifespan and survival assays were performed using either a replica set protocol, as previously described [12], or through traditional methods. Briefly, for replica set protocol: rather than following a single population of animals over time, animals were cultured in 24-well format on HT115 bacteria expressing either empty vector or the indicated RNAi. A sufficient number of plates were created for each condition (combination of strain and RNAi) so that each could be counted at least every other day and then disposed. On average, 25 animals were counted per well. In all cases animals were cultured from L1 to L4 stage of development at 15°C, at which point FuDR was added to a final concentration of 400 μM and the animals were transferred to 20°C for the remainder of their lifespan. For traditional survival assays, proportional survival was charted over time using the Kaplan-Meier estimator. Replicate survival assays were fit via generalized linear modeling to a logit curve, according to the equation y = 1/(1+e∧(a*x-b)). The LD50 for a fit survival curve was then a function of the coefficients: LD50 = b/a. For two sample comparison, p-values were calculated through resampling and refitting, where the difference in resampled LD50s was compared to the difference in the observed LD50s, see [12]. Similar analyses were used on stress survival assays. For each comparison of either a mutant or RNAi treatment compared to control, between two and four independent trials were conducted to measure alterations in lifespan or survival (see figure legends and Dataset S1). Raw mortality data from replicate set lifespan has been deposited into the Dryad repository (doi:10.5061/dryad.pj0p3 Data files: Raw Data of Mortality Observations of Replica Set Experiments).
Synchronized L1 transgenic animals were grown to the L4 stage of development on HT115 bacteria, expressing: empty vector control, daf-16, or pha-4 RNAi at 15°C on 6 cm plates. FuDR was added to a final concentration of 400 μM and the animals were transferred to 20°C. GFP localization was examined on day 2 of adulthood. Worms were mounted on 2% agarose pads and anaesthetized with a mixture of 15 mM tricaine mesylate and 15 mM tetramisole hydrochloride prior to imaging. Microscopy was performed on a Zeiss AxioImager.M2m equipped with an AxioCam MRm camera using high NA Zeiss EC-Plan-NEOFLUAR 20X and 40X objectives, and a Semrock Brightline GFP filter (488 nm/535 nm Ex/Em). Scoring was performed in three separate trials examined under blinded conditions, and each animal was scored according to the rubric shown in (Figure 4).
Synchronized populations were grown to young adulthood at 15°C on 10 cm RNA plates containing HT115 bacteria expressing vector control RNAi. Animals were then rinsed from the plates using M9 and washed three times to remove bacteria. RNA was isolated using Trizol reagent (Life Sciences) following the manufacturer’s protocol. Isolated RNA was converted to cDNA using the Bio-Rad iScript reverse transcriptase kit with oligo dT primers. Negative samples were prepared by omitting the reverse transcriptase enzyme from the reaction.
Analysis of transcript levels via qRT-PCR was performed on three distinct biological replicates per condition, and performed in triplicate in each instance. The total cDNA concentration for each sample was normalized to the levels of act-1 and rpl-32 transcript. mml-1, mxl-2, mxl-1, and mdl-1 transcript levels were normalized to N2 control samples. For analysis of Class I genes: act-1, rpl-32, and pat-10 were used to normalize total cDNA concentration. dod-3, hsp-12.6, stdh-1, and icl-1 transcript levels were normalized to N2 control samples. Statistical analyses of fold-changes from three independent trials were accomplished using Student’s paired t-test.
The primers used are as follows:
act-1:forward 5′-CCA TCA TGA AGT GCG ACA TTG-3′
reverse 5′-CAT GGT TGA TGG GGC AAG AG-3′
rpl-32:forward 5′-AGG GAA TTG ATA ACC GTG TCC GCA-3′
reverse 5′-TGT AGG ACT GCA TGA GGA GCA TGT-3′
mml-1:forward 5′-GGA GAA ATC CGG AAG GCA TTA CTT AC-3′
reverse 5′-CGC CAG AAT ACA AAT CGG GTA TAA TAT C-3′
mxl-2:forward 5′-TCA GAG CCT GCG ACT TCA TG-3′
reverse 5′-GTC GAG GAG AAG TTG GAG CAT-3′
mxl-1:forward 5′-GAC ATG AGT GAC CTC GAA GAT GAC-3′
reverse 5′-CAG GCG AGC TAT CTC TTC TCT G-3′
mdl-1:forward 5′-CGA TCT TTC AAA TGA GTC CGA ACT CC-3′
reverse 5′-GTC TTG CAA TTC AAT GAT GTG ATC TCG-3′
pat-10: forward 5′-GACGGAAAGCTTCACGAAGTTC-3′
reverse 5′-CCTTCGTAAACTGATCCGCAAG-3′
dod-3: forward 5′-AAAAAGCCATGTTCCCGAAT-3′
reverse 5′-GCTGCGAAAAGCAAGAAAAT-3′
hsp-12.6: forward 5′-GGAGTTGTCAATGTCCTCGACG-3′
reverse 5′-GAAGTTCTCCAATGTTCTTGAC-3′
stdh-1: forward 5′-ACAGGATGTCTTCAAAAGGAATATGG-3′
reverse 5′-TTGCTGGGGTGATAGCTTGG-3′
icl-1: forward 5′-GACTACGAGGCTGGAAGAACGATTG-3′
reverse 5′- GTAGGCGAACATCTTGTCTGGGTAC-3′
Statistical analysis of traditional lifespan analysis was performed by two sample tests, which used the Mantel-Cox log-rank statistic. For replica set lifespan and stress survival experiments, data were fit to a two parameter logit survival curve, a generalized linear model with a binomial link function. This curve models the probability of a worm's survival at a given time point t, according to the equation p = 1/(1+e∧(+ax-b)). The parameters a and b, which represent logit slope and intercept, are optimized by iteratively reweighted least squares. For a fit logit curve, the LD50 is b/a, or intercept divided by slope. For maximum life, we took the value of the logit curve where it dropped below 5% survival. We estimate 95% confidence intervals on the LD50 non-parametrically, through resampling of observations (K = 10,000), with replacement. For two sample tests comparing condition LD50s to control, we again used a resampling approach, where labels were swapped between conditions compared, then curves refit, and the resampled difference in LD50 was compared to the observed difference, (K = 10,000).
To compare the effect loss that mxl-2 had in different backgrounds, we compared the ratio of mxl-2 loss of function LD50 over control LD50 between backgrounds. To put confidence intervals on this ratio, and make two sample comparisons, we again used resampling to estimate the distribution (K = 10,000 for each of the four conditions).
The binding site locations for PHA-4, MDL-1, or DAF-16 were identified based on ChIP-seq data generated by the modENCODE project (www.modencode.org). ChIP-seq binding peaks for DAF-16, PHA-4, and MDL-1 were validated in [38]. Binding sites were assigned to specific genes based on location data from WormBase (www.wormbase.org) and then PHA-4 and MDL-1 binding sites were mapped relative to the transcriptional start site (TSS) of nearby open reading frames. Examination of genome-wide binding patterns revealed that MDL-1 and PHA-4 binding was most highly enriched between –700bp and +100bp relative to the TSS. Based on this we assessed the binding of PHA-4 and MDL-1 within this window relative to each gene in the genome. Genes bound by DAF-16 were identified in [37].
For characterizing ontologies among genes bound by MDL-1, we used a list of GO terms relevant to aging mined from WormBase version 210. To evaluate GO term enrichment between a set of bound genes and the genome we used the chi-squared test.
For stress response assays, synchronized L1 animals were cultured on HT115 bacteria expressing the identified RNAi vector at 15°C until the L4 stage of development. FuDR was then added to a final concentration of 400 μM, and the animals were then cultured at 20°C unless otherwise stated. For thermotolerance assays, two day old adult animals were exposed to 35°C for the indicated time period, then removed and allowed to recover for 1–2 hours at room temperature before survival was then scored. For oxidative stress survival assays, two day old adult animals were treated with 7.7 mM tert-butylhydroperoxide (Alfa Aesar) for the indicated time period, then survival was scored. For proteostasis assays, foci formation in AM140 animals was scored on days one, two, and three of adulthood. Scoring was performed on a Zeiss M2BIO upright fluorescence microscope using a 10X objective and an EGFP filter set. Worms were mounted on 2% agarose pads and anaesthetized with a mixture of 15 mM tricaine mesylate and 15 mM tetramisole hydrochloride prior to imaging. For paralysis assays animals were scored in their ability to retreat from gentle nose touch. Paralyzed animals were removed, and scoring continued until the entire population had become paralyzed. Foci accumulation and paralysis were measured in three separate trials each. For stress response assays two separate trials were performed in duplicate.
|
10.1371/journal.pgen.1006130 | A Genomic Map of the Effects of Linked Selection in Drosophila | Natural selection at one site shapes patterns of genetic variation at linked sites. Quantifying the effects of “linked selection” on levels of genetic diversity is key to making reliable inference about demography, building a null model in scans for targets of adaptation, and learning about the dynamics of natural selection. Here, we introduce the first method that jointly infers parameters of distinct modes of linked selection, notably background selection and selective sweeps, from genome-wide diversity data, functional annotations and genetic maps. The central idea is to calculate the probability that a neutral site is polymorphic given local annotations, substitution patterns, and recombination rates. Information is then combined across sites and samples using composite likelihood in order to estimate genome-wide parameters of distinct modes of selection. In addition to parameter estimation, this approach yields a map of the expected neutral diversity levels along the genome. To illustrate the utility of our approach, we apply it to genome-wide resequencing data from 125 lines in Drosophila melanogaster and reliably predict diversity levels at the 1Mb scale. Our results corroborate estimates of a high fraction of beneficial substitutions in proteins and untranslated regions (UTR). They allow us to distinguish between the contribution of sweeps and other modes of selection around amino acid substitutions and to uncover evidence for pervasive sweeps in untranslated regions (UTRs). Our inference further suggests a substantial effect of other modes of linked selection and of adaptation in particular. More generally, we demonstrate that linked selection has had a larger effect in reducing diversity levels and increasing their variance in D. melanogaster than previously appreciated.
| One of the major discoveries in modern population genetics is the profound effect that natural selection on one locus can have on genetic variation patterns at linked loci. Since the first evidence for linked selection was uncovered in Drosophila melanogaster over two decades ago, substantial effort has focused on quantifying the effects and on distinguishing the relative contributions of purifying and positive selection. We introduce an approach to jointly model the effects of positive and negative selection along the genome and infer selection parameters. To this end, we consider how closely linked each neutral site is to different types of annotations and substitutions. When we apply the inference method to genome-wide data from 125 D. melanogaster lines, our model explains most of the variance in diversity levels at the megabase scale and allows us to distinguish among the contribution of different modes of selection on proteins and UTRs. More generally, we provide a map of the effects of natural selection along the genome, and show that selection at linked sites has had an even more drastic effect on diversity patterns than previously appreciated. We also make a tool available to apply this approach in other species.
| Selection at one site distorts patterns of polymorphism at linked neutral sites, acting as a local source of genetic drift. While the qualitative effects of “linked selection” are undisputed, quantifying them and understanding their source has been one of the central challenges in evolutionary genetics over the past two decades [1–17].
Indeed, characterizing the effects of linked selection is of central importance in many contexts. If linked selection introduces substantial heterogeneity in rates of coalescence along the genome, then obtaining accurate estimates of demographic parameters requires a genomic map of these effects [18,19]. Such maps would also serve as improved null models for other population genetic inferences, such as scans for recent targets of adaptation that rely on outlier approaches [20–22]. Moreover, an accurate characterization of the effects of linked selection carries extensive information about the selective pressures that shape genome evolution. Understanding how the effects vary among taxa would also inform long-standing questions about the determinants of levels of genetic diversity and genetic load within species [23,24,25, 26].
Patterns of genetic variation are informative about natural selection at linked sites because the effects of linked selection vary with the mode and parameters of selection. For instance, “classic” selective sweeps, in which a newly-arisen beneficial mutation is quickly driven to fixation, reduce genetic variation at nearby sites over a scale that depends on the strength of selection and rate of recombination [2,3]. Other modes of adaptation, including partial and soft sweeps, cause similar, although more subtle effects [27–31]. Background (purifying) selection against deleterious mutations also reduces diversity levels at linked sites over a scale that depends on the strength of selection and rate of recombination but to an extent that depends on the density of selected sites [5,8,9,32–34].
Until recently, evidence for the effects of linked selection was sought in the relationships between diversity patterns and factors that are expected to influence the strength and frequency of selection [13–15,17]. For example, both positive and negative linked selection should have a greater effect in regions with lower recombination rates, because, on average, a neutral site would be linked to more selected sites. Consistent with this expectation, diversity levels are positively correlated with rates of recombination in Drosophila melanogaster and several other species [4,35,36]. By a similar argument, linked purifying selection should be stronger in regions with a greater density of functional sites (e.g., coding regions) and the effects of sweeps should be greater in regions with more functional substitutions (e.g., non-synonymous substitutions). In accordance with these expectations, diversity levels decrease with the density of amino acid substitutions in Drosophila species [11,12] and in humans [37], and decrease with the density of coding and putatively functional non-coding regions in Drosophila [38], humans [18,35,37] and other species (e.g., [39,40] and cf. [17]).
Beyond providing compelling evidence for the importance of linked selection, these relationships can be used to estimate selection parameters [6,10–12]. These inferences, however, suffer from severe limitations. First, it is difficult to distinguish between the effects of different modes of linked selection, with two decades of effort focused on distinguishing the effects of classic selective sweeps from those of background selection [5,7,10,14,17,31]. Second, even when a specific mode of selection is assumed, some parameters remain poorly identifiable (e.g., the rate and strength of beneficial substitutions in sweep models [10,14]). These inferences also appear to be strongly affected by the genomic scale over which they are evaluated [14].
An alternative approach is to take advantage of spatial diversity patterns along the genome. Pioneering efforts in D. melanogaster used estimates of the genome-wide rate of deleterious mutations, genetic maps, and the spatial distribution of constrained genomic regions, to demonstrate that background selection could account for changes in diversity levels along chromosomes as well as for differences in diversity levels between X and autosomes ([41–43]). More recently, McVicker et al. [18] used ancestral diversity levels along the genome in order to build a map of the effects of background selection along the human genome. The central idea was to calculate the probability that a neutral site is polymorphic, given its genetic distance from conserved coding regions and the rate of deleterious mutation and distribution of selection effects at these regions; selection parameters were then estimated by maximizing the composite-likelihood for neutral polymorphisms along the genome. Although based on limited data, the map inferred by this approach provides an impressive fit to diversity patterns on the mega-base scale. However, the associated estimate of the deleterious mutation rate is unreasonably high, more than four-fold greater than estimates of the total spontaneous mutation rate [44–47], possibly reflecting the absorption of the effects of background selection from other, poorly annotated functional regions or the effects of positive selection [18].
Another recent approach to learn about selective sweeps relies on plots of the average levels of diversity as a function of distance from amino acid substitutions throughout the genome [48–50]. Assuming that some of the substitutions resulted from classic sweeps, we would expect a trough in diversity levels around substitutions, with the depth related to the fraction that were beneficial and the width (in units of genetic distance) reflecting the strength of selection. The rate and strength of classic sweeps can thus be inferred from the shape of the trough. Applying this methodology to data from D. simulans, Sattath et al. [48] found a trough in neutral diversity levels around amino acid substitutions that extended over ~15 kb, but not around synonymous substitutions (which served as a control). The collated plot approach has several limitations, however. First, application of the same approach to human data [49] suggests that background selection, which is concentrated in or near coding regions, may contribute to the troughs in diversity, and thus could bias estimates of positive selection parameters. Second, inferences based on collated diversity patterns account only for the average clustering of amino acid substitutions and not for their spatial distribution around every neutral site.
Here, we combine the advantages of these two recent approaches [18,48] in order to infer selection parameters and build a genomic map of the effects of linked selection, considering background selection and classic selective sweeps jointly. We model the effects of background selection using the annotations for linked sites, and those of classic sweeps by considering linked, putatively functional sites that experienced a substitution. The method is applicable to genome-wide polymorphism data, allowing for information to be combined across samples. As an illustration, we apply our method to genome-wide resequencing data from 125 lines of Drosophila melanogaster (from the DGRP [51]). We also make software available for the approach to be applied more broadly.
We model the effects of background selection and classic sweeps on neutral heterozygosity (i.e., the probability of observing different alleles in a sample size of two), π, at an autosomal position x. In a coalescent framework, the model takes the form
π(x)=2u(x)2u(x)+1/(2NeB(x))+S(x),
(1)
where u(x) is the local mutation rate, Ne is the effective population size without linked selection, B(x) is the local (multiplicative) reduction in the effective population size due to background selection and S(x) is the local coalescence rate caused by classic sweeps. This approximation can be arrived at by considering the probability that a mutation occurs (at a rate 2u(x) per generation) before our pair of lineages are forced to coalesce by either genetic drift (1/2NeB(x)), which includes the effect of background selection, or by a selective sweep (S(x)). While we consider autosomes, the model can be extended to sex chromosomes with minor modifications.
The model for the effects of background selection, B(x), follows Hudson & Kaplan [8] and Nordborg et al. [9] (Fig 1A). We assume a set of distinct annotations iB = 1,…,IB under purifying selection (e.g., exons, UTRs, introns and intergenic regions) and positions in the genome AB = {aB(iB)|iB = 1,…,IB}, where aB(iB) denotes the set of genomic positions with annotation iB. The selection parameters at these annotations are given by ΘB = {(ud(iB),f(t|iB))|iB = 1,…,IB}, where ud is the rate of deleterious mutations and f(t) is the distribution of selection coefficients in heterozygotes. The reduction in the effective population size is then
B(x|AB,ΘB,R)=Exp(−∑iB∑y∈aB(iB)∫ud(iB)t(1+r(x,y)(1−t)/t)2f(t|iB)dt),
(2)
where R is the genetic map, r(x, y) is the genetic distance between the focal position x and positions y (only positions on the same chromosome are considered). The integral reflects the effect that a site under purifying selection at position y exerts on a neutral site at position x. This expression and its combination across sites should provide a good approximation to the effect of background selection so long as selection is sufficiently strong (i.e., when 2Net>>1).
In turn, the model for the effect of selective sweeps follows from an approximation used by Barton [52] and Gillespie [53], among others (Fig 1A). Similarly to the model for background selection, we assume a set of distinct annotations iS = 1,…,IS subject to sweeps, but here we know the specific positions at which substitutions have occurred, AS = {aS(iS)|iS = 1,…,Is}, with aS(iS) denoting the set of substitution positions with annotation iS. The selection parameters at these annotations are ΘS = {(α(iS),g(s|iS))|iS = 1,…,IS}, where α is the fraction of substitutions that are beneficial and g(s) is the distribution of their additive selection coefficients. For autosomes, the expected rate of coalescent per generations at position x due to sweeps is then approximated by
S(x|AS,ΘS,R,N¯e,T)=1T∑iSα(iS)∑y∈a(iS)∫Exp(−r(x,y)τ(s,N¯e))g(s|iS)ds,
(3)
where T is the length of the lineage (in generations) over which substitutions occurred, the positions of substitutions y are summed over the chromosome with the focal site, N¯e is the average effective population size and τ(s,Ne) is the expected time to fixation of a beneficial substitution with selection coefficient s and given an effective population size Ne. We use the diffusion approximation for the fixation time
τ(s,Ne)=2(ln(4Nes)+γ−(4Nes)−1)s,
(4)
where γ is the Euler constant (cf. [28]). This model relies on several simplifying assumptions and approximations. In particular, the term 1/T relies on an assumption of one substitution per site per lineage and neglects variation in the length of lineages across loci. In combining the effects over substitutions, we further assume that the timings of beneficial substitutions are independent and uniformly distributed along the lineage, and that they are infrequent enough such that we can ignore interference among them [54]. The exponent approximates the probability of coalescence of two samples due to a classic sweep with additive selection coefficient s (where 2Nes>>1) in a panmictic population of constant effective size N¯e. (We consider the effects under more general sweep models later.) In principle, we should use the local Ne incorporating the effects of background selection but given the logarithmic dependence of Eq (3) on Ne, we simply use the average.
To infer the selection parameters ΘB and ΘS, we use a composite likelihood approach across sites and samples [55] (Fig 1B). We denote the positions of neutral sites by X and the set of samples by I. We then summarize the observations by a set of indicator variables across sites and all pairs of samples O = {Oi,j(x) | x ∈ X, i ≠ j ∈ I}, where Oi,j(x) = 1 indicates that samples i and j (i≠j) differ at position x and Oi,j(x) = 0 indicates that they are the same. In these terms the composite log-likelihood takes the form
LogL=∑x∈X∑i≠j∈Ilog(Pr{Oi,j(x)|ΘB,ΘS}),
where
Pr{Oi,j(x)|ΘB,ΘS}={π(x|ΘB,ΘS)Oi,j(x)=11−π(x|ΘB,ΘS)Oi,j(x)=0.
(5)
Using composite likelihood circumvents the complications of considering linkage disequilibrium (LD) and the more complicated forms of coalescent models with larger sample sizes. Importantly, maximizing this composite likelihood should yield unbiased point estimates [56,57]. Beyond losing the information in LD patterns and in the site frequency spectrum, the main cost of this approach is the difficulty in assessing uncertainty in parameter estimates (as standard asymptotics do not apply). We therefore use other ways to assess the reliability of our inferences.
To make the composite likelihood calculations (i.e., the calculation of π(x|ΘB,ΘS)) feasible genome-wide, we discretize the distribution of selection coefficients on a fixed grid. Given a grid of negative and positive selection coefficients, tg and sk, g = 1,…,G and k = 1,…,K, the distribution of selection coefficients for each annotation becomes a set of weights on this grid, w(tg| iB) and w(sk| iS). (In principle, the grid could also be annotation-specific.) For background selection, these weights reflect the rate of deleterious mutations with a given selection coefficient and their sum should therefore be bound by the maximal deleterious mutation rate per site. For sweeps, the weights reflect the fraction of beneficial substitutions with a given selection coefficient and their sum should be bound by 1. In these terms, the effect of background selection takes the form
B(x|ΘB)=Exp(−∑iB∑g=1Gw(tg|iB)b(x|tg,iB)),
(6)
where Exp(−b(x|tg, iB)) is the proportional reduction in the effective population size induced by having one deleterious mutation per generation per site with selection coefficient tg at all the positions in annotation iB. By the same token, the effects of sweeps take the form
S(x|ΘS)=1T∑iS∑k=1Kw(sk|iS)s(x|sk,iS),
(7)
where 1Ts(x|sk,iS) is the probability of coalescence per generation induced by sweeps in annotation iS, if all the substitutions in this annotation are beneficial with selection coefficient sk. Thus, by using a grid, we can calculate a lookup table of b(x|tg, iB) and s(x|sk, iS) once and then use it to calculate the likelihood for a given set of weights. Moreover, the interpretation of estimated distributions on a grid is arguably simpler than that of the continuous parametric distributions commonly used (e.g., gamma and exponential), for which densities associated with different selection coefficients are highly interdependent. In the Supplementary Material (S1B Text), we describe additional simplifications in the calculation of b(x|tg, iB) and s(x|sk, iS).
Other parameters are estimated as follows. Consider Eq (1) rewritten as
π(x)=π0⋅(u(x)/u¯)π0⋅(u(x)/u¯)+1/B(x)+S(x;N¯e,T),
(8)
to clearly specify all the additional parameters required for inference. π0≡4Neu¯ is (approximately) the average neutral heterozygosity, given the effective population size in the absence of linked selection and the average mutation rate per site (u¯); π0 is estimated through the likelihood maximization. The local variation in mutation rate u(x)/u¯ is estimated by averaging substitution patterns at putatively neutral sites among closely related species in sliding windows, with a window size chosen to balance true variation in mutation rates and measurement error (see S1B Text). Finally, N¯e is estimated based on the average genome-wide heterozygosity at putatively neutral sites, after dividing out by a direct estimate of the spontaneous mutation rate per site, and T/2N¯e is estimated by (K¯/2)/π0, where K¯ is the average number of substitutions per neutral site on the lineage.
The software package implementing the inference and construction of the map of the effects of linked selection is available online (http://github.com/sellalab/LinkedSelectionMaps). In the Supplementary Material (S1B Text), we describe the steps that were taken to check the proper convergence of the likelihood maximization.
We apply our method to population resequencing data from Drosophila melanogaster. The data analyses are briefly described here, with further details provided in S1A Text. As a proxy for neutral variation, we use synonymous polymorphism within D. melanogaster, based on resequencing data from the Drosophila Genetic Reference Panel (DGRP) [51] consisting of 162 inbred lines derived from the Raleigh, North Carolina population. The rate of synonymous divergence used to control for local variation in mutation rates is estimated using the aligned reference genomes of D. simulans and D. yakuba [58]. As potential targets of selection (annotations), we use coding regions, untranslated, transcribed regions (UTRs), long introns (>80bp) and intergenic regions, downloaded from FlyBase [59] (http://flybase.org, release 5.33), all of which have been inferred to be under extensive purifying selection in D. melanogaster [60–63], and which together cover ~98.5% of the euchromatic genome. Substitutions that occurred in these annotations on the D. melanogaster lineage since the common ancestor with D. simulans are inferred from a three-species alignment of reference genomes from D. melanogaster, D. simulans and D. yakuba [58]. We do not include substitutions in intergenic regions, which are not included in the three-species alignment, and our treatment of missing data, e.g., due to gaps in the alignment, is detailed in S1B Text.
For the genetic map, we rely on estimates of the cM/Mb rates recently published by Comeron et al. [64]. Because our inferences are sensitive to errors in the genetic map in regions of low recombination, we exclude the distal 5% of chromosome arms (in which rates are known to be low in D. melanogaster) and regions with a sex-averaged recombination rate below 0.75cM/Mb.
We perform the inference under a variety of selection models. In the Results, we primarily compare the models incorporating classic sweeps, background selection or both, including all of the annotations listed above using a grid of selection coefficients which consists of five point masses on a log-linear scale, with t and s = 10−5.5, 10−4.5, 10−3.5, 10−2.5 and 10−1.5. Our maps of the effects of linked selection corresponding to the model incorporating both classic sweeps and background selection are available online (http://github.com/sellalab/LinkedSelectionMaps/melanogaster_maps). In the Supplementary Material we study the sensitivity of our results to: selection on synonymous mutations—using a subsets of synonymous differences (S1H Text), the recombination thresholds (S1H Text), the grid of selection coefficients (S1I Text), and to using subsets of annotations (S1I Text) and an upper bound on the deleterious mutation rate (S1E Text).
Our inference yields a map of the expected neutral diversity levels at every position along the genome. One way to evaluate these predictions is to compare them with observed diversity levels (Fig 2). A quantitative comparison at the 1Mb scale suggests that our map accounts for 71% of the variance (R2) in diversity levels of non-overlapping autosomal windows. To address the concern that the high R2 is the result of over-fitting, we perform a leave-one-out cross-validation (LOOCV) analyses [65] in which we divide the genome into non-overlapping 1Mb windows, using only data outside a window to make our predictions about diversity levels in it (S1C Text; Table S2 in S1 Text). This analysis shows that over-fitting has a negligible effect on our prediction, which is to be expected: while our model has many parameters (36), the data set is much larger (consisting of 1.7×106 codons, and levels of linkage disequilibrium are low).
In interpreting the fit, both model misspecification and the stochasticity inherent to the evolutionary process need to be considered. Importantly, even if our model provided an accurate description of the processes generating genetic diversity, we would not expect a perfect fit to the data because of the randomness of the processes being modeled. Notably, our model assumes that a substitution at a given annotation could have occurred with uniform probability at any time along the D. melanogaster lineage and that it had a certain probability of being beneficial with a given selection coefficient. Any evolutionary realization of the model would have that substitution occur at a particular time—more often than not, too far in the past to affect extant diversity patterns—and with a given selection coefficient, thus generating considerable variance in predicted diversity levels at linked sites. In addition, both genealogical and mutational processes are stochastic. Averaging over 1Mb windows partially reduces this stochasticity and in that regard, it is not surprising that our predictions become less precise when we use smaller windows (Fig 2B). However, even with 1Mb windows, we would still expect considerable variance in diversity levels around the expectation.
In addition, although we assume that the genetic maps and annotations are known, there is error in both. Imprecision of the genetic map and imperfect annotations (e.g., our clumping together of all coding, UTR, intronic and intergenic substitutions and regions) decrease our predictive ability. As genetic maps and annotations become better, we should therefore expect our predictions to improve. Another class of assumptions relates to processes that we did not model, including changes in population size [61,66,67]. In spite of many potential factors contributing to noise in our predictions, the fit to data is very good.
In the Supplementary Materials (S1F Text) we compare our predictions to those based on a map of the effects of background selection generated using the methodology developed by Charlesworth [41] and recently extended by Charlesworth [42] and Comeron [43]. This approach differs from ours in several ways, most notably in being based on estimates of selection parameters from the literature, which themselves do not rely on the effects of linked selection on diversity patterns. While it performs impressively well at the 1Mb scale (though not as well as ours) the quality of the predictions becomes much worse than ours as the scale becomes smaller (Table S5 in S1 Text). (Note that Comeron [43] uses rank correlations to evaluate his predictions; the explained variance using rank correlations are much higher than the quantitative predictions we use here, which is why his result might appear comparable at first sight.)
Using R2 values for window sizes varying from 1kb to 1Mb, we can ask which model(s) are best supported. We find that the one combining both background selection and classic sweeps almost always does better than the models with a single mode of selection (Fig 2). Our leave-one-out cross-validation analysis confirms that this finding is not the result of over-fitting in the combined model (Table S2 in S1 Text; see S1C Text for details). Thus, our combined model of the effects of linked selection captures much of the variation in diversity levels at the mega-base scale, and provides an improved null model in scans for targets of positive selection or for the purposes of demographic inference. Because using R2 has its limitations, we use a variety of other statistical approaches to evaluate our inferences in the sections that follow.
We can also use our analysis to learn about the effects of linked selection for different annotations. If a feature is enriched for targets of purifying or positive selection, then we expect to see a reduction in diversity levels around it due to linked selection. Collating diversity levels around all instances of a feature averages over confounding effects at specific genomic positions as well as over the inherent stochasticity in diversity levels, allowing us to isolate the selection effects [18,48–50].
We first consider how diversity levels vary with genetic distance from amino acid and synonymous substitutions (Fig 3). There is a trough in diversity around both, but the one around amino acid substitutions is substantially deeper (Fig 3A). Fig 3B compares the predicted diversity levels around amino acid substitutions based on Sattath et al. [48] and our inference. A rough quantitative comparison suggests that our method fits the data better than that of Sattath et al. (R2 = 62% for our method compared to R2 = 56% for Sattath et al.; see S1G Text for more details). Moreover, the new method also predicts more of the detailed variation in diversity levels, presumably because it accounts for the statistical properties of genome architecture, e.g., the density of coding regions at given genetic distances up or downstream of substitutions.
In principle, our approach should allow us to tease apart the contributions of classic sweeps and background selection to these diversity patterns (Fig 4). Comparing the predictions of each model alone is less informative for this purpose, because when only one is considered, it likely absorbs some of the effects of the other (see next section). In contrast, with the inference based on the combined model, the contribution of each mode should be identifiable from its specific functional forms and annotations. When we focus on the contribution of background selection (blue lines in Fig 4B), we see a reduction in diversity around both synonymous and non-synonymous substitutions because both types of substitutions occur in coding regions, in which background selection effects are strongest (e.g., [18,68]). Moreover, because the density of coding regions and other annotations (blue lines in Fig 4C and Fig S6 in S1 Text) is similar around the two kinds of substitutions, the shape and magnitude of the reductions in diversity are also similar (blue lines in Fig 4B). In contrast to background selection, the reduction around non-synonymous substitutions due to classic sweeps is much greater than for synonymous substitutions (red lines in Fig 4B). This results not only from the focal non-synonymous substitution but also (and primarily) from the greater density of non-synonymous substitutions near a focal non-synonymous substitution than around a synonymous one (red lines in Fig 4C). Whereas the clustering of non-synonymous substitutions around synonymous substitutions primarily reflects the greater density of coding sites, the clustering around non-synonymous substitutions (beyond the focal amino acid substitution) presumably reflects correlated evolution of nearby residues and other adaptive processes (e.g., [69]).
These findings illustrate that, at least as modeled, background selection and classic sweeps are identifiable. Intuitively, the information about classic sweeps at non-synonymous substitutions comes from the comparison of neutral diversity levels between sites near many non-synonymous substitutions versus near few, given a similar density of other annotations. After properly accounting for the effects of classic sweeps, information about the background selection pressure exerted by exons comes from contrasting the diversity levels among regions that vary in the density of codons but are otherwise similar. In practice, we do not learn about these processes in a stepwise fashion, as presented here, but instead maximize the probability of the data considering all of the annotations simultaneously.
We can therefore use these findings to revisit the enduring question of the relative contribution of background selection and classic sweeps to shaping diversity patterns (Fig 4). In particular, the negative correlation between diversity levels and the density of non-synonymous substitutions previously reported in Drosophila [11,12] likely reflects a substantial contribution of background selection in addition to positive selection. In contrast, the greater reduction in diversity levels at non-synonymous compared to synonymous substitutions in Drosophila is almost entirely the outcome of classic sweeps [48]. A caveat is that the parameter estimates obtained from the approach based on collated plots likely absorb some of the effects of background selection and thus overestimate the effects of linked selection due to sweeps (see next section and Tables S6 and S7 in S1 Text). More generally, in interpreting the results, an important consideration is the presence of other modes of selection that are not modeled explicitly, e.g., soft and partial sweeps. As we discuss at greater length below, our inferences about classic sweeps may reflect a mixture of different kinds of sweeps that result in substitutions while our inferences about background selection may reflect a contribution from other modes of linked selection, including sweeps that do not result in substitutions.
We can also consider how well the relationships between diversity levels and various genomic features are explained by models with a single mode of selection. As an illustration, Fig 5A shows that the background selection model does better than the model with classic sweeps at predicting diversity levels far from non-synonymous substitutions. Also visually apparent is that, in contrast to the background selection model, the classic sweeps model explains the narrow, deep trough close to non-synonymous substitutions. The combined model does well at predicting diversity levels both close to and far from non-synonymous and synonymous substitutions, again illustrating the need to consider both modes of linked selection in making inferences.
A similar approach can be used to examine the effects of selection acting on non-coding annotations. Notably, our inference suggests that a substantial fraction of substitutions at UTRs lead to classic sweeps (Table S11A in S1 Text and next section). To examine whether this feature of the model is required to explain the data, we look at average diversity levels as a function of genetic distance from substitutions in UTRs (Fig 5B). Our full model does much better at explaining these observations than a model without sweeps at UTRs. This provides the first evidence, to our knowledge, for sweeps at UTRs (or in any non-coding annotation) in Drosophila, and lends strong support to findings of pervasive adaptation in UTRs based on McDonald-Kreitman type approaches and genetic differentiation (FST) along clines [60,70].
Our approach also provides estimates of selection parameters. We first consider those obtained for classic sweeps, for which the positions of potential targets of selection (i.e. substitutions) are known. For substitutions at non-synonymous sites and to a lesser extent in UTRs, the ability to localize substitutions and to measure diversity levels using nearby synonymous sites provides us with high spatial resolution about selection effects on diversity patterns.
If we exclude background selection from the model, the only notable difference is the addition of a probability mass of strong selection coefficients at amino acid substitutions (~0.3% of substitution with s = 10−1.5), which affects diversity levels on a broad scale, in effect retracing large-scale variation in recombination rate and, to a lesser extent, coding density. When background selection is included in the model, this spatial effect becomes entirely associated with background selection (Fig S3 in S1 Text). This suggests that under a model of sweeps alone, the extra mass is absorbing some of the effects of other modes of selection that are not driven by substitutions.
In turn, under our combined models, the distribution of selection coefficient exhibits two dominant masses: ~4% of substitutions appear to have been strongly selected (s≈10−3.5) and 35–45% of substitutions weakly so (s between 10−5.5–10−6; the ranges reported here and below correspond to grids of selection coefficients with 5 and 11 point masses; see S1I Text). The effects of both masses on diversity levels can be clearly seen in collated plots around substitutions (cf. Fig S8 in S1 Text) and accord with previous studies [48,71]. At UTRs, we find that 25–45% of substitutions are associated with weak to intermediate strength of selection (s≈10−4.5–10−5.5). While the effects of sweeps at UTRs are apparent in Fig 5B, our quantitative estimates are associated with greater uncertainty than those for non-synonymous substitutions because we have lower spatial resolution near substitutions at UTRs (see S1H Text). At long introns, we infer that none of the substitutions were driven by sweeps; this estimate, however, might also reflect low power in these regions, because we measure diversity levels at synonymous sites that are, on average, far from intronic substitutions (see S1I Text).
Intriguingly, our estimates of the fraction of beneficial substitutions in proteins and UTRs accord with those based on extensions of the McDonald-Kreitman test (i.e., between ~40–85% for amino acids and 30–60% from UTRs [12,38,60,72–74]), when previous estimates based on the effects of sweeps on polymorphism data were substantially lower [11,48]. A caveat is that this conclusion only holds when we include the contribution of weakly selected substitutions. Our inference about weakly selected substitutions is based on diversity patterns very close to substitutions (roughly equivalent to 50 bp on average) and at these distances, considerable uncertainty about the genetic map and limited polymorphism data preclude us from distinguishing between selection coefficients ranging between 10-5.5 and 10−6. Because selection coefficients at the lower end of this range could be nearly neutral, the substitutions could partially reflect the fixation of slightly deleterious mutations rather than beneficial ones and more generally compensatory evolution [75]. We note further that our approach is not necessarily expected to agree with McDonald-Kreitman based estimates, which reflect adaptive rates over different time scales (i.e., on the order of Ne in our case [76], as opposed to the time scale of divergence). These reservations notwithstanding, our approach suggests that properly accounting for weakly selected substitutions leads to a convergence of estimates based on linked selection and McDonald-Kreitman based approaches, and provides, to our knowledge, the first corroboration of these elevated estimates.
With recent research highlighting the potential role of modes of adaptation other than classic sweeps, e.g., partial and soft sweeps [27–31,77–80], which we do not model explicitly, it is natural to ask how they might affect our inferences. To a first approximation, the effects of other kinds of sweeps on diversity levels around the selected site can be viewed as a superposition of the effects of classic sweeps with varying selection coefficients at different distances from the selected site (see [31,81] and S1D Text). This property implies that our parameter estimates for classic sweeps can be translated into rates and strengths of other types of sweeps.
As an example, consider our estimates that ~4% of amino acid substitutions were driven by selection coefficients of s = 10−3.5 and ~35% by a selection coefficient of 10-5.5. An approximately similar effect on diversity levels along the genome could be explained by assuming that 39% of substitutions are caused by partial sweeps that are driven to a frequency of x = 0.34 with a selection coefficient of s = 10−3.9, then to fixation with a selection coefficient of s = 10−5.8 (see S1D Text). Similar parameter estimates could also be generated by mixtures of partial and full sweeps, described by the fraction of full and partial sweeps and associated selection coefficients and distributions of frequencies (x) for each kind of partial sweep. In S1D Text, we detail how other kinds of sweeps (soft, from multiple mutations or standing variation, or on recessive alleles) would be recorded by our approach and thus how the effects of mixtures of sweeps would translate into our parameter estimates.
In other words, in the presence of different kinds of sweeps, our parameter estimates reflect the effects of the mixture on diversity levels around substitutions. A given set of estimates designates a continuous class of mixtures and, in principle, one can write down equations for the parametric family of mixtures that would yield the same estimates. Further narrowing down the underlying mixtures, however, will require developing inferences that use other aspects of the data.
Parameter estimates for purifying selection are fairly insensitive to the exclusion of classic sweeps from our model (e.g., Table S5 in S1 Text). When we do not impose an upper bound on the rate of deleterious mutations, we observe two main selection strengths, both of which are localized in exons and UTRs. The dominant one is extremely strong selection (s = 10-1.5), which affects diversity over a spatial scale of ~4Mb (or ~7cM, the distance at which the diversity levels reach 90% of baseline levels). As noted previously, such selection coefficients lead diversity levels to follow large-scale variation in recombination rate and to a lesser extent coding density. In this regard, it is important to note that we have to rely on relatively crude annotations, rather than accounting for the fine-scale location of sites under purifying selection within each annotation. As a result, our inference is likely to capture an average effect over considerably larger spatial scales than is actually the case, thereby leading to somewhat inflated selection coefficients (akin to what is seen for classic sweeps when background selection is not considered).
The strong selection coefficient is also associated with unreasonably high estimates of the deleterious mutation rate, which far exceed direct estimates of the total mutation rate (by 4-9-fold in exons and UTRs; Table S12 in S1 Text) [82]. A plausible interpretation is that these high rates reflect the absorption of linked selection effects that evade direct capture by our inference. For example, they might absorb the effects of sweeps at introns (or intergenic regions) that evade our inference because of the crude annotation of substitutions in these regions. They might also absorb the effects of other modes of linked selection, which are not modeled explicitly. Notably, population genetic models of quantitative traits suggest that the response to changing selection pressures could involve many soft and partial sweeps that do not result in fixation [83,84] and therefore would not be included in our estimates for classic sweeps. The effects of such soft and partial sweeps on diversity levels can be similar to those of background selection [31,81,85,86]. Moreover, because we lack localized annotations for such sweeps (when they do not result in fixation), we would tend to associate them with stronger selection coefficients of background selection, whose effects on diversity are less localized. If this interpretation is correct, then our inference suggests that modes of linked selection other than classic sweeps and background selection have a substantial effect on diversity levels around coding regions.
We also find evidence for somewhat weaker purifying selection (centered around s = 10−3.5) associated with a more realistic deleterious mutation rate (e.g. ~50–60% of the overall mutation rate in exons), but which may still reflect a contribution from other forms of linked selection. These values are in agreement with those obtained for exons by approaches that do not rely on the signatures of linked selection (cf. [42,43], and S1F Text). Purifying selection of this strength should affect diversity levels on spatial scale of ~40 kb (or 0.07cM, defined as above), a footprint that is visible in our analyses of diversity levels around synonymous and non-synonymous substitutions (blue lines in Fig 4B).
In the Supplementary Material (S1E and S1F Text), we present additional analyses that support this interpretation of background selection parameters, based on models in which we impose a biologically plausible upper bound on the deleterious mutation rate and use the modeling approach of Charlesworth [41,42].
We next examine the extent to which linked selection decreases the mean and increases the variance in diversity levels throughout the genome. The average reduction quantifies the effects of linked selection on the effective population size, a key parameter for many aspects of genome evolution [24,25]. The heterogeneity in diversity levels is of interest because it quantifies the deviation from the uniform neutral null model that is implicitly assumed in most, if not all, demographic inferences and scans for targets of adaptation.
We focus on the impact of linked selection in coding regions with recombination rates above 0.1cM/Mb, because our predictions become less reliable in regions with lower recombination rates (see S1H Text). To this end, we sort genomic positions according to their predicted levels of diversity (Fig 6A). For 1600 bins with equal amounts of data, the concordance between observed and predicted levels is extremely high (Spearman ρ = 0.91), indicating that the variation predicted by our model is real (and not due to over-fitting; Table S2 in S1 Text). Sorting based on our predictions, we find substantial variation in the observed diversity levels across bins (approximately five-fold difference between the upper and lower 2.5%; Fig 6B). Moreover, we see that the effects of linked selection are visible across all bins, rather than being restricted to bins with lower expected diversity. In other words, almost no region in the genome is free from the effects of linked selection (with the exception of the correlation coefficient, none of these results are sensitive to the number of bins).
We quantify the average reduction due to linked selection as the ratio of the average observed diversity level, π¯, to the predicted level without linked selection, π0. Doing so indicates an average reduction of 77%-89% in neutral diversity levels genome-wide (excluding low-recombination regions for which the reduction should be greater). Strikingly, even in the upper 1%-tile, linked selection is predicted to have reduced diversity levels by ~60–80%. Given the uncertainty about the parameter estimates associated with strong purifying selection (S1E Text and Table S4 in S1 Text), our inferences about π0 may not be robust, however. Indeed, imposing a plausible bound on the rate of deleterious mutations results in fits that are only marginally worse but dramatically affects our estimates of π0 (reducing it from 4.4 fold times the observed mean to 2.8-fold, with 5 point masses; S1E Text and Table S4 in S1 Text). In brief, this follows from the fact that strong selection affects diversity levels on broad spatial scales, leaving little signal of localization, and thus similar observed diversity levels can result from different combinations of deleterious mutation rates and π0 values. Unfortunately, we cannot observe π0 directly. What we can say, based on our stratification, is that linked selection reduces average diversity levels by at least two-fold (Fig 6A).
Our estimates suggest much stronger effects of linked selection than do previous methods. Notably, when we apply previous methods based on the relationship between diversity levels and rates of recombination or functional divergence [6,10–12,26] (see S1G Text for details), we infer an average reduction in diversity levels that lies between 34–36%, with no reduction in the upper 1%-tile of predicted diversity levels (Fig 6B and Table S12 in S1 Text). Comparing the stratification of diversity levels by the various methods (Fig 6A and 6B) indicates these previous methods do worse at predicting diversity levels, span a smaller range of diversity levels and under-estimate the effects of linked selection; specifically, their predictions of π0 are lower than the upper 1%-tile of observed diversity levels based on our stratification (Fig 6A and 6B). The reason is that by relying on a single genomic feature (e.g., recombination rate) and averaging over others (e.g., non-synonymous divergence), these methods overlook much of the variation in diversity levels caused by linked selection, causing their estimates to suffer from the equivalent of regression toward the mean (the same problem applies to their estimated selection parameters; see S1G Text). A similar “averaging out” effect takes place when we consider a model with background selection or classic sweeps alone (Fig 6A and 6B).
This line of argument implies that even with the combined model, we still underestimate the heterogeneity in diversity levels because of imperfect annotations. Notably, this would be the case if our inferences about background selection are likely absorbing substantial effects of other modes of linked selection but are unable to capture them in full, let alone to do so with high spatial resolution. Thus, the heterogeneity in diversity levels due to linked selection in the Drosophila melanogaster genome is likely to be even greater than we have inferred. Similar speculation about the average reduction in diversity levels is more difficult, given the uncertainty associated with our parameter estimates for background selection (Tables S4 and S10 in S1 Text). What we can say is that our lower bound based on stratification is likely to increase as annotations improve.
Over two decades of research have aimed to quantify the relative contributions of classic sweeps and background selection in shaping diversity patterns. If these were the only modes of linked selection, then we would now have an answer. We have shown that the contributions of background selection and classic sweeps are identifiable using our inference and, with the stated caveats about the effects of partial annotations, we can quantify their relative contributions. Based on the combined model and using the genome-wide average rates of coalescence induced by each mode of selection as a measure of their relative contribution, our findings would suggest that background selection has a ~1.6–2.5-fold greater effect than classic selective sweeps (Table S3 in S1 Text; see S1C Text for details and other metrics).
The question is complicated, however, by the contribution of other modes of linked selection. Our results strongly suggest that inferences about background selection include a major contribution of other modes of linked selection, plausibly the result of sweeps that do not result in substitutions. In turn, our inferences for classic sweeps may reflect a combination of different kinds of sweeps. These results echo other theoretical and empirical results highlighting the importance of other modes of positive selection, notably of partial and soft sweeps [27–31,77–81].
The question about the relative contribution of different modes of linked selection can therefore be rephrased in terms of the contributions of background selection, classic sweeps and other modes of linked selection. If we assume that our combined model fully accounts for the reduction in diversity levels due to linked selection and that the effects of background selection are captured by our inferences excluding the strong selected mass, then 12% of the increase in coalescence rate due to linked selection is the result of background selection (estimates in this paragraph correspond to the model with 5 point masses). Further assuming that our inferences about classic sweeps can reflect any combination of classic and other kinds of sweeps resulting in fixation, and that the remaining effects are the outcome of other modes of linked selection, then we would conclude that roughly 0 to 29% of coalescent events are due to classic sweeps and the remaining 88 to 59%, respectively, are due to other modes of linked selection.
Despite unresolved questions about linked selection, the maps do well at predicting diversity levels at the 1Mb scale (Fig 2), the substantial stratification of diversity levels throughout the genome (Fig 6) and the diversity patterns around different annotations (Figs 3, 4 and 5). This predictive ability is explained in part by the effects of linked selection already well captured by our current approach, e.g., the effects of sweeps that result in substitutions. Also important, however, is the robustness of the inferred map of linked selection to model misspecification. For instance, our map performs well even though the effects of background selection may reflect a substantial contribution of other modes of linked selection and despite an averaging effect owing to the imprecise annotations. Moreover, at this scale, the performance is fairly insensitive to variations of the model (e.g., imposing a bound on the deleterious mutation rate), suggesting that these features play a relatively minor role. Thus, while the spatial resolution of maps of linked selection in Drosophila (and other taxa) is expected to improve with better genetic maps, annotations and models, we can already do quite well. One implication is that our approach already generates substantially improved null models for population genetic inferences about demography and scans of selection.
The reliability of our inferences about selection critically depends on well-localized annotations and on the distance between these annotations and the putatively neutral sites used to measure diversity levels. For these reasons, we obtain reliable estimates for sweeps resulting in substitutions at exons and UTRs and distinguish their contribution from other forms of linked selection, but cannot achieve similarly reliable estimates for other modes and annotations. It follows that in applications to other species, we would expect the reliability of estimates to depend both on the quality of annotations and on genome architecture. Human data may be particularly well suited, as there are higher-resolution annotations as well as phylogeny-based information about conservation in both coding and non-coding regions. In addition, properties of the genome architecture, notably the lower density of selected regions [87], may help to distinguish effects of different annotations and modes of linked selection.
In both Drosophila and humans, one area that will need further work is the inclusion of other modes of selection. In that regard, it is interesting to note that our results mirror similar finding in humans: inferences about background selection in McVicker et al. [18] also led to too large a rate of deleterious mutation and work done since suggests that classic sweeps contribute little to the effects of linked selection on genetic variation [49,77,78]. Taken together with other empirical evidence and modeling [27–30,77,79,80,83], these results strongly suggest that other modes of linked selection and of adaptation in particular play a central role in both Drosophila and humans.
It might be difficult to distinguish between different kinds of sweeps based on their footprints around substitutions, especially given the many additional parameters for each if they act in concert (S1D Text). Additional footprints of selection are likely to be needed. Notably, there is likely to be important information about alternative modes of sweeps in diversity levels and patterns of linkage disequilibrium around amino acid polymorphisms [22,80,88].
Another pertinent extension will be to incorporate more realistic demographic assumptions. Like many other methods aimed at quantifying the genome-wide effects of linked selection to date [10,12,18], our model implicitly assumes a panmictic population of constant size. While we focus on a single population, and hence our assumption of random-mating is appropriate, our assumption of a constant size is likely invalid [66,67,89,90]. However, our inference method should be fairly insensitive to changes in the population size, because demographic history should affect different genomic regions similarly, regardless of annotations or other aspects of genomic architecture. Since our method learns about modes of selection and their parameters by contrasting diversity patterns among regions with different properties, it should implicitly control for much of the effects of demography. Having said that, drastic changes in population size could change the efficacy of selection and thus influence our estimates of the distribution of selection coefficients. In addition, regions with different effective population sizes due to linked selection could differ in their transient responses to demographic changes, potentially affecting our inferences. Accounting for these effects is difficult, however. Moreover, existing demographic inferences for North American D. melanogaster are confounded by the pervasive effects of linked selection. The methods developed here offer a way forward in inferring demography in the presence of linked selection as our map of linked selection could be factored into such analyses.
While these extensions will be important, our current application to Drosophila already reveals that the effects of linked selection are greater than previously assumed, by taking into account spatial features of genome architecture that were previously averaged out. Even excluding low recombination regions, our results suggest high heterogeneity in expected diversity levels due to linked selection (Fig 6B) and an overall reduction in diversity levels of at least two-fold. Applying our approach to other taxa will reveal whether linked selection is having a similarly large effect in other species, and is an important contributor to the apparent disconnect between census and effective population sizes [2,23–26].
|
10.1371/journal.pgen.1006122 | A Powerful Procedure for Pathway-Based Meta-analysis Using Summary Statistics Identifies 43 Pathways Associated with Type II Diabetes in European Populations | Meta-analysis of multiple genome-wide association studies (GWAS) has become an effective approach for detecting single nucleotide polymorphism (SNP) associations with complex traits. However, it is difficult to integrate the readily accessible SNP-level summary statistics from a meta-analysis into more powerful multi-marker testing procedures, which generally require individual-level genetic data. We developed a general procedure called Summary based Adaptive Rank Truncated Product (sARTP) for conducting gene and pathway meta-analysis that uses only SNP-level summary statistics in combination with genotype correlation estimated from a panel of individual-level genetic data. We demonstrated the validity and power advantage of sARTP through empirical and simulated data. We conducted a comprehensive pathway-based meta-analysis with sARTP on type 2 diabetes (T2D) by integrating SNP-level summary statistics from two large studies consisting of 19,809 T2D cases and 111,181 controls with European ancestry. Among 4,713 candidate pathways from which genes in neighborhoods of 170 GWAS established T2D loci were excluded, we detected 43 T2D globally significant pathways (with Bonferroni corrected p-values < 0.05), which included the insulin signaling pathway and T2D pathway defined by KEGG, as well as the pathways defined according to specific gene expression patterns on pancreatic adenocarcinoma, hepatocellular carcinoma, and bladder carcinoma. Using summary data from 8 eastern Asian T2D GWAS with 6,952 cases and 11,865 controls, we showed 7 out of the 43 pathways identified in European populations remained to be significant in eastern Asians at the false discovery rate of 0.1. We created an R package and a web-based tool for sARTP with the capability to analyze pathways with thousands of genes and tens of thousands of SNPs.
| As GWAS continue to grow in sample size, it is evident that these studies need to be utilized more effectively for detecting individual susceptibility variants, and more importantly, to provide insight into global genetic architecture of complex traits. Towards this goal, identifying association with respect to a collection of variants in biological pathways can be particularly insightful for understanding how networks of genes might be affecting pathophysiology of diseases. Here we present a new pathway analysis procedure that can be conducted using summary-level association statistics, which have become the main vehicle for performing meta-analysis of individual genetic variants across studies in large consortia. Through simulation studies we showed the proposed method was more powerful than the existing state-of-art method. We carried out a comprehensive pathway analysis of 4,713 candidate pathways on their association with T2D using two large studies with European ancestry and identified 43 T2D-associated pathways. Further examinations of those 43 pathways in 8 Asian studies showed that some pathways were trans-ethnically associated with T2D. This analysis clearly highlights novel T2D-associated pathways beyond what has been known from single-variant association analysis reported from largest GWAS to date. We also identify a novel locus for T2D in the European populations at chromosome 17q21 (rs1058018, p = 3.06 × 10−8).
| Genome-wide association study (GWAS) has become a very effective way to identify common genetic variants underlying various complex traits [1]. The most commonly used approach to analyze GWAS data is the single-locus test, which evaluates one single nucleotide polymorphism (SNP) at a time. Despite the enormous success of the single-locus analysis in GWAS, proportions of genetic heritability explained by already identified variants for most complex traits still remain small [2]. It is increasingly recognized that the multi-locus test, such as gene-based analysis and pathway (or gene-set) analysis, can be potentially more powerful than the single-locus analysis, and shed new light on the genetic architecture of complex traits [3, 4].
The pathway analysis jointly tests the association between an outcome and SNPs within a set of genes compiled in a pathway according to existing biological knowledge [4]. Although the marginal effect of a single SNP might be too weak to be detectable by the single-locus test, accumulated association evidence from all signal-bearing SNPs within a pathway could be strong enough to be picked up by the pathway analysis if this pathway is enriched with outcome-associated SNPs. Various pathway analysis procedures have been proposed in the literature, with the assumption that researchers could have full access to individual-level genotype data [5–9]. In practice, pathway analysis usually utilizes data from a single resource with limited sample size, as it can be challenging to obtain and manage individual-level GWAS data from multiple resources. As a result, pathway analysis often fails to identify new findings beyond what have already been discovered by the single-locus tests. To maximize the chance of discovering novel outcome-associated variants by increasing sample size, a number of consortia have been formed to conduct single-locus meta-analysis on data across multiple GWAS [10–14]. The single-locus meta-analysis aggregates easily accessible SNP-level summary statistics from multiple studies. Similarly, the pathway-based meta-analysis [15–21] that integrates the same type of summary data across participating studies could provide us a greater opportunity for detecting novel pathway associations. Future association studies focusing on identified pathways would have a much-reduced multiple-comparison burden in searching for novel variants with main or complicated nonlinear joint effects on the outcome of interest.
In this paper, we developed a pathway-based meta-analysis procedure by extending the adaptive rank truncated product (ARTP) pathway analysis procedure [9], which was originally developed for analyzing individual-level genotype data. The new procedure, called Summary based ARTP (sARTP), accepts input from SNP-level summary statistics, with their correlations estimated from a panel of reference samples with individual-level genotype data, such as the ones from the 1000 Genomes Project [22, 23]. This idea was initially used in conducting gene-based meta-analysis [24, 25] or conditional test [26]. As will be shown in the Results Section, sARTP usually has a power advantage over its competitors. In addition, sARTP is specifically designed for conducting pathway-based meta-analysis using SNP-level summary statistics from multiple studies. In real applications (e.g., the type 2 diabetes example described below), it is very common that different studies could have genotypes measured or imputed on different sets of SNPs. As a result, the sample size used in the pathway-based meta-analysis on each SNP can be quite different. Ignoring the difference in sample sizes across SNPs in a pathway-based meta-analysis would generate biased testing results.
Pathway analysis generally targets two types of null hypotheses [4], including the competitive null hypothesis [15, 16, 18–20], i.e., the genes in a pathway of interest are no more associated with the outcome than any other genes outside this pathway, and the self-contained null hypothesis [17, 21], i.e., none of the genes in a pathway of interest is associated with the outcome. The sARTP procedure focuses on the self-contained null hypothesis, as our main goal is to identify outcome-associated genes or loci. Also, as pointed out by [27], tests for the competitive null hypothesis often assume that genotype measured at different genes are independent when evaluating the association significance level. This assumption, which is generally invalid in practice, is unnecessary for sARTP when testing the self-contained null hypothesis. One may refer to [27] and [4] for more discussion and comparison of these two types of hypotheses.
The pathways defined in many public databases can consist of thousands of genes and tens of thousands of SNPs. To make the procedure applicable to large pathways, or pathways with high statistical significance, we implement sARTP with efficient and parallelizable algorithms, and adopt the direct simulation approach (DSA) [28] to evaluate the significance of the pathway association.
We demonstrated the validity and power advantage of sARTP through simulated and empirical data. We applied sARTP to conduct a pathway-based meta-analysis on the association between type 2 diabetes (T2D) and 4,713 candidate pathways defined in the Molecular Signatures Database (MSigDB) v5.0. The analysis used SNP-level summary statistics from two sources with European ancestry. One is generated from the Diabetes Genetics Replication and Meta-analysis (DIAGRAM) consortium [13], which consists of 12,171 T2D cases and 56,862 controls across 12 GWAS. The other one is based on a T2D GWAS with 7,638 T2D cases and 54,319 controls that were extracted from the Genetic Epidemiology Research on Aging (GERA) study [29, 30]. The novel T2D-associated pathways detected in the European population were further examined in Asians using summary data generated by the Asian Genetic Epidemiology Network (AGEN) consortium meta-analysis, which combined 8 GWAS of T2D with a total of 6,952 and 11,865 controls from eastern Asian populations [10].
Here we describe the proposed method sARTP for assessing the association between a dichotomous outcome and a pre-defined pathway consisting of J genes. The same procedure can be applied to study a quantitative outcome with minor modifications.
Firstly, we conducted a simulation study to evaluate the empirical size of sARTP and MsARTP. Secondly, we compared empirical powers of different strategies for carrying out pathway-based meta-analysis that integrated summary statistics from multiple studies. We also evaluated whether results from sARTP were consistent with the ones from MsARTP. Thirdly, we compared our method to the recently developed method aSPUsPath [8] that can be used for pathway-based meta-analysis. We used the R package, aSPU (version 1.39), with the default settings given in [8, 17] to conduct the aSPUsPath test.
To demonstrate the consistency between results obtained by sARTP using SNP-level summary statistics and the ones by ARTP using individual-level genotype data, we compared pathway analysis results from three different procedures on the 4,713 candidate pathways using the GERA GWAS data. Details on how those 4,713 pathways were pre-processed are given in the Results of T2D Pathway Analysis Section. We applied sARTP to the SNP-level summary statistics generated from the GERA study, using either an internal or an external reference panel. We also obtained the pathway p-values by directly applying the ARTP method to the individual-level GERA GWAS data. Fig 1 shows the comparison among p-values from these three analyses, and demonstrates that all three approaches can generate very consistent results.
We developed a general statistical procedure sARTP for pathway analysis using SNP-level summary statistics generated from multiple GWAS. By applying sARTP to summary statistics from two large studies with a total of 19,809 T2D cases and 111,181 controls with European ancestry, we were able to identify 43 globally significant T2D-associated pathways after excluding genes in neighborhoods of GWAS established T2D loci. Using summary data generated from 8 T2D GWAS with 6,952 cases and 11,865 controls from eastern Asian populations, we further showed that 7 out of 43 pathways identified in the European populations were also significant in the eastern Asian populations at the FDR of 0.1. The analysis clearly highlights novel T2D-associated genes and pathways beyond what has been known from single-SNP association analysis reported from largest GWAS to date. Since the new procedure requires only SNP-level summary statistics, it provides a flexible way for conducting pathway analysis, alleviating the burden of handling large volumes of individual-level GWAS data.
We have developed a computationally efficient R package called ARTP2 implementing the ARTP and sARTP procedures, so that it can be used for conducting pathway analysis based on individual-level genetic data, as well as SNP-level summary data from one or multiple GWAS. The R package also supports the parallelization on Unix-like OS, which can substantially accelerate the computation of small p-values when a large number of resampling steps are needed. The ARTP2 package has a user-friendly interface and provides a comprehensive set of data preprocessing procedures to ensure that all the input information (e.g., allele information of SNP-level summary statistics and genotype reference panel) can be processed coherently. To make the sARTP method accessible to a wider research community, we have also developed a web-based tool that allows investigators to conduct their pathway analyses using the computing resource at the National Cancer Institutes through simple on-line inputs of summary data.
Single-locus analysis of GWAS usually has its genomic control inflation factor larger than 1.0. Some proportion of the inflation can be attributed to various confounding biases, such as the one caused by population stratification, while the other part can be due to the real polygenic effect. In the pathway analysis it is important to minimize the confounding bias at the SNP-level summary statistic. Otherwise a small bias at the SNP level can be accumulated in the pathway analysis, and lead to an elevated false discovery rate. Here we try to remove the confounding bias by adjusting for the genomic control inflation factor observed at the GWAS study. This approach is conservative because part of the inflation can be caused by the real polygenic effect. Recently, [42] developed the LD score regression method to quantify the level of inflation caused solely by the confounding bias. Adjusting for the inflation factor estimated by this method, instead of the genomic control inflation factor, can potentially increase the power of the pathway analysis. However, the LD score regression method relies on a specific polygenic risk model, and its estimate might not be robust for this model assumption. More investigations are needed to evaluate the impact of this new inflation adjustment on the pathway analysis.
There are several other strategies to increase the power of pathway analysis besides increasing sample size [4]. One area of active research is to find better ways to define the gene-level summary statistic using observed genotypes on multiple SNPs, so that it can accurately characterize the impact of the gene on the outcome [43–46]. In our proposed procedure, we adopt a data driven approach to select a subset of SNPs within a gene that collectively show the strongest association evidence. Because of this, we have to pay the penalty of multiple-comparison in the final pathway significance assessment. However, it is well recognized that SNPs at different loci can have varied levels of functional implications. We can potentially reduce the burden of multiple-comparisons and thus improve the power of the pathway analysis, by prioritizing SNPs according to existing genomic knowledge and other data resources. For example, [47] recently proposed a new gene-level summary statistic based on a prediction model that was trained with external transcriptome data. The gene-level summary statistic is defined as the predicted value that estimates the component of gene expression regulated by a subject’s genotypes within the neighborhood of the considered gene. Pathway analysis procedures using this kind of biologically informed gene-level summary statistic can be easily incorporated into the ARTP2 framework.
The sARTP method can be easily expanded to adopt other multi-locus statistics in accumulating association within a gene, as long as they can be written in terms of SNP-level score statistics and their variance-covariance matrix. For example, the current ARTP2 package provides the option for conducting the pathway meta-analysis using the joint test statistics proposed by [31].
When conducting pathway analysis with individual-level genetic data, we could run into a computing memory issue if the study has a large sample size and the pathway consists of a large number of genes and SNPs (S4 Fig). The ability of performing pathway analysis using summary data provides a convenient and efficient solution in those situations. We can first calculate the SNP-level summary statistics based on the individual-level genetic data, and then randomly sample a small proportion of the original data as an internal reference to estimate the variance-covariant matrix for score statistics at considered SNPs. Based on our experiments, using 500 or more subjects to form a reference panel would be good enough to generate accurate pathway p-values. As shown in Fig 1, the testing results using this approach are very consistent with those based on individual-level genotype data.
The sARTP approach can be applied directly to SNP-level meta-analysis results. This is very convenient as meta-analysis results are in general easily accessible. But we want to emphasize that it is important to know the set of the SNPs studied by each participating study in order to apply sARTP properly, as the SNP coverage information is essential for accurately estimating the variance-covariance matrix of SNP-level score statistics. GWAS consortia usually do not post the SNP coverage information when releasing their meta-analysis results. Many statistical packages designed for conducting multi-locus analysis based on meta-analysis results often assume the uniform coverage [15–18, 24, 25, 48]. As we already have demonstrated in the context of pathway analysis, this type of over-simplification could lead to inflated false positive rate.
The proposed procedure assumes that all participating studies are conducted with subjects with the same ancestry background. If this is not the case, a simple approach is to use the Fisher’s method to combine pathway p-values estimated on different ethnic populations. However, if there were no evidence for the existence of cross ethnic risk heterogeneity, it would be more powerful to assume a fixed effects model on the SNP-level association when performing the pathway analysis. In that case, since the LD structures in different ethnic populations are different, we need a separate reference panel for each ethic group to derive the corresponding variance-covariance matrix of the score statistics. The current ARTP2 package needs to be modified to accommodate such a more complicated case.
As already demonstrated by many successful GWAS meta-analysis, increasing the sample size through combining results from multiple studies is a very effective way to improve our chance for new findings. For the same reason, pathway-based meta-analysis can provide us with new opportunities to uncover biological pathways that are previously undetectable due to the limitation on the sample size. With more summary data from meta-analysis becoming increasingly available, we expect the ARTP2 package would be a valuable tool for further exploring the genome in search for the hidden heritability.
The URLs for data and software presented herein are as follows:
DIAbetes Genetics Replication And Meta-analysis (DIAGRAMv3), http://diagram-consortium.org/
Genetic Epidemiology Research on Aging (GERA, dbGaP Study Accession: phs000674.v1.p1), http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000674.v1.p1
Molecular Signatures Database (C2: curated gene sets), http://software.broadinstitute.org/gsea/msigdb/collections.jsp#C2
BioMart (Homo sapiens genes NCBI36 and GRCh37.p13), http://feb2014.archive.ensembl.org/
IMPUTE2, https://mathgen.stats.ox.ac.uk/impute/impute_v2.html
GWAS Catalog, http://www.ebi.ac.uk/gwas/
1000 Genomes Project (Phase 3, v5, 2013/05/02), ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/
aSPU, https://cran.r-project.org/web/packages/aSPU/index.html
GTEx Portal v6, http://gtexportal.org/home/
GeneCards Human Gene Database, http://www.genecards.org/
Ingenuity Pathway Analysis, http://www.ingenuity.com/
LocusZoom, http://locuszoom.sph.umich.edu/locuszoom/
ARTP2 package, https://cran.r-project.org/web/packages/ARTP2/
Web-based tool of ARTP2, http://analysistools.nci.nih.gov/pathway/
|
10.1371/journal.pbio.2005570 | The taste of ribonucleosides: Novel macronutrients essential for larval growth are sensed by Drosophila gustatory receptor proteins | Animals employ various types of taste receptors to identify and discriminate between different nutritious food chemicals. These macronutrients are thought to fall into 3 major groups: carbohydrates/sugars, proteins/amino acids, and fats. Here, we report that Drosophila larvae exhibit a novel appetitive feeding behavior towards ribose, ribonucleosides, and RNA. We identified members of the gustatory receptor (Gr) subfamily 28 (Gr28), expressed in both external and internal chemosensory neurons as molecular receptors necessary for cellular and appetitive behavioral responses to ribonucleosides and RNA. Specifically, behavioral preference assays show that larvae are strongly attracted to ribose- or RNA-containing agarose in a Gr28-dependent manner. Moreover, Ca2+ imaging experiments reveal that Gr28a-expressing taste neurons are activated by ribose, RNA and some ribonucleosides and that these responses can be conveyed to Gr43aGAL4 fructose-sensing neurons by expressing single members of the Gr28 gene family. Lastly, we establish a critical role in behavioral fitness for the Gr28 genes by showing that Gr28 mutant larvae exhibit low survival rates when challenged to find ribonucleosides in food. Together, our work identifies a novel taste modality dedicated to the detection of RNA and ribonucleosides, nutrients that are essential for survival during the accelerated growth phase of Drosophila larvae.
| Insects that undergo complete metamorphosis grow only during the larval stage of development. In many species, this period is restricted to a few days, during which larvae might increase their weight up to several hundred-fold. Drosophila melanogaster, for example, grow from a tiny first-instar larva of about 10 μg to a wandering third-instar larva weighing about 2 mg over a period of only 4.5 days. The main macronutrients known to be critical for this period of rapid growth are amino acids and sugars. In this study, we identify ribonucleosides and RNA as a new, additional type of nutrient necessary for rapid larval growth and survival. We show that larvae harbor taste neurons that express taste receptors necessary for sensing ribonucleosides and RNA. Larvae lacking these taste receptors show high mortality rates when exposed to a complex food environment that requires the location of ribonucleoside-containing food. We hypothesize that the ability to taste RNA evolved as a new taste modality in larvae of insects that go through a rapid growth period because ingestion of ribonucleosides, as opposed to de novo synthesis, provides a survival advantage during a period of extreme growth.
| Taste discrimination is a common trait of all animals, the most crucial being the ability to distinguish between palatable and mostly nutritional chemicals from aversively perceived, often harmful, and generally bitter-tasting compounds. Calorically nutritious food compounds fall into 3 categories, fats, proteins, and carbohydrates, and their consumption is dependent not only on availability but also on internal physiological states of the animal, such as overall nutrition status, anticipated need for energy expenditure, developmental stage, and reproductive status. To achieve discrimination between different nutrients, different subsets of cells in the taste sensory system express specific receptors for the detection of chemicals belonging to these nutrient groups [1].
Growth of most arthropods, but also some vertebrates, is characterized by a period of rapid body weight gain within a few days. A mouse increases its weight from time of birth (approximately 1 g) to the time of weaning (about 3 weeks; about 10 to 13 g) by about a factor of 10. Growth dynamics of many insect larvae are even more dramatic. A freshly hatched Drosophila larva weighs about 9.5 μg [2] but grows over a period of only 108 hours to more than 1.5 mg at the time of puparium formation [3], which translates to doubling of weight about every 14 hours. To support this rapid growth, animals are required to consume large amounts of all essential macronutrients, especially carbohydrates, proteins, and fats. In the natural diet of Drosophila, fruits provide the carbohydrates, while fatty acid and proteins are obtained mostly from colonizing microorganisms, such as yeast. Standard Drosophila food used in most laboratories is generally composed of cornmeal, malt extract, and yeast extracts (see Material and methods), which contain all these and other macro- and micronutrients in abundance.
In contrast to flies, which express and employ 8 sugar Gr genes for the perception of sweet taste, larvae use a single sugar receptor, Gr43a, for the detection of fructose and sucrose, the most abundant sugars in most fruits [4]. In addition, an Ionotropic Receptor (IR) was recently shown to be involved in sensing a limited number of amino acids [5]. Receptors for other macronutrients—including many amino acids, fats, and other compounds potentially critically important for rapid larval growth—have not been identified to date.
Here, we report the discovery of a novel taste modality—the taste of ribonucleosides and RNA—and the identification of cognate receptors. We show that larvae have a strong attraction to and feeding preference for ribose, ribonucleosides, and RNA. Feeding on these molecules is mediated by members of the highly conserved taste receptor subfamily Gr28. Live imaging experiments using the Ca2+ sensor Calcium Modulated Photoactivatable Ratiometric Integrator (CaMPARI) show that Gr28a-GAL4-expressing taste neurons respond to ribose, inosine, and uridine, as well as RNA itself, but not to 2-deoxyribose or to adenosine, guanosine, or cytidine. Holidic (synthetic) medium (HM) lacking inosine and uridine slows larval growth and causes high mortality, and supplementing this medium with RNA rescues both phenotypes. Moreover, when provided with a choice of HM with and without inosine and uridine, wild-type larvae readily select complete HM, leading to fast growth and high survival, while Gr28 mutant larvae fail to do so, resulting in slow growth and low survival. In summary, we have identified the cellular and molecular basis for the taste of ribonucleosides and RNA. We suggest that Drosophila larvae—and possibly other insect larvae—need to feed on RNA precursors to sustain the rapid increase in body weight, which is doubled almost twice a day.
While assessing sugar specificity of Gr43a, we made the surprising observation that larvae are also strongly attracted to arabinose (Fig 1). When given the choice between agarose containing L-arabinose and agarose alone, larvae showed a strong preference for arabinose (Fig 1A). In flies, arabinose is detected by receptors encoded by the sugar Gr subfamily [6, 7], none of which are expressed in larvae [8, 9]. Instead, larvae sense the main fruit sugars sucrose and fructose through Gr43a, which is expressed in pharyngeal taste neurons, as well as nutrient-sensing neurons in the brain [8]. Other nutritious sugars, such as trehalose and glucose, are sensed postprandially by Gr43a-expressing neurons in the brain, presumably after the conversion of a fraction of these sugars into fructose [8, 10]. To examine the possibility that arabinose is sensed either through Gr43a or any of the main sugar receptors used by adult flies, the low expression of which might have been missed in previous studies, we examined arabinose preference in Gr43a mutant larvae or larvae lacking all 8 sugar Gr genes (octuple mutant strain; Fig 1A [11]). However, neither Gr43a mutant nor octuple mutant larvae showed any significant loss in arabinose preference in the two-choice preference assay (Fig 1A).
L-arabinose is present as a minor component in heteropolysaccharides, such as hemicellulose and pectin [12, 13], but it is not a major sugar in fruit and cannot be metabolized by flies [14, 15]. To assess whether larvae can use arabinose as an energy source, we compared survival rates of second-instar larvae kept on agarose-containing arabinose to larvae kept on plain agarose or nutritious sugar-containing agarose (Fig 1B). Median (50%) survival for larvae on plain agarose was 3 days. Larvae kept on nutritious sugar–containing agarose survived significantly better, with more than 75% of larvae still alive after 3 days, indicating that the consumed sugar provided energy and decreased mortality. In contrast, only about 30% of larvae kept on arabinose-containing agarose survived to the 3-day time point. These observations suggest that arabinose cannot be a nutritionally relevant ligand, and we posit that larvae instead detect a molecule structurally related to arabinose, but one that is nutritious and essential for larval growth.
Arabinose is closely related to ribose, the carbohydrate backbone of RNA. We therefore investigated whether ribose and RNA can elicit a similar preference. Indeed, larvae strongly preferred ribose and RNA over plain agarose in the two-choice preference assay, whereas 2-deoxyribose, the sugar moiety of DNA, did not elicit a preference (Fig 1C). Moreover, neither Gr43a nor the sugar Gr genes were required to sense either ribose or RNA, as larvae with respective mutations showed robust preference for both substrates (Fig 1C). We next tested how ribose- or RNA-containing agarose affected larval survival (Fig 1B). In contrast to arabinose, neither of these compounds reduced survival time. However, they did not serve as an efficient energy source either because median survival time was not significantly different from larvae kept on plain agarose (Fig 1B). Taken together, these findings suggest that larvae can sense RNA through a taste or internal chemosensory receptor, presumably recognizing the ribose moiety in the RNA backbone.
We next sought to identify the bona fide receptor(s) that mediate taste preference for RNA and ribose. Most Gr genes expressed in larvae have not been functionally characterized during that life stage, but expression and function of many of them have been investigated in adults. The vast majority of these genes (i.e., encoding bitter taste receptors) are expressed in bitter taste neurons of taste bristles in the labial palps and legs of the fly, and it has been shown that bitter Gr proteins form multimeric receptor complexes that are activated by a vast array of chemicals perceived as repulsive [16–19]. In addition, Gr21a and Gr63a, which are also expressed in larvae, have noncanonical roles in a small subset of olfactory neurons in the fly, where they function as carbon dioxide receptors [20, 21]. Because bitter taste receptors and Gr21a/Gr63a are likely to have similar roles in larvae [9], we focused on the Gr28 gene clade (Gr28a, Gr28b.a, Gr28b.b, Gr28b.c, Gr28b.d, and Gr28b.e), members of which show broad expression in larvae as well as adult flies [22], whose roles in chemosensation have remained enigmatic. These receptors are also highly conserved across most arthropod orders [23–25], suggesting important functions that are shared across a range of species. Some of these Gr genes have been implicated in ultraviolet (UV) light sensing in larvae [26], as well as temperature sensing in flies [27], but no chemical ligands for any of these Gr28 proteins have been identified. To investigate a possible role for proteins encoded by the Gr28 gene family in RNA and ribose sensing, we first examined expression of respective GAL4 lines more closely in larvae, focusing on chemosensory organs and the digestive system, as well as the central nervous system (CNS) (Fig 2). Four of the six Gr28 genes were expressed in small numbers of external (Gr28a, Gr28b.a, Gr28b.e) and pharyngeal taste (Gr28a and Gr28b.d) neurons. Other sites of expression included neurons in the proventriculus (Gr28b.a, Gr28b.e), cells in the gut (all but Gr28b.d), multidendritic neurons in the larval body wall (Gr28a, Gr28b.c, Gr28b.d), and many neurons in the CNS (Gr28b.a, Gr28b.b, Gr28b.d, and Gr28b.d). Expression of Gr28-GAL4 drivers in the taste system did not overlap with Gr43aGAL4, which is expressed in a different set of pharyngeal neurons (S1 Fig).
We next investigated possible effects on preference for arabinose, ribose, and RNA across a range of concentrations in larvae lacking all Gr28 genes (ΔGr28; [28], Fig 3A and S2). Indeed, ΔGr28 mutant larvae completely lost preference for ribose and RNA and showed significantly reduced preference for arabinose, phenotypes that were rescued when reintroducing a genomic construct containing the entire Gr28 locus (Fig 3A). These observations indicate that one or several members of the Gr28 gene cluster mediate the detection of ribose and RNA.
To examine whether a single Gr protein can establish ribose sensing, we subjected ΔGr28 mutant larvae expressing each of the Gr28 genes under the control of Gr28a-GAL4. All Gr28 genes were able to restore at least some preference for ribose (Fig 3B). Notably, Gr28a, Gr28b.a, and Gr28b.e, all of which are expressed in taste neurons (Fig 2A), and Gr28b.b, rescued ribose preference to levels comparable to w1118 controls (Fig 3B). We then asked whether ribose sensing could be conveyed upon other larval taste neurons. We choose the well-characterized sugar-sensing, pharyngeal taste neurons expressing Gr43a, a receptor narrowly tuned to the sugars fructose and sucrose [8]. We examined two-choice preference behavior of ΔGr28 mutant larvae expressing each of the 6 Gr28 genes individually under the control of the Gr43aGAL4 driver (Fig 3C). Indeed, 3 genes—Gr28a, Gr28b.a, and Gr28b.e—endowed such larvae with the ability to sense and preferentially feed on ribose containing agarose in a manner indistinguishable from Gr28+ control larvae, and a fourth gene (Gr28b.d) mediated reduced ribose preference. The two other Gr28 genes, Gr28b.b and Gr28b.c, failed to convey any preference, just like Gr28 homozygous mutant larvae (Fig 3C). Of note, neither Gr28b.b nor Gr28b.c is expressed in taste neurons of wild-type larvae (Fig 2B), and while they rescued ribose sensing in Gr28a neurons, they failed to do so in heterologous Gr43a pharyngeal fructose-sensing neurons. One possibility is that Gr28 taste neurons express a cofactor (i.e., a chaperone or coreceptor) absent in Gr43a neurons and that some, but not all, Gr28 proteins are completely dependent on such a factor for taste receptor function. Taken together, our experiments established that larvae possess a taste modality for ribose and RNA and that individual Gr28 proteins are able to mediate ribose and RNA sensing.
To establish a role for the Gr28 proteins in ribose detection at the cellular level, we measured responses of terminal organ (TO) taste neurons using the fluorescent Ca2+ sensor CaMPARI [29]. In the presence of a high concentration of Ca2+ and simultaneous exposure to blue light, CaMPARI undergoes an irreversible conformational change, leading to a shift in emission properties from green to red fluorescence [29]. This feature allows application of ligand under free-moving conditions of the animal, while neural activation can be analyzed subsequently by quantifying green to red conversion ratios using fluorescent microscopy (see Material and methods). We presented Gr28a-GAL4; UAS-CaMPARI larvae with various ligands while illuminating them with blue light and found that ribose and RNA activated Gr28a-GAL4 neurons, while 2-deoxyribose failed to do so (Fig 4A). Consistent with behavioral assays of ΔGr28 mutant larvae, neurons lacking the Gr28 locus were not activated by either RNA or ribose, phenotypes that were rescued in the presence of a Gr28a transgene expressed in these neurons.
To further explore the role of the Gr28 proteins and to assess whether they can mediate ribose sensing to heterologous taste neurons, we co-expressed UAS-CaMPARI with Gr28 genes in fructose-sensing pharyngeal sweet taste neurons using Gr43aGAL4 and measured cellular responses to ribose and RNA (Fig 4B). We chose Gr28a, Gr28b.a, and Gr28b.e because they were competent to mediate ribose preference behaviorally when expressed in these neurons (Fig 3C). Indeed, Gr43aGAL4 fructose-sensing neurons now responded to ribose and RNA in the presence of any of the 3 Gr28 genes. The fructose response in these neurons was similar to that in neurons of control flies (i.e., not expressing a Gr28 gene), indicating that the function of the endogenous fructose receptor was not impaired. Response to RNA or ribose was lower than responses to fructose, however, suggesting that fructose is more potent in activating Gr43a than ribose or RNA is in activating the Gr28 proteins. Alternatively, activation of Gr28 proteins might be suboptimal in heterologous taste neurons due to the absence Gr28 neuron-specific cofactors or facilitators (see above). Regardless, these experiments, along with the two-choice feeding experiments (Fig 3C), show that expression of a single Gr28 protein can convey ribose- and RNA-sensing properties to a taste neuron normally not responding to these compounds.
In contrast to fructose or sucrose, ribose does not act as a major energy source (Fig 1B). Thus, we considered roles of RNA and its precursors as essential nutrients for larval growth. Inosine and uridine are components of HM, a synthetic Drosophila food medium composed of approximately 40 pure chemicals that can sustain larval growth, adult development and fertility [30]. To examine the contribution of these compounds for growth and viability, we examined developmental progression and survival rate of larvae kept on HM, HM lacking inosine and uridine (HMΔ), and modified HM in which RNA was added back to HMΔ (HMΔ + RNA; Fig 5A; see Material and methods). Confirming the findings of Piper and colleagues, HM supported larval growth and adult development with the same survival rate as larvae kept on standard cornmeal food (SCF), albeit at a slightly reduced pace. In contrast, lack of ribonucleosides in the medium (HMΔ) sharply increased mortality rate and significantly extended the larval growth phase, while adding back RNA (HMΔ + RNA)—but not ribose (HMΔ + ribose)—restored both survival and developmental time. Thus, RNA or the RNA precursors uridine and inosine are essential for rapid growth and viability. Given that ribonucleosides contain a ribose moiety, we expected that they are ligands for the Gr28 proteins and could activate TO taste neurons (Fig 5B). To test this, we first performed CaMPARI imaging experiments of wild-type larvae and found that uridine and inosine activated Gr28a-GAL4 neurons, but the other 3 ribonucleosides (guanosine, cytidine and adenosine) did not (Fig 5B). Second, we subjected wild-type and ΔGr28 mutant larvae to the two-choice preference assay and found that larvae were attracted to uridine and inosine in a Gr28-dependent manner (Fig 5C).
The experiments presented thus far identified a previously unknown taste modality for RNA (Figs 3 and 4) and 2 RNA precursors, inosine and uridine, and established a requirement for these compounds during larval life (Fig 5). Thus, we sought to determine whether the Gr28 proteins provided a competitive advantage when larvae were presented with a challenging food environment. We devised an assay in which wild-type and Gr28 mutant larvae were required to find nucleosides in HM food during all larval life stages (Fig 6). Eggs were placed in a feeding arena that consisted of 21 wells, only 9 of which contained complete HM food while the remaining 12 contained HMΔ food (Fig 6A). Control experiments using this setup confirmed a requirement for ribonucleosides (see above), regardless of whether a Gr28 locus was present or not (Fig 6B, compare solid versus light bars). When larvae were provided with the challenging food arena (HM/HMΔ), ΔGr28 mutant larvae showed a large increase in mortality (red checkered bar), while wild-type larvae or ΔGr28 larvae containing the Gr28 genomic rescue construct (black and green checkered bar) showed the same high survival rate as larvae kept on HM food (solid bars). Taken together, these findings establish that larvae can discriminate between HM food based on the presence or absence of inosine and uridine and that they use this ability to increase fitness and survival when presented with a challenging food environment.
It has been assumed that carbohydrates, amino acids/proteins, and fats are the only major macronutrient chemicals, which serve not only as energy source but also provide the essential cellular components during growth and development. Thus, the ability to sense these 3 types of nutrient compounds by taste cells has evolved, albeit to some extent independently, in all major animal orders [1]. Here, we have established that ribonucleosides and RNA are not only potent taste ligands for larvae but also essential macronutrients necessary for the unparalleled growth and body weight gain during larval stages of Drosophila. Given the high level of conservation of the Gr28 genes, it is intriguing to speculate that ribonucleosides and RNA might also be critical nutrients in other insects characterized by a rapid larval growth phase.
We have shown that larvae sense and are attracted to ribose-containing compounds, a behavioral feature mediated by the Gr28 protein family, a set of 6 evolutionarily highly conserved Gr proteins [22]. In contrast to sugar receptors and bitter receptors [6, 7, 11, 18, 19, 31–34], the Gr28 proteins appear not to function in combination with other Gr proteins, based on multiple lines of evidence. First, many cells and neurons express only a single member of the Gr28 gene family (Fig 2), although it remains possible that other Gr genes are co-expressed in some cells. Second, functional rescue experiments indicate that many Gr28 protein can restore preference for RNA and ribose in ΔGr28 mutant larvae when either expressed in Gr28a taste neurons (all 6) or in Gr43a fructose-sensing neurons that do not express any Gr28 gene (3 of 6) and are therefore also unlikely to express a potential coreceptor that might be present in all Gr28 taste neurons. However, we cannot rule out that Gr43a or another Gr present in these cells is incorporated into a Gr28-based multimeric ribose and RNA receptor complex. At the very least, our data indicate that Gr28 proteins convey specificity for these ligands. We note that a single Gr protein (Gr28b.d) conveys temperature sensing in neurons of the fly antenna [35], although in this case, no chemical ligand for this sensory modality has been identified to date. Surprisingly, while both inosine and uridine are potent ligands for Gr28, the 3 other ribonucleosides (guanosine, adenosine, and cytidine) are not. It will be interesting to determine why some, but not all, of the ribonucleosides can activate the Gr28 proteins. One possibility is that the amino group present in guanosine, adenosine, and cytidine prevents or interferes with binding of these chemicals to the Gr28 proteins.
Many Gr28 genes are also expressed in the larval brain and the gut (Fig 2), and it is possible that these cells contribute to preference of ribose-containing substrates via a postprandial mechanism. Three lines of evidence suggest that peripheral taste is the major driver for sensing these chemicals. First, larvae respond within 2 minutes to ribose and RNA (S2 Fig). Our previous studies on the fructose receptor Gr43a have shown that such rapid decision-making is mediated by peripheral taste neurons, whereas postprandially mediated fructose sensing is a much slower process, requiring about 8 minutes to establish a clear preference [4]. Second, Gr28a-expressing taste neurons are activated by these ligands (Fig 4A), and third, some Gr28 proteins can convey ribose and RNA preference to other taste neurons (Figs 3C and 4B). Together, these observations suggest that Gr28-expressing taste neurons are activated by ribose and ribonucleosides and that this activation leads to the rapidly established preference for these ligands.
The discovery that RNA and ribonucleosides are an essential nutrient resource recognized via a distinct taste modality represents a precedent. While detected through the ribose moiety (Figs 1, 3 and 4), our data indicate that the ribose serves only as a proxy for the detection of the nucleobase–sugar complex and not as the critical nutrient component per se, e.g., as a sugar used for energy production. In contrast to RNA, ribose complementation of HMΔ medium did not rescue development time or larval survival (Fig 5A). We propose that RNA or ribonucleosides are sensed by the larval chemosensory system because they are required in large amounts as cellular components with critical roles in gene and protein expression during the accelerated growth phase of larvae. RNA contributes about 4% and 20% of the dry weight of mammalian cells and bacteria, respectively, quantities that are in the same range as fat or polysaccharides (about 7% in both mammalian cells and bacteria). In contrast, DNA contributes only about 1% of the dry weight. Because the larval stage is characterized by unprecedented growth—evidenced by a doubling of body weight almost twice a day over a period of 4.5 days—it appears highly beneficial for larvae to be able to sense this abundant cellular constituent as an appetitive taste stimulus. In the natural environment, RNA is likely obtained from microorganisms that colonize decaying fruit.
RNAs might be of interest to chemosensory evaluation from another viewpoint. Specifically, microRNAs (miRNAs) have recently been implicated in regulating the microbiome in mammals [36], and double-stranded RNAs (dsRNAs)/small interfering RNAs (siRNAs) can cross cell membranes of the gut epithelium, a tissue well known for its ability to sense diverse types of chemicals [37]. Thus, it will be interesting to explore potential roles for Gr28 proteins in RNA sensing and transport in the larval gut, where all but 1 of the 6 genes are expressed (Fig 2).
This paper represents the first clear evidence for a chemical compound acting as a ligand for members of this enigmatic Gr protein subfamily. While our study showed a specific role for these receptors in larval feeding on ribose-containing substrates, their function in adult flies remains to be investigated. Using proboscis extension reflex (PER) assays, we have found no evidence that adult flies respond to ribose, in either an appetitive paradigm or a feeding suppression paradigm (S3 Fig). This is not surprising given that adults, in contrast to larvae, have a much lower requirement of cell proliferation and growth, which is restricted to the female germline and stem cells in a few organs of the fly. Expression analyses have shown that the Gr28 genes are broadly expressed in all taste organs (labial palps, tarsi and pharyngeal taste neurons), and most of them appear to be expressed in bitter taste neurons [22]. Bitter taste receptors are mutlimeric complexes that are activated by non-nutritious and often toxic chemicals and when activated suppress appetitive taste behavior [16–19]. Thus, in fly taste neurons, Gr28 subunits most likely combine with other Grs to form receptor complexes for such ligands. It will be interesting to see whether these ligands share any structural features with ribose.
Gr28 proteins have been reported to have functions in other sensory modalities, such as temperature sensing in flies and light sensing in the larvae. Specifically, Gr28b.d is expressed in 3 neurons of the aristae [22] and was later shown to be important for avoidance of warm temperatures [27]. Gr28b.d conveys thermosensitivity to a number of other cell types, suggesting that this protein acts on its own in the absence of other Gr proteins. Members of the Gr28 protein family were also implicated in light avoidance of larvae, which is mediated by multidendritic neurons in the body wall, where several of the Gr28 genes are expressed [22] (Fig 2). The Caenorhabditis elegans Gr28 ortholog, lite-1, is necessary in worms for the avoidance of visible and UV light [38–40]. Two different mechanisms have been suggested for how LITE-1 senses light: one study proposed that light avoidance is an indirect chemosensory response, related to the avoidance of hydrogen peroxide of worms [40], while another group suggested that LITE-1 directly absorbs light through 2 tryptophan residues [41]. In any case, the Gr28 proteins and its related cousin in C. elegans represent a decidedly atypical type of chemoreceptor that appears to be involved in diverse sensory modalities not associated with taste. Thus, future work will be necessary to identify additional ligands for Gr28 proteins and reveal the many roles of these receptors in physiology and behavior of insects and worms.
Flies were raised on SCF at 25°C on a 12-hour light–dark cycle. SCF in 1.5 L of water is composed of 10.88 g of agar (Drosophila agar type II Genesee; 62–103), 78 g of corn meal (Genesee, 66–101), 165 g malt extract (Alternative Beverage, MUN-UL), 41.25 g of yeast extract (Genesee, 62–106), 4.69 g of propionic acid (VWR, TCP0500-500mL), 0.075 g chloramphenicol (Sigma-Aldrich, C0378), and 2.11 g of tegosept (Sigma-Aldrich, PHR1012).
Chemicals used for two-choice feeding preference assay and CaMPARI imaging were fructose (Sigma-Aldrich, F0127), 2-deoxy-d-ribose (Sigma-Aldrich, 31170), ribose (Sigma-Aldrich, R7500), arabinose (Sigma-Aldrich, 10850), t-RNA (from brewer’s yeast; Sigma-Aldrich, 10109525001), cytidine (Sigma-Aldrich, C4654), guanosine (Sigma-Aldrich, G6752), adenosine (Sigma-Aldrich, A9251), DMSO (Sigma-Aldrich, D8418), uridine (Sigma-Aldrich, U3750), and inosine (Sigma-Aldrich, I4125).
Third-instar feeding-stage larvae were collected from food vials by washing them out using water. They were placed along the midline of a feeding arena (plastic petri plate 60 × 15 mm, Falcon) containing freshly prepared 1% agarose on one side and 1% agarose mixed with tastant on the other side. They were left feeding, and their location (agarose versus agarose plus tastant, respectively) was recorded after 2, 4, 8, and 16 minutes. For simplicity, the preference index (PREF) was calculated for the 8-minute time point based on the number of larvae on either half of the plate. Dose-response curves for all sugars were performed using a concentration range from 25 mM to 500 mM (S2 Fig). For preference tests of wild-type and mutant larvae (Figs 1, 3, 4 and 5), a concentration of 100 mM was used for arabinose, ribose, 2-deoxyribose, and inosine, while 50 mM was used for uridine. Stock solutions for all the tastants were prepared in water before mixing in the agarose. For ribose and 2-deoxyribose, the stock solution was treated with charcoal and filtered so as to remove the unrelated odor. “PREF” indicates the number of larvae on agarose plus tastant minus the number of larvae on the agarose only, divided by the total number of larvae. A PREF score of 0 indicates no preference, while a score of +1 (or −1) indicates all larvae preferred (or avoided) tastant over agarose alone.
Survival time of second-instar larvae kept on a 60 × 15 mm petri plate (Falcon) filled with 1% agarose containing various carbohydrates at 100 mM concentration and RNA at 0.5 mg/mL. Survival was monitored daily until all larvae died. Dead larvae were removed daily to avoid scavenging.
For calcium imaging, we used the slide preparation method as described by Alves and colleagues [42]. A single live larva was placed in 25 μl of distilled water between a cover slip and a perforated slide (a hole of 0.5 cm diameter was made on the slide with a circular drill bit). The larva was expressing CaMPARI calcium sensor described by Fosque and colleagues [29]. This preparation was placed on an inverted Nikon A1 confocal microscope and observed with 20× objective. To activate the neurons, a 25 μl solution (2× of final concentration) of the ligand was injected through the hole in the slide. After about 5 seconds, a pulse of photoconversion (PC) light of 405 nm with a power of approximately 1.8 w/cm2 was delivered to the larvae for 10 seconds. Post activation, the neurons were observed for conversion from green to red emitted wavelengths. The changes were calculated as the ratio of red/green fluorescence (FRED/FGREEN). Data were acquired using the Nikon NIS element acquisition and analysis package. Data are expressed as mean ± SEM. To determine preexposure values, the images were taken without applying ligand or PC light. To determine PC-light–only values, images were taken with applying PC light but without applying ligand.
HM was prepared based on diet reported by Piper and colleagues [30]. A detailed list of chemicals is listed in S1 Table. For HMΔ food, inosine and uridine were removed. For supplementation experiments, HMΔ food was complemented by addition of RNA and ribose final concentrations of 0.5 mg/mL and 0.5 mM, respectively. For survival experiments, HM food and its derivatives were presented in 60 × 15 mm plastic petri plates (Falcon 5 ml/plate; Fig 5) or in 21-well microtiter plates (350 μl/well; Fig 6). These plates were then embedded in a 60 × 15 petri plate using 3% agarose. After plates were at room temperature, 40 eggs were placed on each plate, which were sealed with a perforated lid and placed in a humid chamber maintained at 25°C in a 12-hour light–dark cycle. Larvae location was checked once per day, and dead larvae were removed to avoid scavenging.
In all figure legends, n indicates the number of experiments, unless otherwise noted. For each experiment, data are presented as mean ± SEM. Statistical analysis was performed using two-tailed nonparametric Mann-Whitney U test to compare 2 different groups of samples. p < 0.05 was considered to be statistically significant. Statistical analyses were conducted using Prism 6.0 software (GraphPad Software).
PER assays were carried out as described in Slone and colleagues [7] with the following modifications. Male and female flies were collected on the day of eclosion and kept on standard corn meal food for 5 to 6 days at 25°C. Prior to performing PER assays, flies were starved for 24 to 26 hours at 25°C in empty vials with a water-saturated cotton ball. Flies were immobilized by cooling briefly on ice and wing-mounted dorsally on a microscope slide using double-sided Scotch tape. Legs were taped to the slide. Mounted flies were allowed to recover for 60 to 90 minutes at room temperature in a humidified chamber. Flies were then allowed to drink water until satiation to ensure that PER responses were nutrient derived. Flies showing no response to water were excluded. Each fly was tested with a given tastant by briefly applying the taste solution to the labellum. Each fly was tested 3 times for each taste solution. A PER response was scored as positive (1) if the proboscis was fully extended, otherwise it was scored as negative (0). PER response scores (%) from a single fly were 0% (0/3 responses in the 3 applications), 33% (1/3), 66% (2/3), or 100% (3/3). Flies were allowed to drink water after each taste application. Taste solutions were delivered with a 20 ml pipette. Stock solutions of sucrose (Macron, Cat No. 8360–06) and ribose (Sigma, Cat No. R7500) were prepared in Millipore Q water and kept at 5°C. Stock solutions were diluted to the final concentration using Millipore Q water prior to each experiment.
|
10.1371/journal.pntd.0001486 | Development and Characterization of a Reverse Genetic System for Studying Dengue Virus Serotype 3 Strain Variation and Neutralization | Dengue viruses (DENV) are enveloped single-stranded positive-sense RNA viruses transmitted by Aedes spp. mosquitoes. There are four genetically distinct serotypes designated DENV-1 through DENV-4, each further subdivided into distinct genotypes. The dengue scientific community has long contended that infection with one serotype confers lifelong protection against subsequent infection with the same serotype, irrespective of virus genotype. However this hypothesis is under increased scrutiny and the role of DENV genotypic variation in protection from repeated infection is less certain. As dengue vaccine trials move increasingly into field-testing, there is an urgent need to develop tools to better define the role of genotypic variation in DENV infection and immunity. To better understand genotypic variation in DENV-3 neutralization and protection, we designed and constructed a panel of isogenic, recombinant DENV-3 infectious clones, each expressing an envelope glycoprotein from a different DENV-3 genotype; Philippines 1982 (genotype I), Thailand 1995 (genotype II), Sri Lanka 1989 and Cuba 2002 (genotype III) and Puerto Rico 1977 (genotype IV). We used the panel to explore how natural envelope variation influences DENV-polyclonal serum interactions. When the recombinant viruses were tested in neutralization assays using immune sera from primary DENV infections, neutralization titers varied by as much as ∼19-fold, depending on the expressed envelope glycoprotein. The observed variability in neutralization titers suggests that relatively few residue changes in the E glycoprotein may have significant effects on DENV specific humoral immunity and influence antibody mediated protection or disease enhancement in the setting of both natural infection and vaccination. These genotypic differences are also likely to be important in temporal and spatial microevolution of DENV-3 in the background of heterotypic neutralization. The recombinant and synthetic tools described here are valuable for testing hypotheses on genetic determinants of DENV-3 immunopathogenesis.
| Infectious virus clones are valuable tools for studying how changes in viral genetic codes affect viral biology. Dengue virus is the most important mosquito-borne virus worldwide, yet dengue virus infectious clones have historically been challenging to make and manipulate, making it very difficult to study the variety of genetic changes observed in dengue viruses. Here we describe the construction of a panel of five dengue virus serotype 3 (DENV-3) clones using a novel strategy not previously employed in dengue research. This strategy uses genetic fragments and synthesized genes to introduce genetic changes while minimally affecting the virus. Each of the five recombinant clones was designed to express genetically distinct DENV-3 envelope proteins derived from strains circulating in different regions of the world. We used the recombinant viruses, coupled with DENV-3 sera from geographically defined human cases, to study the impact of E variation on neutralization outcomes. Our data demonstrate that the recombinant viruses varied significantly in their neutralization outcomes, depending on sera. While it has long been presumed that infection, and vaccination, with one serotype confers lifelong protection against all variants of that serotype, our results indicate that this assumption requires a more rigorous assessment by the DENV community.
| Dengue virus (DENV) is an enveloped (+) RNA virus in the family Flaviviridae, genus Flavivirus transmitted by the bite of Aedes spp. mosquitoes. DENV occurs throughout the tropics and subtropics and infects approximately 50 million individuals annually. There are four distinct serotypes, DENV-1–DENV-4. While prospective studies have found that most infections are asymptomatic, a proportion of infected persons will develop symptoms that include fever, rash and myalgia [1], [2] with 2% or less developing the severe disease syndromes of dengue hemorrhagic fever/dengue shock syndrome (DHF/DSS) [2], characterized by hemorrhage, vascular leakage, hypovolemia and, if untreated, shock, end organ failure and death [3]. Approximately 15,000–30,000 persons die annually from DHF [1]. DHF/DSS has been classically associated with secondary infections that occur in the context of pre-existing heterotypic immunity - leading to hypotheses that DHF/DSS is an immune mediated phenomenon driven by cross-reactive DENV antibodies and/or or DENV specific CD8+ T-cells (for reviews see: [4], [5]. Virus genotype also clearly plays an important role in severe disease pathogenesis, as. Multiple studies of DENV molecular epidemiology have found associations between circulating virus genotype and disease severity [6]–[12]. However, the genetic basis of these virulence differences has not been deciphered.
One of the fundamental barriers to DENV vaccine development has been concern that a DENV vaccine must be broadly protective against all four serotypes or recipients will risk secondary-like infection and the severe disease associated with naturally acquired secondary infection. Most vaccine trials have assessed protection against all four serotypes using prototype or vaccine related virus isolates [13] and studies need to address the degree to which intra-serotype genotypic differences may affect antibody-mediated immunity to any of the DENV serotypes, including DENV-3. While genotype specific genetic differences are scattered across the viral genome, the envelope glycoprotein (E) is the main target of neutralizing human antibody and is one logical first choice for assessing the genetic basis of differential antibody mediated neutralization of DENV-3 infection. The E glycoprotein exists as a homo-dimer with 3 distinct domains – I, II, and III [14]–[17], that, on the mature DENV virion, are arranged in a flat herringbone pattern with icosahedral symmetry [14]. Domains I (EDI) and II (EDII) are linearly discontinuous and fold to form a central eight-stranded ß barrel (domain I) with a lateral protrusion (domain II) that contains the highly conserved fusion loop required for virion fusion with endosomes. Domain III (EDIII) is a continuous peptide that extends from domain I and forms an Ig like fold that is believed to be the ligand for an as yet unidentified cellular receptor.
A successful dengue vaccine should induce broadly protective antibodies against all geographic variants of each serotype. The dengue community has long held that primary infection with one serotype confers long-lasting immunity to that serotype, irrespective of the infecting virus genotype. This is based principally on early human challenge trials [18] and multiple observational studies that have shown that, within a particular region, re-infection with the same serotype generally does not occur. Geographic partitioning of DENV genotypes significantly limits our understanding of the role of strain variation in protective immunity, as the vast majority of DENV infected persons in endemic regions never travel to regions where other DENV genotypes are circulating. Several recent findings indicate that genotypic variation may be important in immunity. Recent studies of DENV-3 strain variants using recombinant proteins and whole virus have found that neutralization mAbs raised against one DENV-3 genotype have limited neutralization activity against heterologous genotypes [19]–[22]. After primate vaccination, studies with polyclonal immune sera have also demonstrated variable neutralization of DENV3 strains [23]. In a study of pediatric dengue cases in Thailand, investigators observed significant differences in the ability of sera to neutralize reference and clinical strains of DENV3 [24]. A recent WHO report on dengue neutralization testing highlighted the need for evaluating vaccine induced immune responses using contemporary strains representing the different serotypes and genotypes of dengue [25].
DENV-3 consists of four distinct genotypes: I, II, III and IV, each originally associated with a specific geographic region [26]. Currently genotype I and II are circulating in Asia, genotype III is circulating in the Indian subcontinent, Africa and Latin America, and genotype IV appears to have been displaced but occurred throughout the Caribbean in the 1960s and 70s [7], [26]–[31]. Here we described the construction of a four-fragment DENV-3 infectious clone platform and a panel of isogenic DENV-3 recombinant viruses that captures DENV-3 E glycoprotein genotypic heterogeneity. While our approach is novel for flaviviruses, human coronavirus (CoV) investigators have used a similar system to introduce large, synthesized genomic elements into recombinant viruses to investigate genetic variability in CoV biology and pathogenesis (see [32]–[35] for examples). The CoV systems are a powerful tool for expanding understanding of genetic differences in CoVs and the application to Flaviviruses may prove similarly powerful. We subsequently tested the isogenic recombinant viruses against a panel of immune sera from people exposed to primary or secondary DENV infections. These data demonstrate a role for natural epitope variation in virus neutralization and escape. The molecular clone should also prove to be a valuable tool for studying a variety of other aspects of DENV-3 biology, pathogenesis, immunopathogenesis, epitope mapping and evolution.
The Institutional Review Board of the University of North Carolina at Chapel Hill approved the protocol for recruiting and collecting blood samples from people. Written informed consent was obtained from all donors.
Vero E6 cells (ATCC CRL-1586) were maintained in MEM supplemented with 10% FCS (Gibco), non-essential amino acids (Gibco), L-glutamine (Gibco) and Anti-Anti antibiotic mix (Gibco) at 37°C in 5% CO2. C6/36 cells (ATCC CRL-1660) were maintained in MEM supplemented with 5% FCS, non-essential amino acids, L-glutamine and Anti-Anti at 28°C in 5% CO2.
The cloning strategy for the DENV-3 clone is illustrated in Figure 1A, and based on strategies employed with CoVs to circumvent sequence instability problems in E.coli [36], [37]. The clone parent is a 1989 Sri Lankan DENV3 isolate (genotype III) designated UNC3001 (submitted to GenBank). To isolate the DENV-3 sub-clones, reverse transcription was performed with AMV reverse transcriptase (Roche) and oligodeoxynucleotide primers according to the manufacturer's recommendations using primer BsmbIDen. Following cDNA synthesis, the cDNA was amplified by PCR with Expand Long TAQ polymerase (Boehringer Mannheim Biochemical) with cycle settings based on the size of the amplicon. The Dengue genome was amplified from cDNA and cloned as a set of four fragments (Figure 1 and Text S1). The first fragment, A, was PCR amplified using primer set DEN#1 and DEN2kb−. These primers created a T7 RNA promoter at the 5′ end of the fragment and a BsmBI restriction site at its 3′ end, respectively. The PCR product was gel isolated (Qiagen QIAquick Gel Extraction Kit) and then cloned into the pCR-XL TOPO cloning vector (Invitrogen).
The second fragment, B, was amplified using primers DEN2kb+ and DENBGL4−. The DEN2kb+ primer introduced a BsmBI site that allowed for the directional ligation of fragments A and B (Figure 1 and Text S1). The DENBGL4− primer introduced silent changes in the Dengue genome between nucleotides (nt) 3150 and 3160 to create a unique BglI site without altering the amino acid sequence. Fragment C was amplified with primers DENBLG3+ and DEN7kb−. This primer set duplicated the BglI site at the 3′end of the B fragment and a naturally occurring BglI site at nt 7031. The PCR amplicons for both fragments B and C were gel isolated and cloned into the pCR-XL TOPO cloning vector.
Fragment D was amplified with primers DEN5kb+ and BsmBIDen. This PCR product, which went from approximately nt 5100 to the 3′ end of the Dengue genome, contained two BglI sites; one at nt 7032 and the other at nt 10186. The BglI site at nt 10186 was removed using overlapping PCR. Two amplicons – 3′ and 5′, were generated using primers Dengue15 and Den10198 and primers Den10166 and BsmBIDen, respectively. These two amplicons were joined in an over-lapping extension PCR reaction. The resulting product was digested with SapI and ligated to SapI digested DEN D fragment. This final cDNA fragment, which now had the BglI site at 10186 knocked out, was gel isolated and cloned into the Big Easy v2.0 Linear cloning vector (Lucigen).
Four to six clones of each fragment were sequence verified. The four DEN cDNAs were isolated from plasmids and directionally ligated to create a full-length cDNA of the dengue viral genome. This full-length cDNA contained only the introduced nucleotide changes, all of which were silent, and could be transcribed with T7 polymerase (Ambion). This RNA produced infectious dengue virus when electroporated into Vero E6 cells.
To construct E glycoprotein variant clones (Figure 1B), synthesized envelope genes (nucleotides 913–2416 of the Dengue genome) were delivered in puc57 plasmids (Bio Basic). The portion of these envelope genes that needed to be inserted into the A plasmid, was PCR amplified with either a puc57 forward or reverse primer and the Den2kb− primer (Text S1). These products were digested with BstEII and BsmBI and ligated into the A plasmid which had been digested with the same enzymes. Dengue B plasmids containing the envelope variants were generated by first PCR amplifying the synthetic genes with the Den 2kb+ primer and primer EGENE− (Text S1) and the parent B fragment with primer EGENE+ and DENBGL4− (Text S1). These products were then digested with BsaI and ligated together. Finally, the ligations were gel purified and cloned into the pCR-XL TOPO cloning vector.
To replace the parent clone prM/M gene with a genotype I prM/M gene, RNA from our lab stock genotype I virus UNC3043, was reverse transcribed with random hexamers and the cDNA was PCR amplified with primers Dengue01+ and Denv900. The resulting amplicon was digested with BstAPI and PflMI. This product was ligated into the DEN A plasmid corresponding to the Indonesia 1982 genotype I E gene that had been digested using the same enzymes. The resulting plasmid DEN A was sequence verified and used to construct the genotype I recombinant virus.
Each plasmid was transformed and propagated in E. coli TOP10 competent cells (Invitrogen) and grown on LB plates with selective antibiotics (A, B, and C containing plasmids selected with kanamycin, D with chloramphenicol) at 28.5°C for 24 hours. Individual colonies were picked, screened and sequenced. The plasmids were subsequently grown to high concentration in selective LB, plasmid purified (Qiagen Mini-Spin Kit) and digested as follows according to manufacturers instructions: DEN A with SpeI (NEB) followed by calf intestine phosphotase (NEB) and BsmBI (NEB) yielding a 2.0 kb fragment; DEN B with BglI (NEB) and BsmbI yielding a 1.1 kb fragment; DEN C with BglI yielding a 3.9 kb fragment; and DEN D with BglI and BsmbI yielding a 3.0 kb fragment. Fragments were gel-isolated (Qiagen Gel Extraction Kit) on 0.8% agarose gel, mixed in equivalent copy number and ligated with T4 ligase (NEB) overnight at 4°C. Full-length transcripts of DENV-3 cDNA constructs were generated in vitro as described by the manufacturer (Ambion, Austin, Tex; mMessage mMachine) with the following modifications: For 30-µl reaction mixtures supplemented with 4.5 µl of a 30 mM GTP stock, resulting in a 1∶1 ratio of GTP to cap analog and incubated at 37°C for 2 hours. Vero cells were grown to 75% confluence, trypsinized and resuspended in RNAse free PBS at 107 cells/ml. RNA transcripts were mixed with 800 µl of the Vero cell suspension in an electroporation cuvette, and four electrical pulses of 450 V at 50 µF were given with a Bio-Rad Gene Pulser II electroporator. The transfected Vero cells were seeded at 5×106/ml in 75-cm2 flask and incubated at 37°C for 4 days. Two to five ml of supernatant from electroporated Vero cells were passaged on day 4 to 75% confluent uninfected Vero cells in a 75 cm2 flask. Fresh media was added to a final volume of 15 ml. Seven day supernatants were harvested, supplemented to 30% FBS, clarified by centrifugation and frozen at −80°C or passaged serially to amplify a working virus stock.
At the time this study was initiated, there were 164 unique, full-length DENV-3 envelope genes available in Genbank, and these sequences were added to 11 Sri Lankan DENV-3 sequences from our laboratory. The 175 envelope amino acid sequences were aligned using ClustalX version 1.83 [38], and one representative sequence was selected for each DENV-3 genotype. The representative sequence was chosen based on amino acid conservation within the genotype cluster, with sequences closest to consensus with no outlier amino acids selected as the representative. Representative sequences chosen were: Genotype I Indonesia 1982 (GenBank accession# DQ401690.1); Genotype II Thailand 1995 (GenBank accession# AY676376); Genotype III Cuba 2002 (GenBank accession# AY02031); and Puerto Rico (PR) 1977 (GenBank accession# AY146761). All viruses used in the subsequent experiments were passage three propagated in Vero cells. All passage three clones were sequence verified using previously described methods [39].
To assess viral replication kinetics, each of the DENV-3 clones was inoculated in triplicate onto 95% confluent monolayers of Vero or C6/36 cells in 6 well plates at a multiplicity of infection (m.o.i) of 0.01 ffu/ml. Cells were incubated at either 37°C for Vero or 27°C for C6/36 cells under maintenance media conditions for the cell line for 60 minutes, after which the innocula were removed and cells washed twice in 3 ml of PBS. Each monolayer was covered in a total volume of 5 ml media. After 60 min, 200 ul of cell supernatant, designated as the Day 0 sample, was taken in duplicate with equal volume media replaced. Samples were supplemented with 30% FCS, clarified by centrifugation and stored at −80°C. Samples were taken in the same manner every 24-hrs for 6 additional days. Virus titers were determined as described below.
Sera were collected from adult volunteers with histories of DENV infection [40] and one anonymous donor with dengue infection confirmed by serology (sample 109). Sera were characterized by flow cytometry at UNC [41], PRNT60 at the NIH, Bethesda, MD, or PRNT90 at CDC San Juan to confirm past exposure to primary or secondary DENV infections and also to identify the serotype responsible for primary infections. We note that we cannot establish the infecting virus genotype of our experimental sera on neutralization patterns alone. However, only genotypes I, II, and III are currently circulating, and our samples almost certainly capture genotype II (Thailand) and III (Latin America) based on donor travel history.
The FRNT procedure is based on a method previously described by Whitehead [42]. Briefly, twenty-four well plates were seeded with 5×104 Vero cells in MEM supplemented with 5% fetal bovine serum (FBS) and grown for 24 hours. Growth media was removed. For virus titration, virus stocks were diluted serially ten-fold from 10−1 to 10−6 and 200 ul of each dilution added to individual wells. After 1 hr incubation on a rocker at 37°C, the wells were overlaid with 1 ml 0.8% methylcellulose in OptiMEM (Gibco) supplemented with 2% FBS (Cellgro) and antibiotic mix (Gibco Anti-Anti). Plates were incubated 5 days at 37°C, 5% CO2. On day 5, overlay was removed, cells washed with PBS, fixed in 80% methanol and either stored at −80°C or developed. To develop plates, fixed monolayers were blocked for 10 minutes with 5% instant milk PBS, followed by incubation with anti-flavivirus MAb 4G2 diluted 1∶1000 in blocking buffer for 1 hr at 37°C. Wells were washed with PBS and incubated with horseradish peroxidase (HRP) conjugated goat anti-mouse Ab (Sigma) diluted 1∶500 in blocking buffer for 1 hr at 37°C. Plates were washed once in PBS and foci developed by the addition of 100 ul of TrueBlue HRP substrate (KPL). Foci were counted on a light box and viral titers calculated by standard methods. For FRNT, MAbs or human sera were serially diluted five-fold from starting dilutions of 1∶5 or 1∶10. Each dilution was mixed with approximately 30 focus forming units (ffu) of virus to a final volume of 200 ul, incubated for 1 hour at 37°C, 5% CO2 and added in triplicate to 24 wells plates and processed as above. Mean focus diameter was calculated from ≥20 foci/clone measured at 5× magnification.
Multiple alignments were performed using ClustalX version 1.83 [38] and phylogenetic trees of the envelope protein sequences were conducted using Mr. Bayes version 3.12 (Huelsenbeck JP, 2001). Briefly, 175 amino acid envelope sequences were imported into ClustalX and the alignment was performed using default parameters. Structural models of the informative sites were generated using MacPymol (Delano Scientific) and the crystal structure of DENV-3 envelope (PDB 1UZG) [16]. Mean focus sizes were compared by one-way analysis of variance (ANOVA) followed by Dunnett's test for multiple comparisons. Growth curve and FRNT counts were entered into Graphpad Prism (Version 5.00 for OSX, GraphPad Software, San Diego California USA, www.graphpad.com). FRNT50 values were calculated by sigmoid dose-response curve fitting with upper and lower limits of 100 and 0 respectively. All error bars show 95% confidence intervals unless otherwise specified. Mean FRNT50 values were compared by one-way ANOVA followed by Tukey HSD multiple comparison test with significance level alpha (P) set at <0.05.
The parent DENV-3 clone is a genotype III variant isolated from a Sri Lankan DF patient in 1989 (Figure 2) (See materials and methods). Full-length flaviviruses genomes have been previously described as unstable and toxic in traditional E. coli clone systems [43]–[46]. To disrupt the putative toxic regions and facilitate creation of chimeric DENV-3 clones, the genome was cloned into segmented, sequential fragments. The fragments and junctions in the final platform were chosen to through multiple trials to maximize insert and plasmid stability in E. coli. Clone junctions were based on type IIS restriction enzyme sites (BsmBI and BglI) (Figure 1A) that allow directional assembly into full-length cDNAs as described in Materials and Methods. After digestion and purification of individual cDNAs, the full-length cDNA was assembled by in-vitro ligation, transcripts were electroporated into cells and recombinant viruses were recovered from first passage Vero cell culture supernatant. Sequence analyses verified indicator mutations within the cDNA clone fragments and no nucleotide mutations were detected in the entire genome of the recombinant virus after three passages in Vero cells (data not shown).
To evaluate the role of DENV3 E protein sequence variation on antibody interactions, representative E genes from genotype I, II, III and IV viruses (Figure 2, Text S1) were selected from 175 published DENV-3 sequences. Each E gene was selected to represent a genotype whose sequence most closely matched a consensus E sequence generated for each genotype. Genotype I is a 1982 Indonesia isolate, genotype II is a 1995 Thailand isolate, genotype III a 2002 Cuba isolate, and genotype IV a 1977 Puerto Rico isolate. A total of 32 informative sites were identified across the representative genotypes (Materials and Methods), forming nine clusters on the surface of the E glycoprotein, relatively evenly distributed through domains I, II and III (Text S1).
To generate clones that would allow testing of variable neutralization, these representative sequences were synthesized by Bio Basic and inserted into the parent clone background, replacing the parent E gene (Figure 1B). Three of the four variant clones were successfully recovered with correct replacement of the E gene alone. One variant, however, Indonesia '82 (genotype I), required the replacement of the parent SL '89 genotype III preM/M gene with a genotype I preM/M gene, supporting earlier studies that co-evolutionary changes in preM/M may be essential for efficient E gene function in select instances [47]. Full-length sequencing of all passage three recombinant virus clones used throughout these experiences found only one nucleotide mutation in one of the five clones, a silent C to T pyrimidine transition mutation at genomic position 7043 in the genotype I virus.
Because some DENV clinical isolates do not reliably form plaques on Vero cell monolayers, viral growth on Vero cell monolayers was instead characterized through focus formation (see Materials and Methods). All five clones formed foci on Vero cell monolayers. The parent clone, SL '89 (III) and Cuba '02 (III) produced moderate sized and relatively uniform foci after 5d growth on a Vero cell monolayer (Table 1). Clones with Indonesia '82 (I) E genes produced marginally smaller foci, while Thailand '95 (II) and PR '77 (IV) foci were markedly smaller than those formed by the parent clone (Table 1). The striking difference in plaque phenotype underscores the importance of structural proteins in basic viral biology, and may be due to either E gene differences in the virus envelope or prM-E mismatch in the virus clones, though identifying the particular genetic differences causing the phenotype is beyond this paper's scope.
The growth kinetics of the panel of recombinant viruses were characterized in mammalian Vero cells and C6/36 mosquito cells, both of which are a commonly used for DENV propagation and quantification. Both cell lines were infected with the parent and clones at a multiplicity of infection (MOI) of 0.01 FFU/cell and grown for 216 hours. In Vero cells, the growth curves for the parent virus and the five clones were similar, with all preparations producing focus-forming virus after 24 hours and peak viral titers achieved between 120 hours and 168 hours (Figure 3A). Peak log viral titers ranged from 6.68 log FFU/ml for the parent clone to 5.10 log FFU/ml for the genotype II clone. Early growth was slower in the genotype I Indonesia recombinant virus, but ultimately reached peak titers equivalent to the other recombinants. Growth kinetics in C6/36 cells were similar to those in Vero cultures except that virus was not detected until 48 hrs post infection (Figure 3B). Peak titers were generally similar, though the parent clone had a single peak log titer of 7.70 log FFU/ml that was significantly higher than the other virus samples. The remaining peak titers ranged from 6.30 log FFU/ml to 6.66 log FFU/ml and did not differ significantly. The genotype I Indonesia clone did show slower kinetics than the other clones, particularly early in infection (Figure 3B). Overall, inter-genotypic E variability had minimal impact on the viruses' growth in tissue culture.
To assess the role of DENV E glycoprotein variation on viral neutralization by human polyclonal sera, the isogenic clones were tested against a panel of late convalescent (>2 years) human anti-DENV primary and secondary sera collected from individuals in North Carolina who had been infected during foreign travel [40](Table 2 and Text S1). The majority of the neutralization tests were repeated in independent experiments, with highly reproducible FRNT50 values (Text S1). The original infecting virus is not known for any of these sera.
Eight primary anti-DENV-3 serum samples were tested against the parent and isogenic recombinant viruses with variable E genes from the different DENV-3 genotypes (Text S1). The clones did not show differential neutralization patterns against three of the sera; 003, 005 and 103 (Figure 4A, B, and E). Serum sample 003 was taken from a traveler who acquired a primary DENV-3 infection in Thailand. FRNT50 titers for 003 ranged from 1∶59 for Cuba'02 (III) to 1∶203 for Indonesia '82 (I) (Text S1). Serum sample 005 was taken from a traveler who acquired a primary DENV-3 infection in Puerto Rico. Calculated FRNT50 were similar to those observed for 003, with titers ranging from a low titer of 1∶31 against PR '77 (IV) to a high of 1∶118 against the Sri SL '89 (III) clone and Indonesia '82 (I) (Text S1). Serum sample 103 was from a traveler infected with DENV-3 in Nicaragua in 1995. FRNT50s ranged from a low 1∶42 (PR '77 (IV)) to a high of 1∶117 for Thailand '95 (II) (Text S1). While these FRNT50 values are consistent with most accepted cutoffs for true homotypic neutralization, they are consistently low, with six of the fifteen clone titers in this group less than 1∶60 (Text S1).
More importantly, we found significantly variability neutralization profiles against the five recombinant viruses neutralized with the five remaining homotypic sera tested (Figures 5C, 5D, 5F, 5G and 5H). Though serum sample 011, from an El Salvador infection, neutralized all five clones, we found a 9-fold difference (P<0.05) between the calculated lowest and highest neutralizing titers, with a low neutralizing group consisting of Indonesia '82 (I), 1∶133, and PR '77 (IV) 1∶157 and a second, high neutralizing group included the remaining clones Cuba '02 - 1∶701, Thailand '95 - 1∶1091 and SL '89 -1∶1172 (Text S1). Serum sample 033 (Figure 4D), from an infection in India, was similarly potent, with four of the five clone titers greater than 1∶780, but with the PR '77 (IV) recombinant virus again showing a significantly lower neutralization titer at 1∶304 (Text S1). Against samples 105 (Figure 4F) and 118 (Figure 4G), from infections in Thailand and Nicaragua respectively, the neutralization titers differed five (105) and six (118) –fold between recombinant viruses expressing Thailand '95 or PR '77 E glycoprotein (P<0.05)(Text S1). The most extreme neutralization differences between the clones were seen using serum 109 (Figure 4F), from a Sri Lanka donor. This serum, with titers of 1∶177 and 1∶280, efficiently neutralized the Indonesia '82 and Thailand '95 clones respectively, while the genotype III clones were neutralized at much lower dilutions of 1∶40 (Sri Lanka) and 1∶15 (Cuba) (Text S1). Thus, we observed significant variation in neutralization across DENV-3 genotypes for five of the eight primary homotypic sera tested.
Human anti DENV secondary sera are known to be broadly neutralizing across serotypes, and we would expect it to show relatively high and broad FRNT50 values and resist intra-genotypic variability. To test this assumption, each of the five clones were tested against 009, serum from a patient who had a secondary DENV infection, in India or Sri Lanka in 2000. All of the clones were efficiently neutralized at relatively high titers, though the highest (Indonesia '82) and lowest (Cuba '02) did differ significantly (Text S1, Figure 4H), though this difference was less than three-fold.
Heterotypic primary anti-DENV serum may have low-level serotype-cross neutralizing activity, and in one study was shown to be protective for heterotypic infection in some cases [48]. To assess the role of E glycoprotein variation in heterotypic cross-neutralization, the clone panel was tested against representative primary anti-DENV-1, -2, and -4 sera (Table 2, Text S1). Sample 001 was collected after a primary DENV-2 infection acquired in Sri Lanka in 1996. 001 had low level but detectable FRNT50s that ranged from 1∶11 to 1∶78 (Figure 5A, Text S1). Serum 006 FRNT50 titers ranged from 1∶7 to 1∶54 (Figure 5B, Text S1), and sample 102, collected after a DENV-4 infection in Honduras, had a similarly scaled FRNT50 range of 1∶12 to 1∶58 (Figure 5C, Text S1). However, while repeat FRNT against the clone panel with homotypic sera yielded highly reproducible neutralization titers, repeat FRNTs were not reproducible for heterotypic sera (Text S1), significantly limiting any conclusions that might be drawn from variable heterotypic neutralization.
CJ Lai et al. described the first full-length infectious DENV clone for DENV-4 isolate 814669 (isolated from a patient in the Dominican Republic in 1981 [49]) in 1991 [46]. At that time, the authors noted the full-length DENV cDNA was unstable in E. coli. This was overcome by using a two-fragment system that divided the toxic genomic regions. Subsequent DENV-2 New Guinea C [45] and DENV-4 West Pacific '74 clones [44] employed similar fragment based strategies to overcome genomic stability problems, though a single plasmid DENV-2 clone has also seen considerable use ([50]–[52] for examples).
Blaney et al., using the DENV-3 clinical isolate Sleman '78, described the first, and, until now, only, DENV-3 infectious clone in 2004 [53]. Though based on a full-length cDNA plasmid, successful propagation of the plasmid DNA required inserting a 30 nt linker region containing termination sequences in each of the forward and reverse open reading frames near the E/NS1 junction. To date, the parent Sleman '78 clone has principally been used as a backbone for vaccine candidates [43], [54], [55].
Clearly, instability and toxicity have been the principle challenges of developing tractable DENV infectious clones. The smaller DENV cDNA sub-clone platform we employ offers several advantages. The individual fragments are highly stable in E coli and they can be manipulated individually without affecting distant sites on the genome and allow for fragment re-assortment between DENV strains. The type IIS restriction enzymes BglI and BsmbI generates unique 5′ and 3′ overhangs and prevents spurious self-assembly of the sub-clones, a technical problem with all palindromic cutting restriction enzymes [56]. Finally, multiple mutations can be incorporated simultaneously into separate fragments, circumventing iterative mutation and sequencing of the entire molecular clone and allowing for reassortment of fragments.
With the exception of the genotype I E gene, the parent molecular clone backbone was receptive to heterotypic E sequences. We suspected that genotype specific prM/M-E interactions between the genotype III parent prM/M and genotype I E accounted for failure to recover viable genotype I E chimera. Genotype I prM/M differs from genotype II, III, and IV in 2 positions. The first is a histidine to lysine mutation at prM/M position 55. This region is predicted to form a strand between two parallel beta sheets that interacts with the E fusion loop [47] and the polymorphism at position 55 likely explains why the original genotype I clone was not viable. The second difference was a leucine to phenylalanine mutation at position 128. This polymorphism conserves the hydrophobic character of the residue, hence we think it unlikely that this mutation affected the original genotype I chimera's viability. Replacing the parent prM gene with a genotype I prM established a viable clone and argues that future constructs should include prM and E from the same genotype. However, overall, the clone platform was remarkably stable: full length sequencing of passage three of all five of the clones found only one (silent) nucleotide mutation in one - Indonesia '82 - of the five clones. The recombinant viruses grew to equivalent peak titers compared to the parent clone, though Indonesia '82 showed delayed growth kinetics in both cell lines. Different plaque phenotypes emerged with E glycoprotein changes. While chimeric construction may affect interactions between E and the non-structural proteins or directly change RNA-RNA interactions, these effects are likely subtle, given the relatively similar clone growth kinetics in tissue culture, and are unlikely to directly affect chimeric clone neutralization by polyclonal sera.
Forty years ago Halstead and others first reported variable neutralization between clinical DENV-3 isolates [57] when they observed that mouse immune sera raised against DENV-3 strain H-87 poorly neutralized low passage wild-type DENV-3 isolates from Thailand. The authors hypothesized that the observed differences in neutralization were due to within serotype antigenic differences. Shortly thereafter, Russell et al. reported similar findings for human immune sera [58]. They found that both human convalescent sera and mouse hyper-immune sera against Tahitian and Caribbean DENV-3 poorly neutralized H-87 and a Thailand 1965 clinical isolate with differences in 50% hemagglutination inhibition (HI) titers varying by more that 10-fold. The authors argued that the different titers were evidence of genetic subtypes within DENV-3, at the time a novel idea, although the genetic basis for this variable phenotype was unclear. The authors also argued that Caribbean strains would be poor vaccine candidates because of their antigenic properties did not elicit broadly neutralizing homotypic antibodies. Despite this early observation of variable neutralization within serotypes, the phenomenon remained largely unexplored, in part because few tools existed to isolate antigenic variation in an otherwise stable genetic background.
More recently, Zulueta et al. [22], found that human sera from acute genotype III DENV-3 infections were essentially non-reactive with recombinant genotype IV EDIII but appropriately reactive with genotype III EDIII. However, this study's findings were significantly limited by the use of pooled acute human sera and binding assays, rather than neutralization assays and individual human polyclonal serum samples. In a related set of experiments, Cuban researchers tested convalescent sera collected from twenty DF and DHF cases from the 2001/2002 Cuban DENV-3 epidemic against a panel of six DENV-3 isolates collected between 2000 and 2002 [19]. The sera PRNT50 titers against clinical isolates from before and after that epidemic differed by nearly 10-fold, with the patients' sera more effectively neutralized virus from after the epidemic than before. However, their observed differences are based on neutralization against wild type viruses representing only genotypes III and IV and only three of the seven viruses used were sequenced. Finally, Thomas et al. [59], using previously characterized human DENV sera, found that PRNT50 titers were significantly affected by both virus strain and tissue in which the virus was propagated. While these experiments strongly hint at E gene dependent differences in polyclonal antibody neutralization, they do not directly test variability in the neutralization of isogenic DENV-3 viruses encoding clearly defined E gene differences by late convalescent sera.
Our results significantly advance both the pioneering early studies of Halstead and Russell as well as the more recent work cited above, all of which collectively argue that antigenic variability in DENV-3 genotypes significantly influences intra-serotypic neutralization responses in in vitro assays. With our panel of sera and E variant clones, we found both dramatically large, up to 19-fold, differences in FRNT50 values and FRNT50 titers as low as 1∶15 for homotypic sera (Text S1). Our data indicate that variation in E strongly drives these phenotypes, as all other viral proteins were isogenic.
Prospective studies of DENV transmission have found that low titer pre-existing neutralizing Ab (by PRNT) in endemic areas does not uniformly protect from homotypic infection [24], and a prospective study of maternal antibody in newborns found that 50% neutralization titers of <1∶50 are often not protective against homologous virus strains, even in endemic settings [60]. Finally, a recent human challenge study in DENV-3 vaccinated subjects found that a PRNT titer 1∶57 in one vaccinated volunteer was only partially protective, and another volunteer developed both fever and viremia with a pre-existing anti-DENV-3 titer of 1∶16 [61]. Current vaccine trials define 50% or 60% neutralization titers of >1∶10 [62], [63] or 1∶20 as evidence of immunity, potentially lower than the hypothesized protective thresholds suggested by the studies cited above. Some recent vaccine studies by Durbin et al. Guy et al. have begun to test vaccinee sera against representative genotypes [64], [65]. However, Durbin et al. used early convalescent sera - 42 days post vaccination, which is likely to be more broadly neutralizing is too early post-vaccination to capture the durable, long-term antibody response. Guy et al. similarly evaluated vaccinated vaccine sera against DENV genotypic variants, but used primate rather than human sera and the authors did not specify when the samples were collected post vaccination. The magnitude of the neutralization differences we report may be enough to lead to partial protection or loss of protection in vaccines, depending on the infecting genotype. It is also possible that, in the context of live virus vaccination, broad within serotype protection is conferred even with low titer antibodies, and that genotypic differences will not matter in the context of protection. That said, Genotype IV stands out in our experiments as relatively non-reactive with homotypic human immune sera (Figure 4B, 4E, 4F) and raises the question of whether vaccination could potentially create an immunologic “niche” in human hosts that could be exploited by sylvatic or geographically and genetically distant genotypes within a serotype.
Our findings serve as a point of departure for studying the important epitopes in the human antibody response to DENV infection, most of which have not yet been defined. Clones that selectively alter the antigenic clusters distributed across E (Text S1) will facilitate initial mapping of the epitopes responsible for differential neutralization. Ideally, identifying the key neutralizing epitopes in the human polyclonal immune response will, in turn, inform rational vaccine and possibly therapeutic monoclonal antibody design - optimizing epitopes to elicit potent neutralizing antibodies. Although only speculative, the DEN3 molecular clone may also prove invaluable for identifying epitopes and antibodies responsible for enhancing dengue infection.
|
10.1371/journal.pgen.1004425 | The POU Factor Ventral Veins Lacking/Drifter Directs the Timing of Metamorphosis through Ecdysteroid and Juvenile Hormone Signaling | Although endocrine changes are known to modulate the timing of major developmental transitions, the genetic mechanisms underlying these changes remain poorly understood. In insects, two developmental hormones, juvenile hormone (JH) and ecdysteroids, are coordinated with each other to induce developmental changes associated with metamorphosis. However, the regulation underlying the coordination of JH and ecdysteroid synthesis remains elusive. Here, we examined the function of a homolog of the vertebrate POU domain protein, Ventral veins lacking (Vvl)/Drifter, in regulating both of these hormonal pathways in the red flour beetle, Tribolium castaneum (Tenebrionidae). RNA interference-mediated silencing of vvl expression led to both precocious metamorphosis and inhibition of molting in the larva. Ectopic application of a JH analog on vvl knockdown larvae delayed the onset of metamorphosis and led to a prolonged larval stage, indicating that Vvl acts upstream of JH signaling. Accordingly, vvl knockdown also reduced the expression of a JH biosynthesis gene, JH acid methyltransferase 3 (jhamt3). In addition, ecdysone titer and the expression of the ecdysone response gene, hormone receptor 3 (HR3), were reduced in vvl knockdown larvae. The expression of the ecdysone biosynthesis gene phantom (phm) and spook (spo) were reduced in vvl knockdown larvae in the anterior and posterior halves, respectively, indicating that Vvl might influence ecdysone biosynthesis in both the prothoracic gland and additional endocrine sources. Injection of 20-hydroxyecdysone (20E) into vvl knockdown larvae could restore the expression of HR3 although molting was never restored. These findings suggest that Vvl coordinates both JH and ecdysteroid biosynthesis as well as molting behavior to influence molting and the timing of metamorphosis. Thus, in both vertebrates and insects, POU factors modulate the production of major neuroendocrine regulators during sexual maturation.
| Hormones play major roles in initiating major developmental transitions, such as puberty and metamorphosis. However, how organisms coordinate changes across multiple hormones remains unclear. In this study, we show that silencing the POU domain transcription factor Ventral veins lacking (Vvl)/Drifter in the red flour beetle Tribolium castaneum leads to precocious metamorphosis and an inability to molt. We show that Vvl regulates the biosynthesis and signaling of two key insect developmental hormones, juvenile hormone (JH) and ecdysteroids. Vvl therefore appears to act as a potential central regulator of developmental timing by influencing two major hormones. Because POU factors are known as a major regulator of the onset of puberty, POU factors play a major role during sexual maturation in both vertebrates and insects.
| Many organisms, including amphibians, echinoderms, marine invertebrates, vertebrates, and insects, undergo dramatic morphological and behavioral changes when they enter metamorphosis or puberty. In holometabolous insects, or insects that undergo complete metamorphosis, the larva molts several times before transforming into a pupa and ultimately into an adult. In mammals, puberty is also associated with morphological changes and reproductive maturation. These dramatic transformations are orchestrated by neuroendocrine changes that occur during postembryonic development of an organism. While the central nervous system (CNS) is known to regulate these endocrine changes [1]–[6], the link between the CNS and the endocrine centers remains poorly understood.
Members of the POU family have been shown to influence the neuroendocrine system during puberty and early development of vertebrates [7]–[13]. POU proteins have a highly conserved POU homeodomain and regulate gene expression by binding to high-affinity octamer sites [8], [14]. Many known POU factors are expressed in cell- or region-specific patterns within the developing CNS, suggesting a role in neural development. In addition, POU factors regulate the onset of puberty in mammals [13], [15].
Because POU domain transcription factors have been found in both vertebrates and invertebrates, their neuroendocrine functions may be conserved across species. Here, we investigated a homolog of the POU domain transcription factor, Ventral veins lacking (Vvl)/Drifter, in the holometabolous insect, Tribolium castaneum. Vvl has been shown to regulate the development of the tritocerebrum, CNS, peripheral nervous system and trachea in Drosophila embryos [16]–[22]. In addition, Vvl has been shown to regulate the expression of Diapause hormone-pheromone biosynthesis-activating neuropeptide (DH-PBAN) coding gene in the silkworm, Bombyx mori, and in the cotton bollworm, Helicoverpa armigera [23], [24]. More recently, it has been shown that in Drosophila, the endocrine glands, which synthesize metamorphic hormones, and trachea, share the same developmental origin, and that the progenitor cells that give rise to these structures express Vvl [25]. These studies suggest that Vvl, like vertebrate POU factors, may act within the neuroendocrine organs to regulate the biosynthesis of metamorphic hormones.
Many insects undergo three distinct phases of development: the larval, pupal and adult stages. The larval stage is primarily a feeding stage marked by several larval-larval molts (shedding of the cuticle) that enables the larva to grow to a sufficient size. Ecdysteroids and JH are two major developmental hormones involved in the transition from a larva to a pupa [3]. Ecdysteroidogenesis occurs in the prothoracic gland and involves the conversion of cholesterol into precursors of 20-hydroxyecdysone (20E), the primary ecdysteroid involved in molting. These precursors are ecdysone (E) in Drosophila and beetles, or 3-dehydroecdysone in Manduca sexta and a number of other lepidopteran species [26]–[29]. The secretion/synthesis of ecdysone is in turn triggered by prothoracicotropic hormone (PTTH), a neuropeptide released from the corpora cardiaca [3], [30], and by insulin signaling [31], [32]. Inside the prothoracic gland, these signaling pathways regulate the expression of several Halloween family genes, such as spook (spo), phantom (phm), disembodied (dib) and shadow (sad), which code for cytochrome P450 enzymes necessary for catalyzing a series of reactions that ultimately convert cholesterol into E [26], [33]–[35]. Mutations in these genes lead to lowered ecdysteroid titers in Drosophila [33]. The expressions of these genes are regulated dynamically by both PTTH and insulin signaling [34] and correlate with the ecdysteroid titer. Once released into the hemolymph, E is converted to 20E in the peripheral tissues by shade (shd) [36].
At the target tissues, ecdysteroids act by binding to the nuclear hormone receptor Ecdysone receptor (EcR), a heterodimer with the RXR homolog Ultraspiracle (Usp). Knockdown of EcR or Usp expression leads to disrupted molting and metamorphosis [37]–[40]. Once 20E binds to EcR, a series of ecdysone response genes are activated. Among the first genes to be activated are so-called primary response genes, which include E74 and E75 [41]–[44]. Subsequently, these early response genes activate the expression of delayed early genes, such as HR3 [45], [46]. Silencing the expression of these ecdysone response genes leads to disrupted molting and if silenced during the final instar, the larvae typically arrest their development during the prepupal period, indicating that metamorphosis is incomplete [47]–[49]. These studies show that ecdysteroid signaling is essential for molting and the completion of metamorphosis.
The nature of this ecdysteroid-induced molting depends on JH, a sesquiterpenoid hormone that is secreted from the corpora allata [50]. JH is known as a “status quo” hormone because it prevents progression to the next life stage after a molt [51]. In holometabolous insects, JH is present at high levels during the larval stages (instars) and prevents progression to the pupal stage during a molt. JH titers are known to be regulated by a complex interplay between JH synthesis, degradation and sequestration by JH binding proteins [52]. Only when JH levels drop in the last instar can ecdysone induce metamorphosis. That JH decline is essential for the initiation of metamorphosis has been illustrated through several distinct approaches. First, in many insects, application of JH during the penultimate instar induces a supernumerary molt [53]. [54]–[56]. Second, previous studies have shown that the removal of the JH-producing corpora allata can induce precocious metamorphosis with larvae exhibiting pupal characteristics before entering the final larval instar [57]. In addition, recent studies have demonstrated that the basic helix-loop-helix (bHLH)- Per-Arnt-Sim (PAS) domain protein encoded by Methoprene-tolerant (Met) plays a key role in mediating sensitivity to JH and likely acts as a JH receptor in Tribolium [58]–[60]. When Met is knocked down in Tribolium, larvae undergo precocious metamorphosis, a few larval molts earlier than normal [59], [61], [62], and develop into miniature adults. Finally, silencing of the JH biosynthesis gene, JH acid methyltransferase 3 (jhamt3), in Tribolium leads to precocious metamorphosis, again after fewer larval molts than normal, resulting in the formation of miniature adults [63]. The expression of this gene tracks the JH titer closely [63], and JHAMT is thought to act on the rate-limiting step for the series of biochemical reactions that ultimately results in the formation of active JH [64]. The regulation of JH biosynthesis is clearly an integral part of the regulation of timing of metamorphosis. However, how the expression of the key JH biosynthetic enzymes is regulated remains poorly understood. To summarize, the timing of metamorphosis is regulated by the dynamic titers of ecdysteroids and JH. Ecdysteroids are required for molting and their removal leads to disrupted molting and failure to initiate or complete metamorphosis. In contrast, JH is required to maintain the larval stage and its removal leads to precocious metamorphosis because in its absence, ecdysone induces metamorphosis even if the larva has undergone fewer larval-larval molts than in the wildtype.
In this study, we investigated the role of the POU domain transcription factor Vvl in coordinating the metamorphic hormones in Tribolium. Tribolium is well suited for the study of metamorphic regulation because of its sequenced genome and amenability to RNA interference (RNAi). Moreover, JH in Tribolium plays a prominent role in determining the number of molts prior to metamorphosis, facilitating the study of the role of developmental hormones in the regulation of metamorphic timing [59], [61], [63]. In contrast, Drosophila has a fixed number of instars, and topical application of JH does not lead to additional larval molts [65].
To determine the function of Vvl in Tribolium, we knocked down the expression of vvl and found that this resulted in precocious metamorphosis. These animals also had lowered expression of the JH response gene, krüppel-homolog 1 (kr-h1), which could be restored with topical application of JH. Knockdown of vvl also resulted in the reduced expression of jhamt3, a key regulator of JH biosynthesis. In addition, vvl knockdown led to an inability to molt and a corresponding reduction in ecdysone levels, and the expressions of ecdysone biosynthesis genes and the ecdysone-response gene, HR3. The expression of HR3 but not the molting defects could be rescued by injection of 20E. Our results suggest that Vvl may act as a nexus between JH and ecdysone biosynthesis.
We have identified one single ortholog of Vvl in the Tribolium Genome Base (http://www.beetlebase.org). Phylogenetic analysis of Tribolium POU factors confirms that the Tribolium Vvl clusters with both Drosophila and Bombyx Vvl homologs (Figure S1). To compare the Vvl ortholog from Tribolium to those of other invertebrates and vertebrates, the Tribolium Vvl protein sequence was identified in the Tribolium Genome Base and blasted in Geneious (http://www.geneious.com/). The POU region of the Tribolium Vvl consists of a POU domain and a homeodomain. The amino acid sequence of this region is highly conserved with those found in other species, such as Vvl in Drosophila melanogaster (98% sequence identity), POU-M2 in Bombyx mori (98% sequence identity), POU3F4 in Mus musculus (93% sequence identity), Xenopus laevis (93% sequence identity) and Homo sapiens (93% sequence identity) (Figure 1; Figure S2).
To determine the expression profile of vvl in Tribolium, qPCR was used to amplify vvl in cDNA obtained from the whole body mRNA extracts of the sixth and final instar larvae, and prepupae (Figure 2A). The expression of vvl was the highest on day 0 of the sixth instar and then dropped to a low level by day 3. The expression then increased to a high level at the time of the molt to the final instar. During the final instar, the vvl expression decreased gradually until the larva entered the prepupal stage when vvl expression increased again to a high level. The fluctuations observed are similar to the fluctuations of ecdysteroids and JH observed during the penultimate and final instars in other insects [51]. Thus, vvl expression might be correlated with endocrine signaling.
To further investigate tissue specific expression of vvl, vvl expression was analyzed in the CNS/corpora allata complex, epidermis, fat body and gut of day 0 seventh instar larvae using qPCR. The expression of vvl was the highest in the CNS/corpora allata complex (Figure 2B). Vvl was also expressed in the epidermis. Very low amounts of vvl mRNA were detected in the fat body and the gut. If vvl influences endocrine signaling, then one might expect that its removal would disrupt molting and/or the metamorphic transition.
To investigate the functions of Vvl during Tribolium development, vvl double-stranded RNA (dsRNA) was injected into day 0 fifth instar Tribolium larvae. All animals injected with vvl dsRNA initiated precocious metamorphosis and entered the prepupal stage without molting (n = 15; Figures 3 and 4A; Table 1). In contrast, control larvae injected with ampicillin-resistance (ampr) dsRNA at the beginning of the fifth instar molted at least two more times before initiating metamorphosis, typically at the end of the seventh instar stage (n = 16). After larvae were injected with ampr dsRNA, the larvae took approximately 12 days to enter the quiescent stage; in contrast, when animals were injected with 0.5 µg of vvl dsRNA, the timing of metamorphosis was shifted about four days earlier (Figure 4A). Because the animals injected with vvl did not molt to subsequent larval instars but rather initiated metamorphosis without molting (Table 1), the vvl dsRNA-injected prepupae were much smaller than the ampr dsRNA-injected prepupae (Figure 3A). The larvae injected with vvl dsRNA arrested at the prepupal stage, but the pupal characters, such as compound eyes and gin traps, eventually developed under the larval cuticle (Figures 3B–I). When these animals were sectioned, the old larval cuticle and the newly synthesized pupal cuticle appeared to be attached to each other, indicating that these animals fail to complete apolysis (Figure 3M). The vvl dsRNA-injected animals never developed into adults. To determine whether this effect was indeed due to the specific effect of vvl knockdown, dsRNA targeted to another region of the vvl gene was injected into fifth instar larvae. Similar precocious metamorphosis was observed (Figure S3), indicating that the precocious metamorphosis observed in this study is due to knockdown of vvl and not another gene.
The precocious metamorphosis seen in vvl dsRNA-injected larvae suggested the possibility that JH signaling might be affected. To determine how vvl dsRNA-injected larvae compare with larvae that have reduced expression of the JH receptor Met, Met dsRNA was injected into day 0 fifth instar larvae. Met dsRNA-injected larvae molted precociously, and a comparison of the mean time to metamorphosis showed that the mean time to metamorphosis in vvl and Met dsRNA-injected larvae was not significantly different from each other whereas they metamorphosed significantly earlier compared to ampr dsRNA-injected animals (p<0.005, ANOVA with Tukey HSD). However, in contrast to the vvl dsRNA injected animals, most larvae injected with Met dsRNA successfully completed a single molt before precociously entering metamorphosis (Table 1). Furthermore, the Met dsRNA-injected animals did not arrest their development at the prepupal stage (Figures 3J–L). Nearly 67% of Met dsRNA-injected animals (n = 21) began to develop adult tissues in the head and thoracic regions under the old larval cuticle and eventually eclosed as an abnormal adult, similar to the phenotypes reported by Parthasarathy et al (2008) ([61]; Table 1; Figures 3J–L). These observations suggest that vvl not only influences JH signaling but might also regulate the molting process.
Semi-quantitative RT-PCR was used to verify knockdown of vvl expression in early prepupae after last instar larvae were injected with 0.5 µg of dsRNA. In addition, semi-quantitative RT-PCR and qPCR were used to determine whether knockdown of either vvl or Met leads to changes in Met or expression, respectively. vvl mRNA expression level was lower in animals injected with vvl dsRNA than in the control larvae injected with ampr dsRNA and larvae injected with Met dsRNA (Figures 2C and 2D). In addition, Met expression level was lower in animals injected with Met dsRNA than in those injected with ampr dsRNA or vvl dsRNA (Figure 2C and 2D). These results demonstrate that the vvl and Met expression was effectively reduced and that Met and Vvl do not regulate the mRNA expression of each other.
The precocious metamorphosis suggested that JH biosynthesis might be impaired in vvl knockdown animals. A key JH biosynthesis enzyme is encoded by JHAMT, an enzyme that converts JH acid into JH. This step has been proposed to act as the rate-limiting step of JH biosynthesis [64], and its knockdown in Tribolium results in precocious metamorphosis [63]. The vvl expression profile described above follows closely the published expression profile of jhamt3 [63]. Thus, we examined whether the removal of Vvl affects the expression of jhamt3.
To determine whether Vvl plays a role in JH biosynthesis, qPCR was performed on day 4 fifth instar Tribolium larvae that were injected with either vvl or ampr dsRNA on day 0. The expression of jhamt3 was significantly decreased when the larvae were injected with vvl dsRNA (Figure 5A). In contrast, the expression of Met did not differ between vvl and ampr dsRNA-injected larvae (Figure 5B). Thus, vvl appears to play a crucial role in JH biosynthesis by influencing the expression of jhamt3.
If JH biosynthesis, and not JH sensitivity, is affected by vvl knockdown, topical application of methoprene should restore the normal timing of metamorphosis. We therefore ectopically applied the JH analog, methoprene, to day 0 fifth instar larvae injected with vvl dsRNA. Acetone treatments were used as controls, and ampr and Met dsRNA-injected larvae were also treated similarly for comparison.
All controls injected with ampr dsRNA and treated with acetone molted into pupae and formed normal adults (Table 2; Figures 6A and 6C). Application of 5 ng methoprene to day 0 fifth instar larvae injected with ampr dsRNA caused supernumerary molts (extra larval molts after the eighth instar) in seven out of 14 larvae (Table 2). Five of the methoprene-treated ampr dsRNA-injected animals failed to progress beyond the larval stage, but nine were able to undergo metamorphosis (Figure 6E); however, most larvae that underwent metamorphosis arrested their development and died as either prepupae or pupae (Table 2; Figure 5G). The time to metamorphosis was significantly delayed relative to those treated with acetone (Figure 4B; p<0.005, Student's t-test). When 15 µg of methoprene was applied to day 0 fifth instar larvae, the majority of the larvae stayed in the larval stage and did not pupate (n = 13/14).
In contrast, day 0 fifth instar larvae injected with vvl dsRNA and treated with 5 ng methoprene were still unable to molt (n = 16), and most died as prepupae (n = 10; Table 2). These prepupae ultimately developed pupal-like characteristics underneath the larval cuticle (Figures 6B and 6F). Ectopic application of methoprene delayed the timing of metamorphosis relative to those treated with acetone (Figure 4D; p<0.0001, Student's t-test). When the concentration of methoprene was increased to 15 µg, none of the larvae initiated metamorphosis and eventually died without ever molting (n = 9). The larvae survived on average for 23.9 days without molting.
As a comparison, day 0 fifth instar larvae were also injected with Met dsRNA and treated with acetone or 5 ng methoprene. For both treatments, the majority of Met dsRNA-injected animals underwent metamorphosis and developed into prepupae or eclosed as adults (Table 2; Figures 6A, 6D, 6E and 6H). In agreement with previous studies suggesting that Met is a receptor for JH, there was no significant difference between the timing of metamorphosis in larvae treated with acetone or 5 ng methoprene (Figure 4C; p = 0.61, Student's t-test), indicating that Met knockdown animals are insensitive to JH. Similarly, all Met knockdown animals treated with 15 µg of methoprene metamorphosed precociously, unlike those injected ampr or vvl dsRNA and treated with 15 µg of methoprene (n = 13/14). Taken together, the results support the notion that Vvl acts on JH biosynthesis but not JH reception to regulate the timing of metamorphosis.
To determine whether Vvl influences the expression of a downstream target gene of the JH pathway, qPCR analysis was performed on day 4 fifth instar Tribolium larvae that were previously injected with vvl dsRNA and treated with either acetone (control) or 15 µg methoprene on day 0 of the fifth instar. ampr dsRNA-injected animals treated similarly were used as a comparison. kr-h1 expression was reduced in the vvl dsRNA-injected larvae in comparison to control ampr dsRNA-injected larvae (Figure 5C). In larvae that were ectopically treated with 15 µg methoprene following injection of vvl dsRNA on day 0 of the fifth instar, the expression level of kr-h1 was restored to a level similar to that of the ampr dsRNA-injected animals treated with methoprene (Figure 5D). Thus, knockdown of vvl results in the downregulation of kr-h1 expression which can be rescued with the ectopic application of methoprene. These results further support the notion that JH biosynthesis rather than reception is influenced by Vvl.
The inability of vvl dsRNA-injected larvae to molt indicated a potential disruption of ecdysteroid signaling pathway, which is required for molting. To determine whether Vvl is required for the activation of ecdysone response genes that are associated with molting, we examined the expression of two such genes, E75 and HR3 [48], in vvl dsRNA- and ampr dsRNA-injected larvae four days after injection, just prior to their molt to the sixth instar in control larvae. The expression level of the ecdysone inducible early gene, E75, was not significantly reduced in vvl dsRNA-injected larvae (Figure 5E). However, the ecdysone-inducible gene, HR3, was dramatically reduced in vvl dsRNA-injected larvae (Figure 5F). We do not know why E75 did not decrease significantly in vvl knockdown animals. One reason might be that HR3, further downstream in the transcriptional cascade, is more sensitive than E75 to sustained Vvl-mediated changes in ecdysteroid secretion or action. Nevertheless, our results taken together suggest that ecdysteroid signaling is disrupted in vvl knockdown larvae. Thus, Vvl may have a dual role as both an activator of JH signaling and as a regulator of molting through its influence on ecdysteroid signaling.
To determine whether ecdysteroid titer was affected in response to vvl knockdown, ecdysteroid titer was observed on day 4 of the fifth instar after larvae were injected with either vvl dsRNA or ampr dsRNA on day 0 of the fifth instar. We observed that larvae injected with vvl dsRNA have a lower ecdysteroid titer relative to ampr dsRNA-injected larvae (Figure 7; p<0.01, Student's t-test). Together with the lack of molting phenotype and the lowered expression of ecdysone response genes, the lowered ecdysone titer indicates that Vvl influences ecdysteroid biosynthesis.
A recent study by Burns et al (2012) demonstrated that a GFP enhancer trap line that expresses GFP under the control of the vvl enhancer (strain KT817) expresses GFP in the oenocytes during the embryonic and early larval stages [66]. We examined the late sixth instar larvae and found GFP expression in the abdominal structures that most likely correspond to oenocytes (Figures 8A–8A″). Since oenocytes have been shown to be capable of synthesizing ecdysone from cholesterol in another tenebrionid beetle, Tenebrio molitor [67], [68], we wondered if the reduction of ecdysone titer might result from the reduced expression of vvl in the oenocytes.
To determine whether ecdysone biosynthesis was altered in the prothoracic gland or the oenocytes, larvae injected with vvl or ampr dsRNA were bisected and assayed for ecdysone biosynthesis gene expression. We found that the expression of spo was significantly reduced in the posterior half of the vvl knockdown larvae (Figure 9A). The expression of spo in the anterior portion of vvl knockdown larvae was not significantly different from that of ampr knockdown larvae (Figure 9A). To see if expression of spo was lowered in the oenoctyes of vvl knockdown larvae, day 0 fifth instar KT817 larvae were injected with vvl or ampr dsRNA and oenocytes along with the associated tissues were collected on day 4. We found that the expression of spo was lowered in the oenocytes of vvl knockdown larvae relative to the ampr knockdown larvae (Figure 9F). In contrast, we found that phm expression was reduced in the anterior portion of vvl dsRNA-injected animals relative to those injected with ampr dsRNA (Figure 9B). The posterior expression of phm remained unchanged. The expression of sad, shd and dib did not differ between the vvl and ampr dsRNA-injected animals in both the anterior and posterior portions of the body (Figures 9C–9E). These results suggest that Vvl regulates ecdysone biosynthesis in different glands by influencing the expression of different ecdysone biosynthesis genes.
We also looked for expression of GFP in the anterior portion of the larvae. In Coleoptera, prothoracic glands have been found to be associated with the dorsal tracheal trunk and the ventral tracheal trunk [69]. No obvious expression of GFP was observed in the cells lining the dorsal and ventral tracheal trunks except a few cells, which may correspond to Inka cells (Figures 8B and 8B′). The absence of GFP expression in the PG cells may be due to the KT817 enhancer trap not reflecting the entire repertoire of tissues where the vvl gene is active. To determine whether GFP is expressed in all the cells that express vvl, in situ hybridization was performed on day 1 embryos. In the anterior region, vvl mRNA was detected at the base of the maxillary and labial segments as well as the tracheal pits in the T2 and T3 segments (Figures 8D, 8D′, 8E and 8E′). These locations fail to express GFP in the KT817 strain (Figures 8C and 8C′). Similarly, in the abdomen, vvl mRNA was detected in the tracheal pits as well as the oenocytes even though no GFP expression was detected in the tracheal pits of the KT817 strain (Figures 8C, 8C″, 8D and 8D″). No staining was observed when the control sense probe was used (Figure 8F). These results suggest that there are additional enhancers that drive expression of vvl besides the one that drives GFP expression in the KT817 strain. The expression of vvl at the base of the labial and maxillary segments is noteworthy since in the Drosophila, vvl expressing cells in these segments give rise to the future endocrine glands [25]. While we attempted to stain for vvl in the prothoracic glands during the larval stages, we were unable to obtain convincing staining to due to excess background staining inherent to this stage.
To determine whether exogenous ecdysteroid agonist can rescue molting defects in vvl knockdown larvae, vvl dsRNA-injected larvae were treated with either 20E or RH-2845 two days after injection (Table 3). Injection of 1.5 µg or 0.15 µg of 20E led to death in vvl dsRNA-injected larvae (Table 3). When the concentration was reduced to 0.015 µg, a few larvae (n = 2/5) were able to survive but did not molt and developed into prepupae. ampr dsRNA-injected larvae also had high mortality although even at the highest concentration of 20E (1.5 µg), a few larvae survived to undergo a larval-larval molt. Thus, 20E does not appear to rescue the normal molting phenotype in vvl knockdown larvae. However, the high number of dead larvae relative to water injected controls (Table 3) suggests that vvl dsRNA injected larvae are able to sense 20E even though they cannot molt.
To determine whether 20E treatment can rescue HR3 expression, day 0 fifth instar larvae were injected with vvl dsRNA. Two days later, these larvae were injected with either 0.15 µg 20E or water. We found that larvae injected with 20E had significantly higher levels of HR3 expression (Figure 5G). Thus, exogenous 20E can rescue HR3 expression in vvl knockdown larvae. Our findings show that vvl knockdown leads to lowered ecdysteroid titer, which in turn leads to lowered HR3 expression. However, the failure of 20E to rescue the molting phenotype indicates that vvl knockdown impacts the molting process downstream of EcR/Usp.
Since injection of 20E just two days after dsRNA injection is potentially traumatic for the larvae, we decided to also investigate the effect of topical application of an ecdysteroid mimic. In Tribolium adults, 1 µg of RH-2845 upregulates the expression of ecdysone response genes, EcR and E75, indicating that this compound activates of ecdysteroid signaling [70]. ampr dsRNA-injected larvae molted after 1.6±0.4 days or 3.5±0.6 days when treated with either 0.1 µg or 1 µg of RH-2845, respectively (Table 3). In contrast, larvae injected with vvl dsRNA could not molt even when they were treated with either 0.1 µg or 1 µg of RH-2845 (Table 3). A high rate of mortality was also observed when these larvae were treated with RH-2845 (n = 4/8 and 3/6 for 0.1 µg and 1 µg of RH-2845, respectively). The remaining larvae all became prepupae without molting (n = 4/8 and 3/6 for 0.1 µg and 1 µg of RH-2845, respectively). While the mechanism by which RH-2485 acts as an ecdysteroid agonist remains unknown, the fact that RH-2485 is unable to induce a molt in the absence of Vvl supports the notion that Vvl is in part required for mediating the organism's molting behavior, in addition to its role in regulating ecdysone biosynthesis.
In this study, we determined the effects of knocking down vvl in Tribolium and elucidated a possible mechanism by which the gene interacts with the physiological regulation of development. Use of dsRNA-mediated expression knockdown revealed that suppression of vvl results in precocious metamorphosis. We also propose that Vvl acts upstream of JH signaling and that its expression is required for jhamt3 expression. In addition, we found that Vvl influences ecdysteroid biosynthesis and signaling to regulate molting.
In the present study, we found that fifth instar animals injected with vvl dsRNA underwent premature metamorphosis. The precocious metamorphosis observed in vvl dsRNA-injected larvae indicates that these larvae have disrupted JH signaling which results in precocious pupal commitment (Figure 10). Consistent with this interpretation, we found that knockdown of vvl expression in the fifth instars causes a down-regulation of the expression of the JH-response gene kr-h1 [71]. In addition, we found that vvl knockdown animals were not resistant to methoprene as seen in Met knockdown animals. The delay in metamorphosis observed in these methoprene-treated vvl knockdown larvae suggests that vvl does not affect the sensitivity to JH. Similarly, ectopic application of methoprene was able to upregulate kr-h1 expression in vvl knockdown animals. Because kr-h1 is a direct downstream target of Met [58], [71], the down-regulation of kr-h1 observed in vvl knockdown animals and the rescue of its expression with ectopically applied methoprene indicate that vvl knockdown causes a decrease in JH biosynthesis by acting upstream of kr-h1. Because the expression of jhamt3, a JH biosynthesis enzyme, is decreased in vvl RNAi animals, Vvl likely influences the rate of JH biosynthesis by influencing the transcription of jhamt3 directly or by regulating an upstream regulator of jhamt3 expression. In agreement with this view, vvl and jhamt3 whole body expression profiles are similar (this study; [63]). Our study suggests that this transcription factor plays a key role in regulating the timing of JH biosynthesis. In support of this, we have detected high levels of vvl mRNA in day 0 seventh instar CNS/corpora allata complex. While we were unable to distinguish between the expression of vvl in the CNS and that in the corpora allata, our finding suggests an intriguing possibility that Vvl might function as a mediator between the nervous system and JH production, potentially linking the two systems to coordinate the onset of metamorphosis.
All larvae injected with vvl dsRNA underwent precocious metamorphosis without molting, while Met dsRNA-injected animals almost always molted before metamorphosis (Table 1). In addition, although jhamt3 dsRNA-injected larvae undergo precocious metamorphosis as seen in vvl knockdown larvae, jhamt3 knockdown animals do not show any molting defects and eclose into miniature but externally complete adults [63]. Thus, Vvl suppression might have an additional inhibitory effect on the ecdysteroid signaling pathway, a key regulator of molting. Consistent with this notion, we found that ecdysone titer and the expression of the ecdysone response gene HR3 were reduced in vvl knockdown larvae. This lowered level of HR3 could be rescued by injection of 20E, further supporting the notion that the vvl is required for ecdysteroid biosynthesis. In Bombyx, a homolog of Vvl, BmPOUM2, has been shown to be required for metamorphosis. However, when the expression of this gene was knocked down during the wandering stage of Bombyx, the expression of ecdysone response genes was not altered [72]. This difference may be due to a difference in the species used or because Vvl removal during wandering is too late to detect an effect on ecdysone biosynthesis or response genes.
We also observed that the expression of the ecdysone biosynthesis genes phm and spo were lowered in the anterior and posterior halves of the vvl knockdown larvae, respectively (Figure 9). The latter observation is consistent with the expression of GFP in the oenocytes of the KT817 GFP enhancer trap line, which expresses GFP under the control of a vvl enhancer [66]. In several insect species, including another tenebrionid beetle, Tenebrio molitor, oenocytes are capable of producing ecdysone from cholesterol [67]. We have found that when vvl is knocked down, spo expression decreases. Thus, our observations support the notion that Vvl in the oenocytes regulates ecdysteroid production. However, it is possible that Vvl might target additional ecdysteroidogenic tissues besides oenocytes [68], [73].
We were unable to detect GFP expression in the cells lining the ventral and dorsal tracheal trunk where prothoracic gland cells are typically found [69]. The comparison between embryonic vvl expression and embryonic GFP expression in the KT817 line suggests that vvl expression in the prothoracic glands and oenocytes are regulated by distinct enhancers. The expression of vvl in the labial and maxillary segment of the embryo is similar to what has been observed in Drosophila. In Drosophila, these vvl-expressing cells give rise to the prothoracic glands and the corpora allata. Thus, we suggest that the vvl expressing cells in the labial and maxillary segments likely also contribute to the endocrine organs in Tribolium.
While we cannot be certain about the mechanism by which ecdysone biosynthesis gene expression is modulated in the different glands, our findings suggest that ecdysone biosynthesis genes are regulated differently in the various endocrine structures. This is not surprising considering that in adults, ecdysone biosynthesis genes are activated differently in the accessory glands and the ovaries [35]. The modular nature of gland specific regulation of hormonal production is probably very common. Future studies should investigate the relative contributions of ecdysteroids from the prothoracic gland and other endocrine sources during larval development.
In addition, the inability of exogenous 20E and RH-2485 to rescue the molting defects despite being able to rescue HR3 expression suggests that Vvl might affect a component of the ecdysis pathway, such as ecdysis triggering hormone or eclosion hormone [74]–[77]. Our observation of vvl-GFP expression in the Inka cell-like cells in the KT817 line is also consistent with this notion since Inka cells have been shown to produce ecdysis triggering hormone in a several insects [77], [78]. The expression of ecdysis triggering hormone is influenced by circulating levels of ecdysteroids and EcR [76], [79], [80]. Preliminary findings show that Vvl binds to EcR/Usp through a GST-pulldown assay (Figure S4). This suggests a possibility that Usp and EcR heterodimerize with Vvl to regulate the expression of ecdysis triggering hormone (Figure S4). However, additional studies, such as ChIP assays, are necessary to definitively determine how Vvl modulates target genes.
Our findings show that Vvl regulates both ecdysone and JH signaling, the two hormonal signaling pathways that regulate metamorphosis (Figure 10). Thus, Vvl may couple the two hormonal pathways to coordinate the timing of metamorphosis in insects. Specifically, we propose that a switch in the expression of Vvl signals the initiation of metamorphosis by regulating the dynamic fluctuations of both of these developmental hormones. Whether Vvl regulates the two hormones separately or relays a change from one to another remains unclear at present.
It has been suggested that puberty in humans and other mammals is another form of metamorphosis. Although seemingly different, both processes are marked by dramatic changes in physiology. Puberty in humans is initiated by endocrine changes in response to increased pulsatile release of gonadotropin releasing hormone (GnRH) from the hypothalamus. In vertebrates, POU factors have been shown to regulate neuropeptide expression. In particular, GnRH expression is regulated by POU factors, which bind to the promoter and enhancer sequences of the GnRH gene. For example, the POU domain transcription factor Oct-6 (POU3F1) represses GnRH activity, delaying the onset of puberty [10]. In the absence of Oct-6, GnRH is expressed, triggering pubertal development.
Our study shows an intriguing similarity between the transcriptional regulation of the key neuroendocrine regulators of insect metamorphosis and mammalian puberty. Most certainly, these two processes have evolved independently and are not homologous. However, given the substantial sequence conservation of POU domain proteins across various metazoan taxa, future studies should investigate whether POU factors play a role in reproductive maturation in other metazoan taxa. Since the corpora allata and the prothoracic gland along with the trachea all express vvl during Drosophila embryogenesis, it has been suggested that these Vvl-specified structures have a common origin in a proto-arthropod [25], [81]. In light of the endocrine functions we identified in our study, we suggest that vvl had an ancient endocrine function, and its expression in these endocrine structures was subsequently retained as corpora allata and prothoracic glands evolved.
Wildtype Tribolium strain GA-1 was obtained from Dr. Richard Beeman (USDA ARS Biological Research Unit, Grain Marketing & Production Research Center, Manhattan, Kansas). The KT817 GFP enhancer trap line was obtained from the Kansas State University Tribolium Stock Center [82]. All beetles were raised on organic whole-wheat flour fortified with 5% nutritional yeast at 29°C and 50% humidity.
Tribolium RNA from various larval instars was isolated by homogenizing the tissues in TRIzol (Life Technologies, Carlsbad, CA). The RNA sample was treated with DNAse and converted to complementary DNA (cDNA) using reverse transcriptase (Thermo Scientific, Waltham, MA). The vvl gene sequence was obtained from NCBI (NM_001145913). Primers were designed to amplify particular regions of interest in the synthesized cDNA using polymerase chain reaction (PCR) (Table 4).
The PCR product was subsequently cloned into the TOPO-TA cloning vector (Life Technologies, Carlsbad, CA), and cells were transformed with this plasmid. The purified plasmid DNA was sequenced to confirm the identity of the cloned gene and the accuracy of the insertion into the TOPO vector. The plasmid DNA was linearized using Spe1 and Not1 restriction enzymes (New England Biolabs, Ipswich, MA) and subsequently used for single-stranded RNA (ssRNA) synthesis. ssRNA was synthesized using MEGAscript T3 and T7 kits (Life Technologies, Carlsbad, CA), according to the manufacturer's instructions. The complementary ssRNA were then combined and annealed to form a 2 µg/µl dsRNA solution using a standard annealing protocol [83]. The final annealed product was analyzed via gel electrophoresis to ensure proper annealing.
To characterize the role of vvl, day 0 fifth instar Tribolium larvae were injected with dsRNA. Using a pulled 10 µl glass capillary needle connected to a syringe, dsRNA was manually injected into each animal. dsRNA was injected until the abdomen began to stretch (approximately 0.25 µl). Controls were injected with the similar volume of bacterial ampr dsRNA. Animals were also injected with Met dsRNA as a comparison to the vvl knockdown animals.
To determine the effect of JH on vvl knockdown animals, a subset of the injected larvae was also topically treated with 0.5 µl JH analog methoprene (5 ng or 15 µg) (Sigma-Aldrich, St. Louis, MO) dissolved in acetone. The same amount of acetone was applied to control larvae. All solutions were applied to the dorsal side of the animals immediately following injection with dsRNA (day 0 fifth instar) to mimic a constant level of JH. All experimentally treated animals were maintained at 29°C and 50% humidity. The animals were examined every other day and characterized in comparison to ampr dsRNA-injected animals treated with acetone. Initiation of metamorphosis was identified by observation of the characteristic J-shape (“J-hangers”) of the larva, immobile legs and the position of the larval eyes, which begin to migrate towards the posterior of the head segment.
To determine the effect of vvl knockdown on ecdysone sensitivity, larvae injected with either vvl or ampr dsRNA as day 0 fifth instar larvae were treated with either 20E or an ecdysteroid mimic, RH-2485 methoxyfenozide (Sigma-Aldrich, St. Louis, MO) on day 2 of the fifth instar. For 20E, larvae were injected with 0.3 µl of 5 µg/µl, 0.5 µg/µl or 0.05 µg/µl 20E dissolved in water. Water injection was used as a control. For RH-2485, 0.5 µl of 0.2 or 2 µg/µl RH-2485 in acetone was topically applied to the larva. The latter amount of RH-2485 has previously been shown to activate ecdysone response genes in adult Tribolium [70]. Acetone was used as a control. Treated larvae were checked daily for signs of molting, prepupal development or death.
To determine the effect of vvl knockdown on ecdysone titer, day 0 fifth instar larvae were injected with ampr or vvl dsRNA. On day 4, six larvae were collected and frozen at −80°C. Pooled larvae were homogenized in 100% methanol, centrifuged, and the supernatant was dried, then resuspended in 35 µl Grace's media. Triplicate 10 µl samples were assayed for ecdysteroids, and averaged. The radioimmunoassay was conducted as previously described using an antibody generously provided by Dr. Lawrence Gilbert that cross-reacted equivalently with ecdysone and 20-hydroxyecdysone [84]. Ecdysone was used as a standard.
To determine the developmental expression profile of vvl, RNA was isolated from the whole body of sixth and seventh instar GA-1 strain larvae, and prepupae. Total RNA was isolated from the whole body of day 4 fifth instar larvae using Trizol extraction and treated with DNAse to remove genomic DNA. cDNA was synthesized from 1 µg of RNA as described above. Each biological sample consisted of pooled RNA from five fifth instar larvae or three seventh instar larvae or prepupae. Three biological replicates of each treatment were prepared, and SsoAdvanced SYBR Green Supermix (Bio-Rad Laboratories, Hercules, CA) was used for qPCR analyses.
To assess the effect of vvl on the expression of E75, HR3, kr-h1, Met and jhamt3, day 0 fifth instar GA-1 strain larvae were injected with vvl dsRNA or ampr dsRNA, and their RNA was isolated. A subset of larvae was also treated with 0.5 µl of methoprene (30 µg/µl) or acetone. The effects of vvl or Met knockdown on Met and vvl expression, respectively, were also assayed during the prepupal stage. For this, two animals injected with either vvl, Met or amprdsRNA were pooled per biological sample, and three biological replicates were created.
The effect of vvl on the expression of ecdysone biosynthesis genes spo, phm, shd, dib and sad, was assayed by injecting day 0 fifth instar larvae with vvl dsRNA or ampr dsRNA, and splitting the larvae into the anterior (containing the thorax) and posterior (containing the abdomen) halves. RNA was pooled from five animals per biological sample, and three biological samples were prepared as described above. In addition, to determine whether vvl knockdown influences spo expression in the oenocytes, day 0 fifth instar KT817 strain larvae were injected with either vvl or ampr dsRNA. Oenocytes were visualized by GFP and isolated from day 4 fifth instar larvae. RNA was pooled from 15 larvae per treatment.
To determine the tissue specific expression of vvl, CNS with the associated structures including the CA, the fat body, the gut and the epidermis were isolated from 20 day 0 seventh instar GA-1 strain larvae. The tissues were pooled and RNA was extracted as described above. cDNA was created from 1 µg RNA.
To determine the effect of 20E on HR3 expression in vvl dsRNA-injected animals, day 0 fifth instar GA-1 strain larvae were injected with vvl dsRNA. Two days later, larvae were injected with either 0.3 µl of water as a control or 0.3 µl of 0.5 µg/µl 20E. After 6 hrs, larvae were harvested for RNA isolation and subsequent cDNA synthesis as described above. Three larvae were pooled per biological replicate and three biological replicates were created. Each sample was run in triplicate with no-template controls. rp49 was used as an internal control for all qPCRs.
In order to perform semi-quantitative RT-PCR, early prepupae were collected. Last instars injected with ampr, vvl, or Met dsRNA were collected as they prepupated. Three animals for each treatment were pooled in Trizol and prepared for total RNA isolation, and cDNA was synthesized from 1 µg of RNA as described above. PCR products were run on a gel and visualized using a UV-illuminator.
To create the probes, vvl gene was cloned into a TOPO-TA cloning vector and restriction digested as described above. The purified linearized plasmid DNA was used to create probes as described previously [85]. To examine the expression of vvl in GA-1 strain embryos, day 1 eggs were collected and dechorionated in 25% bleach. After heptane/formaldehyde fixation, eggs were cracked in heptane/methanol. Embryos were individually dissected out of the egg shell. Standard in situ hybridization protocol was followed. The sense probe was used as a control. To visualize GFP expression in KT817 embryos, eggs were dechorionated and fixed as above. Embryos were manually dissected out and mounted in 70% glycerol in PBS.
|
10.1371/journal.pntd.0003194 | Cathepsin B in Antigen-Presenting Cells Controls Mediators of the Th1 Immune Response during Leishmania major Infection | Resistance and susceptibility to Leishmania major infection in the murine model is determined by the capacity of the host to mount either a protective Th1 response or a Th2 response associated with disease progression. Previous reports involving the use of cysteine cathepsin inhibitors indicated that cathepsins B (Ctsb) and L (Ctsl) play important roles in Th1/Th2 polarization during L. major infection in both susceptible and resistant mouse strains. Although it was hypothesized that these effects are a consequence of differential patterns of antigen processing, the mechanisms underlying these differences were not further investigated. Given the pivotal roles that dendritic cells and macrophages play during Leishmania infection, we generated bone-marrow derived dendritic cells (BMDC) and macrophages (BMM) from Ctsb−/− and Ctsl−/− mice, and studied the effects of Ctsb and Ctsl deficiency on the survival of L. major in infected cells. Furthermore, the signals used by dendritic cells to instruct Th cell polarization were addressed: the expression of MHC class II and co-stimulatory molecules, and cytokine production. We found that Ctsb−/− BMDC express higher levels of MHC class II molecules than wild-type (WT) and Ctsl−/− BMDC, while there were no significant differences in the expression of co-stimulatory molecules between cathepsin-deficient and WT cells. Moreover, both BMDC and BMM from Ctsb−/− mice significantly up-regulated the levels of interleukin 12 (IL-12) expression, a key Th1-inducing cytokine. These findings indicate that Ctsb−/− BMDC display more pro-Th1 properties than their WT and Ctsl−/− counterparts, and therefore suggest that Ctsb down-regulates the Th1 response to L. major. Moreover, they propose a novel role for Ctsb as a regulator of cytokine expression.
| The emergence of resistance to the available drugs against cutaneous leishmaniasis emphasizes the need of new chemotherapeutic approaches. Cysteine proteases from Leishmania are important virulence factors and, therefore, interesting drug targets. Studies on inhibitors against these enzymes during Leishmania major infection in mice had shown that host equivalents of these proteases are also affected, namely cathepsin B and cathepsin L. The inhibition of cathepsin B resulted in immune-mediated protection, while inhibition of cathepsin L caused susceptibility to the parasite. In the present study, we investigated the effect of cathepsin deficiency on the signals used by dendritic cells to orchestrate the T helper (Th)-mediated immune response against L. major and the control of parasite proliferation within infected macrophages. The results demonstrate that cathepsin B-deficient dendritic cells express higher levels of the antigen-presenting MHC class II molecules than WT and cathepsin L-deficient cells. Surprisingly, dendritic cells and macrophages deficient for cathepsin B showed higher expression of the protective Th1-inducing cytokine IL-12. Therefore, we propose a novel role of this protease as a regulator of cytokine expression. Altogether, these findings suggest that cathepsin B down-regulates the Th1 response to L. major, and, in its absence, antigen-presenting cells express signals protecting against the parasite.
| Worldwide, 2 million new cases of leishmaniasis occur every year. It is endemic in 98 countries, where 350 million people are considered to be at risk, largely affecting “the poorest of the poor” [1], [2]. The cutaneous form of leishmaniasis is characterized by lesions that heal over the course of months or years, and leave permanent scars that can be disfiguring or disabling [1], [2]. Control of Leishmania within the host is mediated by innate and adaptive immune responses. Experimental mouse models of Leishmania major infection first documented the relevance of Th1/Th2 polarization for resistance and susceptibility to the disease in vivo [3]. In resistant mouse strains, such as C57BL/6, a contained and self-healing development of the disease appears, mediated by a protective Th1 immune response. Th1 cells secrete IFN-γ, which induces the nitric oxide (NO)-mediated killing of the amastigote form of the parasite within phagosomes in macrophages. In contrast, infection of BALB/c mice with L. major causes a non-healing Th2 form of the disease, characterized by expression of the cytokines IL-4, IL-13, and IL-10.
The key role of dendritic cells (DC) in inducing cell-mediated immune responses against leishmaniasis has been extensively documented [4], based on their capacity to migrate to draining lymph nodes after capture of Leishmania parasites and to induce Th cell polarization. Several subsets of DC have been reported to perform this function, including Langerhans cells [5], dermal DC [6], lymph node resident DC [7], and monocyte-derived DC [8], [9]. In order to instruct Th cell polarization, DC use three main signals: (1) antigen presentation via MHC class II molecules, (2) the expression of co-stimulatory molecules and (3) cytokine secretion. Quantitative and qualitative differences in these signals are crucial for Th cell polarization [10]. Among these signals, IL-12 is a key cytokine for the development of a protective Th1 immune response. Neutralization of IL-12 by antibodies leads to susceptibility to Leishmania infection in otherwise resistant mice [11], [12]. Conversely, treatment of BALB/c mice with IL-12 resulted in a protective Th1 response [13]. DC have been reported to be the primary source of IL-12 in lymphoid tissues [14], with variations depending on the DC subset, maturation status, and whether promastigotes or amastigotes are used [15]. Macrophages, on the other hand, are considered as main host cells for Leishmania parasites, where freshly inoculated promastigotes find a niche for differentiating into amastigotes and proliferating. Macrophages are not able at all to produce IL-12 in response to L. major [16], [17], even if they are further stimulated with lipopolysaccharide (LPS) [18], [19], reflecting the extent of the silencing that L. major induces in its host.
Silencing of infected cells has been attributed to different virulence factors. Some of them are cysteine proteases [20], which impair NF-κB signaling in macrophages [21] and are also important for autophagic and differentiation processes in the parasite [22]. Therefore, they are interesting targets for drug development [23], [24]. However, they also have homologs in mammals. Few studies have addressed the effects that unspecific inhibition of host cathepsins would have on the immune response against L. major. Maekawa et al. reported that treatment of L. major-infected BALB/c mice with the Ctsb inhibitor CA074 triggered a protective Th1 immune response [25], [26]. Treatment of these mice with the Ctsl inhibitor CLIK148, on the other hand, caused a stronger Th2 response [27], even in resistant mouse strains [28]. The authors showed that the inhibitors had no direct effect on the proliferation of the parasite but that the host cell cathepsins were inhibited, and hypothesized that lack of Ctsb or Ctsl would lead to different patterns for proteolytic processing of L. major antigens. It had remained unclear, however, how the inhibition of Ctsb and Ctsl activity could have such effects in Th polarization. Thus, further investigation is needed to understand the involvement of Ctsb and Ctsl in the immune response during leishmaniasis.
In the present study, we used cathepsin B (Ctsb−/−)- and cathepsin L (Ctsl−/−)-deficient bone marrow-derived DC (BMDC) and bone marrow-derived macrophages (BMM) to determine the role of these proteases for the signals that DC use to instruct Th cell polarization in response to L. major promastigotes: the expression of MHC class II and co-stimulatory molecules, and the production of cytokines. Furthermore, we studied the effect of these cathepsins on parasite proliferation in infected cells. Our results indicate that in this model of infection, cathepsin B plays a significant role not only in the expression of antigen-presenting MHC class II molecules, but also in the regulation of IL-12 production.
Complete RPMI medium was prepared by supplementing RPMI 1640 medium (with phenol red or phenol red-free, as indicated; Invitrogen, Darmstadt, Germany) with heat-inactivated fetal calf serum (FCS, 10% v/v; PAA Laboratories, Pasching, Austria), L-glutamine (final concentration 2 mM; Biochrom AG, Berlin, Germany), HEPES (pH 7.2, 0.01 M; Invitrogen, Darmstadt, Germany), penicillin G (0.2 U/ml; Sigma-Aldrich, Taufkirchen, Germany), gentamicin (0.05 mg/ml; Sigma-Aldrich), and 2-mercaptoethanol (0.05 mM; Sigma-Aldrich). In addition, for generation of BMM, a conditioned medium was used containing Dulbecco's Modified Eagles Medium (DMEM; Invitrogen), heat-inactivated FCS (10% v/v; PAA Laboratories), heat-inactivated horse serum (0.5%; Invitrogen), 2-mercaptoethanol (0.05 mM; Sigma-Aldrich), nonessential amino acids (Invitrogen), HEPES (0.01 M; Invitrogen), L-glutamine (4 mM; Biochrom) and L929 supernatant (15% v/v). L. major promastigotes were cultured in a biphasic medium consisting of a solid base of rabbit-blood agar (Elocin-lab, Gladbeck, Germany) plus a liquid phase of RPMI medium without phenol red.
BMDC and BMM were generated from bone marrow progenitors as described previously [23], [29], [30] from female BALB/c, C57BL/6, C57BL/6 Ctsb−/− and C57BL/6 Ctsl−/− mice (6–12 weeks old). The generation of Ctsb−/− and Ctsl−/− mice has been described previously [31]–[33]. Briefly, total bone marrow cells were flushed from femurs and tibiae. To generate DC, the cell number was determined by trypan blue staining, and 0.2×106 cells/ml bone marrow cells were cultured in complete RPMI 1640 medium in the presence of recombinant murine granulocyte-macrophage colony-stimulating factor (GM-CSF, 0.04 µg/ml; Invitrogen) at 37°C, 5% CO2. Cultures were fed with complete RPMI medium supplemented with GM-CSF on days 3 and 6. At day 8, the non-adherent cells were collected, washed with complete RPMI medium and resuspended at 2×106 cells/ml in complete RPMI medium. BMM were generated by culturing 0.67×106 cells/ml of total bone marrow progenitors as described above in conditioned DMEM at 37°C and 5% CO2. On day 6, the culture medium was removed carefully and replaced with cold RPMI complete medium, and the petri dishes were kept on ice for 10 min. Thereafter, the macrophages were removed with a cell scrapper, washed with fresh complete medium without phenol red, and resuspended at 2×106 cells/ml.
As quality control, the morphology of the obtained BMDC and BMM was analyzed. Part of the cells was used for cytospin preparations stained with Diff-Quik II dye (Medion Diagnostics, Düdingen, Switzerland) according to the manufacturer's instructions, and observed under the light microscope. Furthermore, the expression of the phenotypic markers CD11c in DC and F4/80 in macrophages was assessed by flow cytometry as described below.
The morphology of BMDC and BMM from WT, Ctsb−/− and Ctsl−/− mice was additionally analyzed by TEM. Samples of the obtained cells were prepared for TEM using OsO4 and uranyl acetate as contrasting agents, following the protocol previously described by Schurigt et al. [24].
The L. major isolate MHOM/IL/81/FE/BNI was maintained by continuous passage in BALB/c mice, and promastigotes were grown in vitro in blood-agar cultures as described previously [34] at 27°C, 5% CO2 and 95% humidity. In order to preserve maximal infectivity, only promastigotes passaged 5 to 8 times were used for in vitro infection experiments. In addition, two different transgenic L. major strains were used in some experiments: a luciferase-transgenic strain (Luc-tg) previously described [30], and an eGFP-transgenic strain (eGFP-tg). For the preparation of L. major antigen (LmAg), stationary-phase WT promastigotes were washed three times in phosphate-buffered saline (PBS), resuspended at 1×109/ml in PBS, and subjected to three cycles of freezing in liquid nitrogen and thawing. The aliquots were stored at −80°C and each aliquot was thawed not more than twice. Heat-killed parasites (HK) were prepared by incubating a parasite suspension of 1×109/ml in RPMI medium for 30 min at 80°C.
The enhanced-green fluorescent protein (eGFP)-coding region was cut from pEGFP-N1 (Clontech, Saint-Germain-en-Laye, France) by BamHI-NotI (Promega, Mannheim, Germany) and cloned into the Bglll-NotI-restricted Leishmania expression vector pLEXSY-hyg2 (Jena Bioscience, Jena, Germany), which contains a marker gene for selection with hygromycin. The generated plasmids were linearized by SwaI (New England Biolabs, Frankfurt, Germany), and the parasites were transfected by electroporation. eGFP and HYG were integrated into the 18S rRNA locus of L. major by homologous recombination. For in vitro experiments, promastigotes were grown in blood-agar cultures supplemented with 50 µg/ml hygromycin under the same conditions as WT and Luc-tg L. major promastigotes. In order to maintain their virulence, eGFP-tg L. major parasites were passaged in female BALB/c mice. The stability of the integrated eGFP without further selection by hygromycin was assessed in vitro and in vivo by flow cytometry.
Intracellular amastigotes in BMM from Ctsb−/− and Ctsl−/− mice, and their WT C57BL/6 counterparts, were measured with the method described by Bringmann et al. [30]. Briefly, 200 µl of a 2×105 cells/ml suspension of BMM in phenol red-free complete medium were seeded into 96-well plates with clear bottoms (Greiner Bio-One, Frickenhausen, Germany) and were incubated for 4 hours to allow cell adhesion. The medium was then removed, and 200 µl of a 3×106 cells/ml suspension of Luc-tg L. major promastigotes were added at an infection ratio of 1∶15 and incubated for 24 hours at 37°C, 5% CO2. Any remaining extracellular parasites were eliminated by washing 3 times with medium, and 200 µl of the phenol red medium were added. After further 24 hours incubation at 37°C, 5% CO2, 50 µl of the luciferin-containing lysis buffer Britelite Plus (PerkinElmer, Waltham, USA) were added to each well. The plate was incubated in the dark for 5 min at room temperature (RT), and the resulting luminescence was measured as counts per second (CPS), with a Victor×Light 2030 luminometer (PerkinElmer).
5×105 BMM from WT, Ctsb−/− and Ctsl−/− mice were seeded in duplicates into chambered cover glasses, in a final volume of 250 µl of complete medium without phenol red, and incubated at 37°C, 5% CO2 for 4 hours to promote cell adhesion. The culture medium was removed, replaced by an equivalent volume of eGFP-tg L. major promastigotes at an infection ratio of 1∶15, and the cells were further incubated for 24 hours at 37°C, 5% CO2. The cells were then washed 3 times with warm PBS; part of the cells were incubated with Hoechst solution 0.5% v/v (Immunochemistry Technologies, Bloomington, USA) for 15 min at 37°C protected from the light, followed by washing 3 times with warm PBS and addition of 250 µl of complete medium. Then, they were observed under a fluorescence microscope (Leica Microsystems). The rest of the cells were incubated in fresh medium for further 24 hours, stained and observed under the fluorescence microscope as described above. The amount of cells and L. major bodies were quantified with the Cell Counter plug-in from the ImageJ software [35], and the average parasite count per infected cell was calculated as a geometric mean (G) using the formula , where represents the sequence of parasites counted for every infected cell.
1×106 BMDC/ml were harvested at day 7 of culture, plated in 6-well plates and incubated overnight at 37°C, 5% CO2. For some experiments, BALB/c BMDC were pre-incubated with 10 µM CA074Me (Bachem, Bubendorf, Switzerland), 10 µM CLIK148 (kindly provided by Prof. Tanja Schirmeister), or 10 µM of Z-Arg-Leu-Arg-α-aza-glycyl-Ile-Val-OMe (ZRLR, kindly provided by Dr. Timo Burster, University of Ulm, and Dr. Ewa Wieczerzak, University of Gdansk) for 4 hours prior to infection. eGFP-L. major promastigotes were harvested, washed 3 times in warm PBS, added to the BMDC at a 1∶5 infection ratio, and further incubated at 37°C. After 2 hours of exposure of the BMDC to the parasites, the cells were washed with warm PBS and resuspended in fresh medium at a concentration of 1×106 cells/ml. Part of the cells was fixed in paraformaldehyde (PFA, 4%; Applichem, Darmstadt, Germany). The remaining cells were incubated for a total of 4 hours or 24 hours post infection, fixed, and the amount of infected cells at the different time points was determined by flow cytometry, together with the expression of maturation markers as described next.
BMDC infected with e-GFP L. major or stimulated with LmAg (30 µl LmAg/ml, equivalent to 30 parasites per BMDC) were fixed with 4% PFA and resuspended in FACS buffer containing the following antibodies (Ab): phycoerythrin-cyanine 7 (PECy7)–conjugated anti-CD11c (BD Biosciences, Heidelberg, Germany), phycoerythrin (PE)-conjugated anti-CD86 (BD Biosciences), allophycocyanin (APC)-conjugated anti-MHC class II (Miltenyi, Bergisch Gladbach, Germany). For some assays, BMDC were infected with WT L. major promastigotes instead, and fluorescein isothiocyanate (FITC)-conjugated anti-CD40 (Biolegend, San Diego, USA) and FITC-conjugated anti-CD80 (eBioscience, San Diego, USA) Ab were used. Data was obtained using the MACSQuant flow cytometer (Miltenyi) and analyzed using FlowJo (Tree Star Inc., CA, USA). The expression of F4/80 in BMM was determined using FITC-conjugated anti-F4/80 Ab (Biolegend).
The expression of intracellular IL-12 was analyzed in BMDC after 24 hours of stimulation with LPS (1 µg/ml) at 37°C, 5% CO2, in the presence or absence of 10 µM CA074Me or 10 µM ZRLR, and brefeldin A (3 µg/ml, eBioscience). The cells were then incubated for 20 min in 4% PFA fixation buffer, permeabilized for 20 min at 4°C using 0.1% saponin, 1% FCS permeabilization buffer, and incubated for 1 hour with (PECy7)-conjugated anti-CD11c and PE-conjugated anti-IL-12(p40/p70, BD Biosciences. Data was obtained using the MACSQuant flow cytometer. Furthermore, cells from polarization assays described below were fixed with 2% formaldehyde for 20 min at 4°C, permeabilized for 20 min at 4°C, and stained with the following Ab diluted in permeabilization buffer: Pacific Blue-conjugated anti-CD4 (Biolegend), FITC-conjugated anti-IFN-γ (BD Biosciences), PE-conjugated anti-IL-4 (BD Biosciences), and allophycocyani-conjugated anti-IL10 (Biolegend). Data was obtained using a LSR-II flow cytometer (BD Biosciences, San Jose, USA). All results were analyzed using the software FlowJo.
BMM from WT and cathepsin-deficient mice were seeded and infected as described for the proliferation assay. After 24 hours of incubation, the cells were washed with phenol-free complete RPMI medium to eliminate any extracellular parasites and incubated for further 48 hours in the absence or presence of 1 µg/ml LPS. The supernatants were collected, and the concentration of nitrite () was determined by addition of 100 µl of culture supernatant to 100 µl of Griess reagent (Sigma-Aldrich) and incubation for 15 min at RT. The resulting absorbance at 540 nm was measured with the Multiskan Ascent ELISA reader (Thermo Electronic Corporation). The nitrite concentrations were determined using sodium nitrite (NaNO3) as a standard, and reflect the NO levels released by macrophages.
1×106 BMDC were seeded in a final volume of 1 ml in 24-well plates, and were stimulated with 5×106 L. major WT promastigotes (infection ratio 1∶5), LmAg (30 µl/ml), LPS (1 µg/ml; Sigma-Aldrich), or CpG ODN 1668 (5′-TCCATGACGTTCCTGATGCT-3′, Qiagen Operon, Cologne, Germany). The cells were further incubated for 24 or 48 hours, and the supernatants were collected. The concentration of the cytokines in the supernatants was determined by sandwich ELISA, using capture-detection Ab pairs purchased from BD Biosciences for IL-12p40, IL-6 and tumor necrosis factor alpha (TNF-α), and R&D Systems for IL-10 (Wiesbaden, Germany) following the suppliers' instructions. In addition, IL-12p70 was measured by using the IL-12p70 ELISA Ready-SET-Go kit from eBioscience according to the manufacturer's instructions. To analyze the cytokine production in BMM, 1×106 cells were seeded in 500 µl into 24-well plates, together with 15×106 L. major WT promastigotes (infection ratio 1∶15), in the presence or absence of LPS (1 µg/ml). The cells were incubated for 24 and 48 hours, and the supernatants were collected. Cytokine measurements by ELISA were performed as described above.
Total RNA from 2×106 BMDC or BMM, stimulated as described above, was isolated using the RNeasy kit (Qiagen, Hilden, Germany) according to the manufacturer's instructions. cDNA synthesis was performed using the iScript cDNA synthesis kit (BioRad, Munich, Germany) and the resulting cDNA was used at a 1∶8 dilution to assess the expression of IL-12a(p35) by real-time PCR. The real-time PCR was performed in a final volume of 25 µl per well using Maxima SYBR Green/Fluorescein qPCR Master Mix (Thermo Scientific, Schwerte, Germany) and run with a CFX96 Touch real-time PCR detection system (BioRad) for 40 cycles. The primer pairs used were: Il12p35 forward: TGGCTACTAGAGAGACTTCTTCCACAA, Il12p35 reverse: GCACAGGGTCATCATCAAAGAC; Il12p40 forward: CGTGCTCATGGCTGGTGCAAA, Il12p40 reverse: ACGCCATTCCACATGTCACTGCC. The housekeeping gene β-actin was used for normalization of the samples: β-actin forward: CATTGCTGACAGGATGCAGA, β-actin reverse: TTGCTGATCCACATCTGCTG. Non-treated (NT) BMM were used as negative control, and LPS-stimulation (1 µg/ml) was used as positive control. Relative gene expression values were calculated with the 2− ΔΔCT method [36] using WT NT BMM at t = 6 hours as a reference.
Lymph nodes and spleens were removed from OVA-specific T-cell receptor (TCR)-transgenic OT-II mice, and kept in ice-cold complete RPMI medium in 60×15 mm petri dishes. Lymphocytes and splenocytes were isolated by mechanical dissociation using the sterile plunger of a 5-ml syringe and a cell strainer (70 µm, BD Falcon, Durham, USA). Red blood cells from spleen suspensions were eliminated with ammonium chloride lysis buffer for 5 min at 37°C. Naïve CD4+ T cells were isolated by negative selection using the CD4+ T-cell enrichment kit (StemCell Technologies, Grenoble, France) following the manufacturer's instructions. The enriched cells (1×104) were co-cultured with day 8 BMDC (5×104) from WT C57BL/6, Ctsb−/− or BALB/c mice, together with 1 mg/ml ovalbumine (OVA, Hyglos, Bernried, Germany) or 100 ng/ml OVA-peptide 327–339 (Activotec, Cambridge, UK.), and LPS (0.1 µg/ml), in U-bottom 96-well plates, with a final volume of 200 µl/well at 37°C, 5% CO2. After 5 days of culture, the cells were harvested, counted, and adjusted to a concentration of 1×106 cells/ml for re-stimulation with phorbol 12-myristate 13-acetate (PMA,10 ng/ml, Sigma-Aldrich), ionomycin (1 µg/ml, Sigma-Aldrich), and brefeldin A (3 µg/ml), for 5 hours at 37°C, 5% CO2. The cells were then washed, fixed in 2% formaldehyde, incubated for 20 min in saponin buffer, and the expression of Th1 cytokines was assessed by staining of intracellular cytokines, and flow cytometry as described above.
5×106 BMM were seeded in 6-well cell culture plates, and incubated for 4 hours at 37°C, 5% CO2 to promote adherence. The cells were thereafter infected with L. major promastigotes using a 1∶15 ratio, and further incubated at 37°C, 5% CO2. At different time points (t = 0, t = 15 min, t = 30 min, and t = 1 h), lysates were prepared as follows: two different buffers were prepared, cytoplasmic cell fractionation buffer (10 mM HEPES, 10 mM KCl, 1.5 mM MgCl2, 0.34 M D-sucrose, 10% glycerin and 1 mM dithiothreitol (DTT), and nuclear cell fractionation buffer (3 mM ethylenediaminetetraacetic acid (EDTA), 0.2 mM ethylene glycol tetraacetic acid (EGTA), and 1 mM DTT). Directly prior to use, both buffers were supplemented with DTT (final concentration 0.5 mM), protease inhibitor cocktail (1∶100 dilution, Sigma-Aldrich), and Na3VO4 (final concentration 1 mM). At each time point, the stimulated cells were washed twice with cold PBS, and resuspended in 90 µl of ice-cold cytoplasmic cell fractionation buffer. 10 µl of 1% Triton X-100 in cytoplasmic cell fractionation buffer were added to the samples, and they were further incubated for 5 min on ice with gentle agitation. The samples were then centrifuged at 2000× g for 5 min at 4°C, and the supernatants were collected as cytoplasmic fraction, and stored at −20°C. The pellets were then washed with 100 µl of cytoplasmic cell fractionation buffer, and the samples were centrifuged again as described above. The supernatants were discarded, and the pellets were resuspended in 60 µl of nuclear cell fractionation buffer. The samples were further incubated for 30 min on ice. Then, they were sonicated on ice (Sonoplus, Bandelin, Berlin, Germany) using two cycles of 20 s each, with 40% of amplitude. The resulting suspensions were collected as nuclear fraction, and were stored at −20°C. The protein concentration of each sample was determined using a microplate bicinchoninic acid (BCA) protein assay kit (Thermo Scientific), following the manufacturer's instructions. 40 µg of protein from each sample were separated by SDS-PAGE (10% acrylamide gels), and transferred to poly(vinylidene difluoride; PVDF) membranes. The membranes were incubated overnight with primary Ab against the p65 subunit of NFκB (1: Santa Cruz, Dallas, USA, 2: Cell Signaling, Danvers, USA), MEK1/2, Lamin A/C (Cell Signaling), and Ctsb (R&D Systems, Minneapolis, USA). For detection, the membranes were incubated for 1 hour at RT with their corresponding horseradish peroxidase (HRP)-conjugated secondary Ab (Cell Signaling), and developed using a chemiluminescence kit (GE Healthcare, Munich, Germany). The membranes were then visualized using a FluorChem Q imager (Biozym Scientific, Oldendorf, Germany).
Values are provided as mean ± standard deviations from at least 3 independent experiments. Statistical significance was determined by the unpaired 2-tail Student's t test (Microsoft Excel Software) comparing, for each treatment, the results from Ctsb−/− or Ctsl−/− cells with their WT counterparts.
We used bone marrow stem cell progenitors from Ctsb−/−, Ctsl−/− and WT (C57BL/6) mice to generate BMDC. At day 8 of culture with GM-CSF, the cells from all these mice displayed a typical myeloid immature DC morphology (Figure 1A, 1–6). These cells presented similar levels of CD11c expression, and comparable yields of CD11c+ cells were obtained (Figure 1B). Similarly, BMM generated from these cathepsin-deficient mice did not show any significant differences in morphology (Figure 1C, 1–6) and levels of F4/80 expression (Figure 1D) in comparison with BMM from WT mice.
Next, eGFP-tg L. major promastigotes were used to analyze the kinetics of parasite uptake and processing by BMDC. After 2 hours of culture with L. major promastigotes at a parasite-to-BMDC ratio of 5 to 1, most of the WT cells (70%±9.8%) were infected with L. major, and rapidly processed the phagocytosed promastigotes (Figure 2A). At 24 hours after infection, only around 13% of the cells remained infected. BMDC from Ctsb−/− and Ctsl−/− mice showed no significant differences neither in the uptake of eGFP-tg L. major promastigotes nor in their kinetics for processing the parasites in comparison with WT BMDC (Figure 2B). These results indicate that cathepsins B and L are not relevant for the generation of BMM and BMDC, and that the capacity of the latter to phagocytize and process L. major promastigotes is not altered by the lack of cysteine cathepsins.
We examined the survival of eGFP-tg L. major in macrophages. At 24 and 48 hours after infection, BMM from WT, Ctsb−/− and Ctsl−/− mice showed no significant differences in terms of percentage of infected cells, and average number of parasites per infected cell (Figures 3A and 3B, and Figure S1). To confirm these results, we also infected BMM from these mice with Luc-tg. L. major promastigotes, and measured the luminescence produced after addition of a luciferin substrate as a read-out for intracellular parasites. Again, no significant differences were found between WT BMM and cathepsin-deficient-BMM (Figure 3C).
In addition, we measured the nitrite concentrations in supernatants of BMM infected with L. major in the presence or absence of LPS, and in response to LPS alone (Figure 3D). We found that 48 hours of infection with L. major promastigotes alone did not result in higher NO production compared to non-infected cells from WT, Ctsb−/− and Ctsl−/− BMM. Furthermore, we found no significant differences in nitrite levels in the supernatants of these cells neither after stimulation with LPS alone nor after infection with L. major promastigotes and further stimulation with LPS. These results show that cathepsin B and cathepsin L are dispensable for the control in vitro of L. major in BMM, and that absence of either of them does not affect the capacity of BMM to produce NO in response to neither L. major nor LPS.
Upon encounter with pathogens, immature DC become activated and mature, up-regulating the expression of the antigen-presenting molecules MHC class II as well as co-stimulatory molecules such as CD86, CD80 and CD40 [37]. We analyzed the maturation profile of BMDC from WT, Ctsb−/− and Ctsl−/− mice 24 hours after uptake of L. major promastigotes. We found that the expression of MHC class II molecules was greatly enhanced in BMDC from Ctsb−/− mice and, to a lesser extent, in BMDC from Ctsl−/− mice in response to L. major compared to WT BMDC (Figure 4A). The expression of the co-stimulatory molecules CD40, CD86, and CD80, on the other hand, was comparable among WT, Ctsb−/− and Ctsl−/− BMDC (Figures 4B and 4C). In addition, we tested the expression of MHC class II and co-stimulatory molecules of BMDC pre-incubated with CLIK148 and CA074Me, a modified form of CA074 with increased cell permeability. BMDC treated with CA074Me showed an up-regulation of MHC class II molecules in response to L. major promastigotes higher than that observed for BMDC treated with CLIK148 or DMSO. On the other hand, we found no effect on CD86 expression, similar to the results obtained with the use of cathepsin-deficient cells (Figure S2 A and B).
These higher expression levels of MHC class II molecules in Ctsb−/− BMDC appeared to be a specific response to living parasites, since stimulation of Ctsb−/− BMDC with LmAg or HK parasites did not result in enhanced expression levels of MHC class II molecules and co-stimulatory molecules in comparison with WT and Ctsl−/− mice (Figure 4). In addition, no significant differences in the expression of co-stimulatory and MHC class II molecules were found in WT and cathepsin-deficient BMDC upon stimulation with LPS. The activation by inflammatory stimuli is known to recruit cathepsins B, L and S to late endosomes [38], [39], and, therefore, it is possible that different stimuli would lead to different profiles of active cathepsins for antigen processing.
Our results demonstrate that L. major-stimulated Ctsb−/− BMDC, on the basis of MHC class II expression, display an enhanced maturation compared to WT and Ctsl−/− BMDC in response to L. major promastigotes. On the other hand, no significant differences in the expression of the co-stimulatory molecules CD86, CD80, and CD40 were observed. This effect was not elicited by stimulation with LmAg, HK parasites, or with LPS.
Another important signal that BMDC use to instruct Th cell polarization is cytokine production. We analyzed the concentrations of the cytokines IL-12p70, IL-12p40, IL-10, IL-6 and TNF-α in supernatants of BMDC 48 hours after infection with L. major promastigotes, or stimulation with LmAg. We found a significant increase in both IL-12p70 and IL-12p40 levels in Ctsb−/− BMDC in response to L. major in comparison with WT and Ctsl−/− BMDC (Figures 5A and 5B). We also found that IL-10 expression was enhanced in Ctsb−/− BMDC (Figure 5C), resembling the production of IL-10 in response to an up-regulation of IL-12 observed in DC after stimulation with LPS, a Th1 inducer. Moreover, the IL-12 up-regulation required living parasites, since stimulation of BMDC with LmAg or HK parasites did not induce higher levels of IL-12 production. In addition, we found no differences in IL-6 and TNF-α production between WT and cathepsin-deficient BMDC in response to L. major parasites (Figure S3). In contrast, Ctsb−/− BMDC presented an impaired IL-12p70 expression in response to CpG in comparison with WT BMDC, which reflects the importance of Ctsb in Toll-like receptor 9 (TLR9) signaling (Figure S4). The IL-12 up-regulation in response to L. major observed with Ctsb−/− BMDC could not be replicated using the inhibitor CA074Me (Figure S2, C), and this inhibitor caused a dose-dependent decrease in IL-12p70 in LPS-stimulated cells (Figure S2, D). However, when BMDC from BALB/c and C57BL/6 mice were pre-treated with the peptide-based cathepsin B inhibitor ZRLR, they up-regulated their expression of IL-12p70 in response to L. major promastigotes resulting in levels comparable to those observed in Ctsb−/− BMDC (Figure S5, A). Pre-treatment of Ctsb−/− BMDC with ZRLR did not cause significant differences in the expression of IL-12 in response to L. major in comparison with DMSO pre-treated Ctsb−/− BMDC. In addition, we found a very similar pattern of increased IL-12p70, IL-12p40 and IL-10 expression in Ctsb−/− BMM in comparison with WT and Ctsl−/− BMM (Figures 5D to F). It should be noticed that the levels of IL-12p40 were considerably lower in BMM than in BMDC.
Moreover, Ctsb−/− BMDC stimulated with LPS also produced slightly more IL-12p70 and IL-12p40, but not IL-10, compared to WT and Ctsl−/− BMDC. However, TNF-α expression was greatly impaired in Ctsb−/− BMDC, and no significant differences in IL-6 production among WT, Ctsb−/− and Ctsl−/− were found (Figure 6). This effect in IL-12 expression was also found in BMDC from BALB/c and C57BL/6 mice pre-treated with ZRLR (Figure S5, B to D). When co-cultured with naïve T cells from OT-II mice, we found that Ctsb−/− BMDC having LPS as a maturation stimulus resulted in higher frequencies of IFN-γ+ T cells, but not of IL-4+ T cells, indicating a Th1 polarization (Figure 7). Furthermore, the enhanced IL-12 production in response to L. major and LPS that we observed was found also at the transcriptional level, since Ctsb−/− BMM presented an up-regulation in the expression of Il12p35 and Il12p40 in response to both stimuli (Figure 8).
Next, we investigated if this up-regulation of IL-12 was dependent on the NF-κB signaling pathway by assessing the translocation of the p65 subunit from the cytoplasm to the nucleus (Figures S6 and S7). We tested separately two different monoclonal antibodies to detect the NF-κB p65 subunit by Western Blot, and used for analysis the protein bands with the expected molecular weight (65 kDa). We found that the different levels of p65 for all the treatments in nuclear fractions from WT and Ctsb−/− had no statistical significance. We should point out that we found with both antibodies multiple protein bands with molecular weights other than 65 kDa. In particular, we observed a non-identified protein with a molecular size around 30 kDa in WT LPS-stimulated BMM but not in LPS-stimulated Ctsb−/− BMM. While these results suggest that NF- κB is not responsible of the Ctsb−/− -mediated regulation of IL-12, more experiments with different approaches would be needed to confirm this observation.
Altogether, our results show that Ctsb−/− BMDC and BMM presented a significant up-regulation of the Th1 promoter cytokine IL-12 in response to L. major and LPS, in comparison with WT and Ctsl−/− cells. This effect could be replicated using the peptide-based cathepsin B inhibitor ZRLR in BMDC from mouse strains susceptible and resistant to L. major. Furthermore, the observed up-regulation of IL-12 was present already at the transcriptional levels, and Ctsb−/− BMDC induced higher frequencies in vitro of Th1-polarized T cells.
In the present study, we investigated the differences in the signals that BMDC use to instruct Th cell polarization, i.e., antigen presentation, expression of co-stimulatory molecules and cytokines, as a response to L. major promastigotes in the absence of Ctsb and Ctsl. In addition, we analyzed the impact of the lack of these proteases on the proliferation of L. major in infected BMM. We found that Ctsb−/− BMDC express higher levels of MHC class II molecules and of IL-12 in response to L. major promastigotes, and that this up-regulation of IL-12 expression was also present in BMM. These results indicate a novel role for Ctsb in the regulation of cytokine expression in response to L. major.
Stem cell progenitors from Ctsb−/− and Ctsl−/− mice are able to generate BMDC and BMM with comparable yields and phenotypes as WT mice. These cells presented similar rates of parasite uptake, and BMDC showed similar kinetics of parasite processing. Moreover, parasite survival was similar in cathepsin-deficient and WT BMM. These results indicate that the up-regulation in MHC class II molecules and IL-12 expression that we observed was not due to differences in the parasite load between WT, Ctsb−/− and Ctsl−/− BMDC.
DC present an incomplete maturation after uptake of L. major promastigotes [40] and L. amazonensis [41]. Previous studies using the Ctsb-selective inhibitor CA074 and the Ctsl-specific inhibitor CLIK148 showed drastic changes in the Th cell response of mice infected with L. major [25], [28], [42], and it was hypothesized that these effects resulted from differences in antigen processing. Later reports showed that cathepsin S is indispensable for the degradation of the invariant chain in antigen-presenting cells [43]–[45], while Ctsl was relevant for the processing of antigens only in cortical thymic epithelial cells [46]. Ctsb and cathepsin D (Ctsd) were shown to be dispensable for the maturation of MHC class II molecules and the presentation of several antigens [31]. However, this study also reported that splenocytic antigen-presenting cells from Ctsb- or Ctsd-deficient mice were actually more efficient to present certain antigens to T-cell hybridomas, in agreement with reports using different inhibitors with murine splenocytes [47] and primary human DC [48]. The latter study described the use of the peptide-based cathepsin inhibitor ZRLR which was shown to have superior specificity towards Ctsb compared to CA074Me. Deussing et al. suggested that some antigenic determinants may present different degrees of susceptibility to degradation by cathepsins before being able to bind to MHC class II molecules, and, therefore, would benefit from the absence of Ctsb or Ctsd [31]. We found higher levels of MHC class II molecules in Ctsb−/− BMDC than in WT and Ctsl−/− BMDC in response to L. major. Similar results were found with the use of the inhibitor CA074Me, while no significant differences were found when we used LmAg or heat-killed parasites as stimulus. The different results observed with promastigote- and LmAg-mediated stimulation could reflect the interaction of the living parasite with the host cell, and the differences in uptake mechanism and subsequent processing [47], [48]. Stimulation with heat-killed parasites led to comparable levels of MHC class II molecules and co-stimulatory molecules as observed with infected BMDC. However, we found no significant differences between WT and cathepsin-deficient BMDC. This could indicate that the higher levels of MHC class II molecules in infected Ctsb−/− BMDC in comparison with WT BMDC are related to the active manipulation that the living parasite exerts in its host cell. Incomplete BMDC maturation, such as after stimulation with Trypanosoma brucei antigens, has been shown to induce activation of genes correlating with the induction of Th2 polarization [49]. In contrast, higher levels of antigen presented [50] are associated with induction of Th1 responses.
Upon in vitro infection with Leishmania promastigotes, BMDC present poor cytokine expression [51]. We found that Ctsb−/− BMDC were able to express significantly higher levels of IL-12 (both p70 and p40 forms) than WT and Ctsl−/− BMDC in response to L. major promastigotes. We did not detect significant differences in IL-6 and TNF-α, which would indicate that the observed effect was not a generalized up-regulation of cytokine expression, but a rather specific mechanism. Likewise, infection of BMM with L. major promastigotes induces poor cytokine expression [17], [19]. Ctsb−/− BMM also presented a significant increase in IL-12 expression, with similar levels of IL-12p70 as Ctsb−/− BMDC, although the up-regulation of IL-12p40 was not as high. Pompei et al. reported a differential release of IL-12 in BMDC and BMM in response to Mycobacterium tuberculosis, and suggested that this was dependent upon the level of engagement of different TLR, particularly TLR9 in BMDC [52]. TLR9 requires processing by endosomal cathepsins to initiate signaling [53], [54]. In agreement with Matsumoto et al. [53], we observed a great impairment in IL-12 expression in Ctsb−/− BMDC upon CpG stimulation. Therefore, the up-regulation in IL-12 expression by Ctsb−/− BMDC and BMM that we observed here is independent from TLR9 signaling.
In addition, we tested the response of Ctsb−/− and Ctsl−/− BMDC to LPS, which is recognized by TLR4. Ctsb−/− BMDC stimulated with LPS did not show a significantly higher expression of MHC class II molecules in comparison with WT BMDC or Ctsl−/− BMDC but they did display higher levels of IL-12. Moreover, the expression of TNF-α was greatly impaired in Ctsb−/− BMDC, in agreement with Ha et al. [55] who reported that LPS-treated BMM secrete significantly less TNF-α in response to LPS upon lack of Ctsb, due to an accumulation of TNF-α-containing vesicles that could not reach the plasma membrane. Schotte et al. reported an impairment of cytokine production in macrophages stimulated with LPS upon treatment with the cathepsin B inhibitor z-FA.fmk [56]. Our results with Ctsb−/− BMDC and BMM do not show an inhibition of IL-12 expression but rather an enhancement. We obtained similar results using CA074Me, the methyl ester form of CA074. Upon uptake by the cell, CA074Me is hydrolyzed to CA074, but if this hydrolysis is incomplete, inhibition of other cysteine proteases besides Ctsb takes place [57]. In contrast, pre-treatment of BMDC from susceptible BALB/c or resistant C57BL/6 mice with ZRLR induced IL-12 expression levels comparable to those observed with Ctsb−/− BMDC. Although we did not have Ctsb−/− mice on a BALB/c background available, these results suggest that the observed role of Ctsb in L. major infection would be independent of the mouse strain.
Upon infection with L. major promastigotes, Ctsl−/− BMDC and BMM did not present significant differences in the production of cytokines in comparison with WT cells. However, they produced higher levels of IL-10 and TNF-α, but not IL-12, in response to LPS. These results alone would not explain the observations made by Onishi et al. [58], where use of the cathepsin L inhibitor CLIK148 caused a Th2-like immune response to L. major in otherwise resistant mice. Again, it should be kept in mind that CLIK148 can also inhibit other cathepsins, including C, K, and S [59], which could have contributed to this response.
Altogether, while previous studies hypothesized that lack of Ctsb or Ctsl activities during L. major infection would lead to changes in Th cell polarization due to differences in antigen presentation [25], [28], [42], our results indicate that Ctsb−/− BMDC up-regulate two of the three types of signals used for instructing Th cell polarization: expression of MHC class II molecules and cytokine expression. Thus, these cells exhibit a “pro-Th1”-like profile. Moreover, co-culture of purified naïve CD4+ T cells with Ctsb−/− BMDC resulted in a higher frequency of Th1-polarized T cells compared to WT BMDC. To the best of our knowledge, the present study is the first indicating a new role of Ctsb as a regulator of cytokine expression in response to L. major. Future work will focus on the implications of these effects in vivo, considering the infection of Ctsb−/− animals, as well as in transfer experiments of Ctsb−/− BMDC into WT animals.
In which way could Ctsb influence cytokine production in BMDC and BMM? Ben-Othman et al. reported that L. major parasites induce macrophage tolerance by a process involving MAPK and NF-κB pathways of the host [60]. These pathways, although initially activated by exposure to the parasite, become silenced when the infection is firmly established, rendering the cell unresponsive to further stimulation with LPS [61]. This silencing has been attributed to different virulence factors, including surface phosphoglycans [16], [62], the metalloprotease GP63 [63], and cysteine proteases from L. mexicana [21]. It is possible that in the absence of Ctsb, one or more of these key signaling pathways are no longer silenced by L. major promastigotes. This would open a range of new questions regarding the involvement of Ctsb in L. major infection, e.g., whether Ctsb interacts directly with the parasites, contributing to processing or activation of one or more virulence factors, or whether Ctsb directly interferes with any intermediate of key signaling pathways, such as NF-kB and MAPK. While Ctsb−/− BMDC and BMM were able to up-regulate IL-12 in response to L. major and LPS, IL-6 was not regulated. IL-6 transcription has been shown to depend on NF-κB [64]. Therefore, we hypothesize that the molecular mechanism behind the involvement of Ctsb−/− in the expression of IL-12 would not be shared by IL-6. Our results suggest that the regulation of IL-12 expression by Ctsb is not NF-κB-dependent, although further work is necessary to confirm this observation and to explore the interaction of Ctsb with other candidate signaling pathways. In addition, finding the location of these interactions would be a key to further understand the mechanisms underlying these processes, e.g., whether proteolytic processing of signaling intermediates takes place in the cytoplasm, or whether cleavage of transcription factors by Ctsb occurs in the nuclear space, as described in thyroid carcinoma cells [65].
The concept of “protease signaling” has gained increasing attention in different research fields [66], especially in the context of therapeutic applications. Yet, more research is needed in order to understand the interplay between proteolytic networks and other signaling pathways. On the basis of the present study, we propose a novel role for cathepsin B during L. major infection: in addition to its involvement in antigen presentation, it is also a regulator of cytokine expression. It is tempting to speculate that pharmacological inhibition of cathepsin B may improve the Th1-mediated clearance of L. major.
|
10.1371/journal.pntd.0005415 | Analysis of Mycobacterium ulcerans-specific T-cell cytokines for diagnosis of Buruli ulcer disease and as potential indicator for disease progression | Buruli ulcer disease (BUD), caused by Mycobacterium (M.) ulcerans, is the third most common mycobacterial disease after tuberculosis and leprosy. BUD causes necrotic skin lesions and is a significant problem for health care in the affected countries. As for other mycobacterial infections, T cell mediated immune responses are important for protection and recovery during treatment, but detailed studies investigating these immune responses in BUD patients are scarce. In this study, we aimed to characterise M. ulcerans-specific CD4+ T cell responses in BUD patients and to analyse specific cytokine-producing T cells in the context of disease severity and progression.
For this case-control study, whole blood samples of BUD patients (N = 36, 1.5–17 years of age) and healthy contacts (N = 22, 3–15 years of age) were stimulated with antigen prepared from M. ulcerans and CD4+ T cells were analysed for the expression of TNFα, IFNγ and CD40L by flow cytometry. The proportions and profile of cytokine producing CD4+ T cells was compared between the two study groups and correlated with disease progression and severity. Proportions of cytokine double-positive IFNγ+TNFα+, TNFα+CD40L+, IFNγ+CD40L+ (p = 0.014, p = 0.010, p = 0.002, respectively) and triple positive IFNγ+TNFα+CD40L+ (p = 0.010) producing CD4+ T cell subsets were increased in BUD patients. In addition, TNFα+CD40L-IFNγ- CD4+ T cells differed between patients and controls (p = 0.034). TNFα+CD40L-IFNγ- CD4+ T cells were correlated with lesion size (p = 0.010) and proportion were higher in ‘slow’ healers compared to ‘fast healers’ (p = 0.030).
We were able to identify M. ulcerans-specific CD4+ T cell subsets with specific cytokine profiles. In particular a CD4+ T cell subset, producing TNFα but not IFNγ and CD40L, showed association with lesion size and healing progress. Further studies are required to investigate, if the identified CD4+ T cell subset has the potential to be used as biomarker for diagnosis, severity and/or progression of disease.
| Buruli ulcer disease (BUD) is a devastating skin disease characterised by nodules, plaque or oedema at early stages, which progress to a characteristic form of ulcer. Without treatment the disease can cause broader tissue destruction, affect bones and cause permanent disabilities. BUD is treated with a combination of two antibiotics over a period of eight weeks. Overall this treatment scheme is effective, but duration of the healing process varies strongly. Specific immune responses are of major importance for controlling other mycobacterial infections (e.g. tuberculosis), but are also used for immune-based diagnostic tests. However, detailed knowledge of specific immune responses in BUD patients is missing. The aim of this study was to characterise the specific cytokine response of CD4+ T helper cells in BUD patients and to correlate specific T cell subsets to clinical disease progression. Using flow cytometry on Mycobacterium ulcerans-stimulated blood samples, we identified specific cytokine-secreting T cell subsets, induced in BUD patients. Proportions of a CD4+ T cell subset, producing a single cytokine were positively associated with lesion size and healing time. Results of this study provide deeper insights in the specific immune response in BUD patients and identify a potential indicator for early diagnosis and disease progression.
| Buruli ulcer disease (BUD), caused by Mycobacterium ulcerans, is a neglected tropical disease with reported cases in 33 subtropical and tropical countries [1]. The majority of cases have been recorded in 12 countries, mainly in Western and Central Africa with Ghana among those countries where BUD is a significant problem for public health [1]. Notably about half of all cases are diagnosed in children and adolescents with dramatic consequences for their health and social life.
At early stages BUD is characterised by painless, subcutaneous papules, nodules, plaques or oedemas. Without treatment, most lesions enlarge progressively and ulcerate developing undermined skin borders [2, 3]. The disease can eventually destroy tissues, affect bones leading to deformation and may cause permanent disabilities [4, 5]. However, in some cases non-ulcerative lesions are stable and do not progressively enlarge or ulcerate.
Treatment outcome has substantially improved with the introduction of a combination therapy with rifampicin (given orally) and streptomycin (intramuscular injection) [1, 6, 7]. However an effective vaccine against infection is currently not available. Studies on protection of Bacillus Calmette-Guérin (BCG) vaccination led to contradictory results [8–10].
In the absence of protective vaccines and with limited knowledge about the exact mode of transmission [11], prevention of BUD is currently not feasible [12]. Therefore disease management relies on early case detection, reliable diagnosis and on early effective treatment. Classical methods of diagnosis such as isolation by M. ulcerans culture or smear microscopy of acid-fast bacilli take many weeks and/or lack sensitivity [13, 14]. The current gold standard for diagnosis is based on PCR dependent detection of the M. ulcerans-specific insertion sequence IS2404 from lesion specimens. This test shows high sensitivity and specificity, but is not helpful in monitoring disease progression.
Immune-based assays provide a reliable tool for the detection of other mycobacterial diseases. In tuberculosis, Interferon-gamma release assays (IGRAs) are important for the diagnosis of Mycobacterium tuberculosis infections [15]. IGRAs detect interferon-gamma (IFNγ) following in vitro stimulation with specific antigens. IFNγ is either quantified from plasma of cultured blood (QuantiFERON) or by detecting IFNγ producing T cells (T-SPOT.TB). A comparable test based on the activation of M. ulcerans-specific T cells is currently not available. However, it is tempting to speculate, that such an immune based test could be beneficial for early diagnosis, prognosis of healing progress and monitoring response to antibiotic treatment in BUD patients.
For most mycobacterial infections, including tuberculosis, acquired cellular immunity is important for protection, but cellular immune responses are not well defined in BUD. It is known that up to one-third of lesions can heal spontaneously [16–18] and formation of granulomas has been reportedly associated with an induction of a proinflammatory and a down-modulation of inhibitory immune responses likely affecting the number of bacilli in the lesions [19, 20]. Therefore, it has been suggested that cellular immunity plays an important role in the immune response against M. ulcerans.
So far only few studies exist characterising immune responses in the context of M. ulcerans infection [21–25]. For instance, a broad analysis of systemic serum chemokines and cytokines revealed suppression of proteins including macrophage inflammatory protein (MIP)-1β, monocyte chemoattractant protein (MCP)-1 and IL-8 indicating immune modulation during M. ulcerans infection [26]. Immune responses to M. tuberculosis infections strongly depend on T helper type 1 cytokines IFNγ and tumor necrosis factor alpha (TNFα). Counteracting, regulatory immune responses based on the inhibitory cytokines Interleukin (IL-)10 and TGF-β are of major importance in several infectious diseases including tuberculosis [27]. In lesions of BUD patients, the key cytokines IFNγ and TNFα and the inhibitory cytokines IL-10 and TGF-β are produced, but the relative expression of these cytokines varied with the stage of the disease [20]. Modulated concentrations of TNFα were reported in serum of BUD patients if compared to controls [28]. Higher levels IFNγ and IL-10 were detected in BUD patients compared to household contacts or non-endemic controls following stimulation with M. ulcerans sonicate, in a study focusing on IFNγ and IL-10 from supernatants of whole blood [29]. However, there are no studies investigating proportions of cytokine producing CD4+ T cells in BUD patients, which allow dissecting regulation of specific cytokine-producing CD4 T cell subsets in the context of BUD, including multiple cytokine producing CD4+ T cells or T cells producing a single cytokine (based on what has been measured).
The analysis of specific cytokine producing CD4+ T cells in active tuberculosis and latently infected tuberculosis patients revealed differences in expression of CD40L- T cells [30]. Expression of CD40L on CD4+ T cells has been associated with a TH1 signature in infections with M. tuberculosis [31, 32]. Elsewhere, differences in IFNγ+TNFα+IL-2+, TNFα-single positive CD4+ T cells or a general increase in multi-cytokine producing T cells expression were observed between tuberculosis patients and contacts [33, 34]. Higher proportions CD40L+IL-2- CD4+ T cells were associated with cystic fibrosis patients infected with Mycobacterium abscessus [35], highlighting the potential of characterizing the precise profile of cytokine produced by CD4+ T cells in mycobacterial infections.
Therefore the aim of this study was to analyse the profile and frequencies of cytokine producing CD4+ T cells after stimulation with M. ulcerans antigen in BUD patients and healthy contacts and to analyse specific cytokine-secreting CD4+ T cell subsets in the context of disease severity and progression.
Ethical approval of the study was obtained from the ‘Committee on Human Research, Publication and Ethics (CHRPE) at the School of Medical Sciences, Kwame Nkrumah University of Science and Technology, Kumasi, Ghana (Ref.: CHRPE/AP/275/14 and CHRPE/AP/301/15). In addition, ethical approval was granted by the Ethical committee of the medical faculty of the Heinrich-Heine-University Dusseldorf (Ref.: 3903). The aims and procedures were explained to participants and/or their parents/guardians prior recruitment into the study. Only compliant patients were recruited and they were free to withdraw at any point during the study. Written consent was obtained from the parents or guardians. In some cases (illiterate guardians/parents) consent was confirmed by thumbprint, a procedure approved by the review board.
Between September 2014 and February 2016, patients with BUD were recruited at the Agogo Presbyterian Hospital in the Asante Akim North District, where there is high incidence of BUD in the middle forest belt of Ashanti region of Ghana [36].
A patient was recruited when the presenting lesion was consistent with the WHO clinical disease definition for BUD and diagnosis later confirmed by M. ulcerans IS2404 PCR. This study is part of a larger study investigating immune modulation in BUD patients with and without concomitant co-infections in children and adolescence. Up to 50% of all BUD cases in Africa are diagnosed in children below the age of 17 years [1, 37] with age being a significant factor for the clinical presentation of the disease [38]. The present study is specifically focusing on children and adolescent. Participants had to be ≤ 17 years of age, to be recruited into the study. Fifty-one of 101 (50.5%) patients were excluded for falling outside the age limit. Patients were also excluded if they i) had a history of BUD, tuberculosis or leprosy, ii) had a history of liver or kidney diseases, iii) were HIV positive (none excluded) or tested positive for any helminth infection, iv) had a recent or current antibiotic use. A flow chart indicating selection of the study group is shown in S1 Fig.
A control group of age- and gender matched healthy contacts was recruited from siblings, relatives or household contacts of participants. All participants were asked about their presenting active BUD disease, family history of BUD exposure, their previous medical history, household demographics and previous treatments.
Clinical BUD cases were confirmed by PCR for the IS2404 repeat sequence specific for M. ulcerans [39]. Lesions were classified by their form: nodule, oedema, plaque or ulcer. Lesion size was determined by measuring the widest diameter of a lesion and surface area lesion in cm2 (to take into account differences in the shape of lesions). Lesions were categorized following WHO guidelines: Category I: lesion <5cm in the widest diameter, Category II: lesion <15 cm, category III: lesion >15 cm, multiple lesions, osteomyelitis. Stool samples were taken for diagnosis of soil-transmitted helminths (STH) and infection with S. mansoni using the Formol-Ether concentration method. Microfilariae were detected using Sedgewick chamber using 100 μl of whole blood [40]. Depending on the initial result, up to 1000 μl of blood was filtered on a nucleopore filter membrane (3 μm) (Whatmann). Filters were stained with Giemsa and analysed by microscopy [40]. Malaria (Plasmodium falciparum) status was determined using the rapid test CareStart Malaria HRP2 pf (Access Bio, Inc.). Haematological parameters were assessed using the Sysmex XS-800i system (Sysmex).
Patients with BUD received a standard treatment regime with 15 mg/kg streptomycin and 10 mg/kg rifampicin daily for 8 weeks, as recommended by the WHO [7]. Patients presented every two weeks during antibiotic treatment and monthly subsequently for monitoring of healing progress until complete healing.
Up to 10 mL of venous blood was collected into heparinised blood collection tubes (BD Biosciences) between 9am and 12noon, prior initiation of antibiotic treatment. Blood samples were transported to the laboratory based in Kumasi and immediately processed (in less than 6 hours after blood has been taken [41, 42]). Surface stains to identify T, B, NK and CD16+ myeloid cells was performed directly in 100 μl of whole blood, diluted with 100 μl RPMI 1640 supplemented with 100 U/mL penicillin and 100 μg/mL streptomycin. Following incubation with the fluorochrome-conjugated antibodies for 30 min, red blood cells were lysed using a red blood cell lysis buffer (Roche) and remaining leucocytes were fixed for 15 min using Cytofix Solution (Biolegend).
For detection of intracellular cytokines, 100 μl of whole blood was stimulated in 96 well round bottom plates with M. ulcerans antigen at a final concentration of 5 μg/mL or with Staphylococcal enterotoxin b (SEB; Sigma) at a final concentration of 15 μg/mL. M. ulcerans antigen sonicate was prepared from a M. ulcerans 1 isolate of African origin. The origin and preparation of the M. ulcerans antigen are described in detail elsewhere [29].
Stimulated blood was incubated for a total of 17.5 hr at 37°C, 5% CO2 and Brefeldin A (Sigma) was added after an initial incubation of 2.5 hr. Whole blood cultured in medium without any stimuli was used as unstimulated negative control. Following incubation, red blood cells were lysed and the remaining cells fixed and permeabilized using a permeabilization Wash Buffer (Biolegend). Fluorochrome-conjugated antibodies (CD4, TNFα, IFNγ, CD40L) were added and incubated for 30 min followed by two additional wash steps.
Stained blood was acquired on a BD Accuri C6 Flow Cytometer (BD Biosciences) and analysed using FlowJo v10 (TreeStar). For further analysis, M. ulcerans or SEB-specific responses were determined by subtracting unstimulated control values from SEB or M. ulcerans stimulated cells. If subtracted values were ≤ 0, values were set to 0.001 for illustration purposes.
Following anti-human antibodies were used for flow cytometric analysis: APC-conjugated CD16 (clone 3G8), PerCP-Cy5.5.-conjugated CD3 (HIT3a), FITC-conjugated CD56 (HCD56), Alexa488-conjugated CD4 (RPTA-4), APC-conjugated TNFα (MAB11), PerCP-Cy5.5.-conjugated CD154/CD40L (24–31; all Biolegend), PE-conjugated IFNγ (2572311, BD Bioscience) and PE-conjugated CD20 (2H7, eBioscience). All staining panels were evaluated using fluorescence minus one and unstained controls.
Since analysed data were not normally distributed (Shapiro-Wilk) and did not meet assumptions for parametric tests, non-parametric tests were applied. For comparison of two groups (e.g. BUD patients vs. contacts) Mann-Whitney U test was used. Correlations were tested using Spearman’s rho analysis. Statistical analyses were performed using SPSS v23 (IBM Corp.) and were taken as significant if ≤ 0.05. Graphical illustration was done using Graphpad v7.0 (GraphPad Software, Inc.).
Lesion types were separated into groups based on their lesion form: non-ulcerative forms (nodule/oedema/plaque) and ulcerative forms. Patients were also separated into two groups based on the time to healing following the start of antibiotic treatment, using a cut-off of 111 days (based on the median healing time = 111.0 days, range 14–337 days) or alternatively based on the time healing was first recognized (cut-off based on the median of 56 days, range 14–231 days).
Blood samples were obtained from 36 BUD patients with a median age of 8.5 years (range 1.5–17 years) and from healthy contacts (median age 7.0 years; range 3.0–15 years). Detailed characterisation of the study groups are provided in Table 1. Haematological parameters were compared between BUD patients and contacts. There was a moderate, but significant increase in the frequency of basophils (Table 1). There was no difference between major lymphocyte subsets including T cells, B cells and NK cells (S2A–S2C Fig), but proportions of CD16+ myeloid cells were increased in BUD patients (S2D Fig). Additional haematological parameters did not differ between the two groups (Table 1). Clinical characteristics of the BUD group are shown in Table 2. Lesion size varied between individual BUD patients (2.3–79.4 cm2). Based on the widest diameter all lesions but one were classified as category I or II lesions (for details see methods section). One lesion with 15.7 cm in the widest diameter was classified as category III lesion. The majority of patients presented with either nodules (50.0%) or with ulceration (30.6%) (Table 2). None of the patients included into the study presented with multiple lesions or osteomyelitis.
To determine the specific cytokine producing profile of M. ulcerans specific CD4+ T cells, whole blood samples were stimulated overnight with M. ulcerans antigen and analysed for TNFα, IFNγ and CD40L producing T cells by flow cytometry. An example of the according gating and cytokine production is shown in S3 Fig. BUD patients had significant higher proportions of CD4+ T cells producing two (IFNγ+TNFα+, TNFα+CD40L+, IFNγ+CD40L+) or all three (IFNγ+TNFα+CD40L+) cytokines when compared to contacts (Fig 1A, upper panels). Only a minor fraction of the BUD contacts showed detectable levels of these cytokine-producing cells in response to M. ulcerans antigen. In addition, frequencies of two subsets characterised as TNFα+CD40L- and CD40L+IFNγ-, were increased in BUD patients (Fig 1B, upper panels). Further characterisation of these subsets revealed, that TNFα+CD40L- CD4+ T cells were almost exclusively IFNγ- (median of IFNγ- = 100%) and were therefore denoted as TNFα+CD40L-IFNγ-. In contrast, CD40L+IFNγ- contained both TNFα+ and TNFα- cells (median of TNFα+ = 42.9%). Neither CD40L+IFNγ-TNFα- nor IFNγ+CD40L-TNFα- single positive CD4+ T cells differed between BUD patients and contacts (p = 0.134 and p = 0.881 respectively).
No significant differences between both groups were observed following stimulation with Staphylococcal enterotoxin b used as polyclonal control stimulation (Fig 1A and 1B, lower panels). In summary, CD4+ T cells of BUD patients showed specific cytokine profile characterised by induction of multiple cytokine producing cells as well as an increase in the frequency of TNFα+CD40L-IFNγ- CD4+ T cells.
BUD patients presented at different stages of disease characterized by different types of lesions and varying lesion sizes (Table 2). Cytokine producing CD4+ T cells were analysed in the context of these clinical presentations to determine their potential as biomarkers. Data were split into two groups comprising patients with non-ulcerative forms (nodules, oedema, plaque) or ulcer. Neither multiple cytokine producing CD4+ T cells producing TNFα+IFNγ+, TNFα+CD40L+, TNFα+IFNγ+CD40L+ differed between the two groups (Fig 2A) nor TNFα+CD40L-IFNγ- CD4+ T cells showed any differences (Fig 2B). However, the size of lesions did not differ significantly between these two groups (p = 0.508). Therefore the lesion size was correlated with cytokine producing CD4+ T cells. There was no correlation between multiple cytokine producing T cells (Fig 2C). However, TNFα+CD40L-IFNγ- CD4+ T cells were positively correlated with surface area of the lesion (rho = 0.456; p = 0.010) (Fig 2D), which was reflected by a significant correlation with the widest diameter of the lesion (rho = 0.529; p = 0.004) (Fig 2E). None of the additional CD4+ T cell subsets presented in Fig 1B (TNFα+IFNγ-, IFNγ+TNFα-, IFNγ+CD40L-, CD40L+TNFα-) was correlated with lesion size (S4 Fig).
Patients were treated with a combination of rifampicin and streptomycin for eight weeks and time until complete healing was monitored. Applying the healing time, patients could be divided into ‘fast’ and ‘slow’ healers (Fig 3A). A cut-off of 111.0 days (median healing time) was applied to classify and distinguish these two groups. Notably, we detected significantly increased proportions of TNFα+CD40L-IFNγ- CD4+ T cells in ‘slow’ compared to ‘fast’ healers, whereas none of the other cytokine producing subsets were significantly different between the two groups (Fig 3B and 3C). Of note, ‘fast’ and ‘slow’ healers did not differ significantly by age (p = 0.193) or gender (p = 0.079) or original lesion size (Fig 3D). There was no significant correlation between healing rate within the first four weeks (expressed as mm/week) and TNFα+CD40L-IFNγ- CD4+ T cells (Fig 3E).
Some lesions do not immediately start to heal following initiation of antibiotic treatment, a phenomenon that has been described in BUD patients. For BUD patients included in this study healing started at a median duration of 56 days and this value was used to distinguish two groups. BUD patients with lesion starting to heal in less than 56 days had significant lower proportions of TNFα+CD40L-IFNγ- CD4+ T cells compared to patients with lesions starting after more than 56 days (Fig 3F).
Of note, is the fact that TNFα+CD40L-IFNγ- CD4+ T cells did not differ between BCG scar positives and negatives (p = 0.815).
In the present study, we show that the TNFα+CD40L-IFNγ- CD4+ T cell subset, induced by M. ulcerans antigen, provides a promising approach for establishing an immune-based assay for monitoring BUD.
We evaluated the expression of T helper type 1 associated cytokines in combination with CD40L following stimulation with M. ulcerans sonicate. In particular multiple cytokine expressing CD4+ T cells (e.g. IFNγ+TNFα+, TNFα+CD40L+, IFNγ+CD40L+) were induced in BUD patients compared to healthy contacts. In addition, proportion of TNFα+CD40L-IFNγ- were also higher in BUD patients making it a unique single cytokine producing subset. Both multiple cytokine producing CD4+ T cells as well as TNFα single positive T cells have been identified to discriminate between latent and active tuberculosis [33]. TNFα-single positive T cells were also the strongest predictor of an active tuberculosis in a larger study [43]. In BUD this unique TNFα-single positive T cell subset may reflect a strong inflammatory rather than a protective response. However, the functional role of TNFα+CD40L-IFNγ- CD4+ T cells in BUD needs to be investigated in more detail.
Diagnostic tests based on immunological responses are routinely used for detecting an infection with M. tuberculosis. IGRA’s are based on the antigen-induced production of IFNγ, but comparable assays are not available for BUD. Our findings may provide an approach to develop such an immune-based assay.
In BUD patients, suppression of immune responses can be recognized locally within lesions [20, 21, 44]. Whether a generalised systemic immune suppression can be detected is not yet clarified with contradictory findings in different studies [21, 23, 29, 45]. In our study, none of the analysed cytokine-producing CD4+ T cell subsets were lower in BUD patients compared with contacts suggesting that patients are able to mount TH1 responses confirming earlier results [29]. If immune suppression can be detected systemically as found by others [21, 46], may depend on the stage of disease, time of sampling, stimulation and method of analysis.
As there is currently no vaccine for BUD and the mode of transmission remains unclear disease management and optimization of treatment will remain highly important for the foreseeable future. Therefore we evaluated cytokine producing CD4+ T cells in the context of several clinical characteristics of BUD. First we compared BUD patients with different types of lesions (ulcer versus non-ulcerative forms). None of the analysed subsets including TNFα+CD40L-IFNγ- CD4+ T cells differed between the two groups.
Since the type of lesions was not necessarily associated with the size of lesion, we also correlated our identified cytokine producing CD4+ T cells subsets with surface area and with the widest diameter of lesions. Interestingly the only parameter showing a correlation was the TNFα+CD40L-IFNγ-, further supporting the role of this subset as biomarker.
An effective treatment based on rifampicin and streptomycin has been established and is recommended by the World Health Organization [1, 7]. The efficacy of the treatment regime is more than 90% [47, 48]. Nevertheless the healing process varies widely. Healing may start directly upon initiation of antibiotic treatment or it may take several weeks until healing becomes obvious and time until complete healing is obtained varies considerably. In addition, lesions may increase in size during or following initiation of antibiotic treatment, a phenomenon, which is mainly attributed to recovery of immune responses and is referred to as ‘paradoxical reaction’ [49–51]. Such paradoxical reactions are reported in more than 20% of cases [49, 52]. Given this, immunological markers, which help to predict healing and potentially ‘paradoxical reactions’ could be beneficial in terms of patient care and optimisation of antibiotic treatment. Since paradoxical reactions are thought to be based on changes in immune response [53], identifying biomarkers is a promising approach.
In the present study complete healing was observed between 14 and 337 days (median 111.0), which is in the range of observed in other studies [47, 48]. Comparing ‘slow’ versus ‘fast’ healers revealed indeed that only TNFα+CD40L-IFNγ- expression differed between the two groups. The difference in time to healing could theoretically be attributed to the fact that larger lesions need longer time for complete healing. However, in our study, the size of lesions did not differ significantly between ‘slow’ and ‘fast’ healers. Therefore the fact that TNFα+CD40L-IFNγ- differed between the two groups could not solely be attributed to an association with the original lesion size. In addition the healing rate within the first four weeks after start of antibiotic treatment did not correlate with TNFα+CD40L-IFNγ-. The time until first signs of healing were also recorded and varied between 14 and 231 days (median 56 days), consequently we analysed two groups based on this median. Indeed, the frequency of TNFα+CD40L-IFNγ- CD4+ T cells was higher in patients, which showed a delayed start of healing. Evaluating the healing progress using biomarkers may also be of importance in the evaluation of novel treatment regimens, such as full oral therapies, which are currently tested [54]. Of note, we did not have sufficient numbers of patients with a paradoxical reaction (N = 3) to evaluate differences in cytokine producing CD4+ T cells as prognostic marker for this. Therefore it needs to be further analysed if TNFα+CD40L-IFNγ- T cells are useful in this context in a larger study with more patients included.
The current study is focusing on children and adolescent. Age is an important factor affecting the clinical presentation of BUD [38] and in Africa more than 50% of all BUD cases are diagnosed in children. Therefore this age group would benefit most of improved biomarkers. However, in addition, TNFα+CD40L-IFNγ- producing CD4+ T cells as biomarker has to be tested in older BUD patients, which are more likely to develop severe forms of BUD [55].
In summary, we present here the first study analysing cytokine producing CD4+ T cells stimulated with M. ulcerans antigen focusing at T helper type 1 CD4+ T cells. Proportions of multiple cytokine as well as TNFα+CD40L-IFNγ- CD4+ T cells differed between BUD patients and healthy contacts and the later one was associated with lesion size and differed between ‘slow’ and ‘fast’ healers. Hence TNFα+CD40L-IFNγ- CD4+ T cells are a potential biomarker, which may allow the development of tests comparable to IGRAs.
|
10.1371/journal.pntd.0005869 | Increased rates of Guillain-Barré syndrome associated with Zika virus outbreak in the Salvador metropolitan area, Brazil | In mid-2015, Salvador, Brazil, reported an outbreak of Guillain-Barré syndrome (GBS), coinciding with the introduction and spread of Zika virus (ZIKV). We found that GBS incidence during April–July 2015 among those ≥12 years of age was 5.6 cases/100,000 population/year and increased markedly with increasing age to 14.7 among those ≥60 years of age. We conducted interviews with 41 case-patients and 85 neighborhood controls and found no differences in demographics or exposures prior to GBS-symptom onset. A higher proportion of case-patients (83%) compared to controls (21%) reported an antecedent illness (OR 18.1, CI 6.9–47.5), most commonly characterized by rash, headache, fever, and myalgias, within a median of 8 days prior to GBS onset. Our investigation confirmed an outbreak of GBS, particularly in older adults, that was strongly associated with Zika-like illness and geo-temporally associated with ZIKV transmission, suggesting that ZIKV may result in severe neurologic complications.
| Shortly following the introduction of Zika virus (ZIKV), a type of flavivirus transmitted by mosquitoes, into Brazil in early 2015, the Brazil Ministry of Health began receiving increased reports of a paralyzing condition known as Guillain-Barré syndrome (GBS). The areas with the greatest number of GBS cases appeared to correlate geographically and temporally with the areas reporting the highest rate of ZIKV infections. This association had been previously observed during a ZIKV outbreak in French Polynesia, however, this had not been systematically examined in a case-control investigation for the ZIKV outbreak in South America. In this investigation, the authors found that the occurrence of GBS in the affected population was nearly four times higher than would be expected, and the risk for GBS was particularly elevated among older adults. GBS was associated with ZIKV-like symptoms and with a combination of ZIKV-like symptoms plus laboratory evidence of a recent flavivirus infection. Taken together, these findings provide strong support for and greater understanding of the link between ZIKV and GBS.
| Guillain-Barré syndrome (GBS) is a peripheral polyneuropathy characterized by acute onset of bilateral, symmetric limb weakness with decreased or absent deep-tendon reflexes. GBS is a progressive illness with clinical nadir occurring generally within 2–4 weeks. The underlying disease mechanism by which GBS develops is thought to be related to an aberrant immune response following an infection or other immune stimulation [1]. The most common known inciting infection is Campylobacter jejuni, though sporadic cases of GBS have been described temporally following a myriad of other viral, bacterial, and parasitic infections [2]. Several emerging and re-emerging arboviruses, including dengue, chikungunya, West Nile, and Zika viruses, have been associated with isolated cases of GBS [3–6]. The onset of GBS symptoms typically manifests within 6–8 weeks, and particularly the following 10–14 days, after exposure [7]. Although clinical outcomes are generally favorable, approximately 20–30% of cases may develop autonomic disturbances and/or neuromuscular respiratory failure, which are the most common causes of death in GBS [8]. Reported mortality rates range from 3–7% in North America and Europe to 13% in parts of Asia [8, 9].
Zika virus (ZIKV), a flavivirus primarily transmitted by Aedes spp. mosquitoes [10–12], was originally identified in Uganda in 1947 [13]. Historically, it has been associated with sporadic cases of rash illness in Africa and Southeast Asia [14–17], though outbreaks of ZIKV began to emerge in the Western Pacific region during the late 2000s. In 2014, an outbreak of ZIKV in French Polynesia was followed by increased reports of GBS. An investigation into the GBS outbreak provided evidence for a possible etiologic association between ZIKV and the cluster of GBS cases [18]. Subsequent reports have further supported this association of ZIKV with severe neurologic sequelae such as GBS and congenital malformations [6, 18–21].
In April 2015, ZIKV was first identified in Brazil, causing an outbreak of exanthematous illness centered in the northeast region [22]. Subsequent to the ZIKV outbreak, clustering of GBS diagnoses was noted in mid-2015 in northeastern Brazil [23–25]. However, to identify risk factors and potential infectious pathogens associated with the reported increase in GBS cases, we performed a case-control investigation to evaluate the relationship between ZIKV and increased reports of suspected GBS. In particular, we sought to establish an etiology for the outbreak of GBS through a case-control investigation using information collected through interviews, and evaluate the relationship of GBS to arboviral infections in the population.
We conducted our investigation in the Salvador metropolitan area during January 16 –February 5, 2016. We identified suspected GBS case-patients reported by physicians and hospitals to the Center for Information and Epidemiologic Surveillance of Bahia (Centro de Informações Estratégicas em Vigilância em Saúde [CIEVS]) with onset of neurologic symptoms during January 1– August 31, 2015. To determine compatibility with a GBS diagnosis, we performed medical record reviews to ascertain characteristics of the clinical illness and diagnostic testing, including cerebrospinal fluid, neuroimaging, and electrodiagnostic test results. Suspected GBS case-patients were classified according to diagnostic certainty of the Brighton Collaboration Criteria case definitions for GBS [26]. Case-patients meeting levels 1–3 of diagnostic certainty, and who were at least 12 years of age at time of interview, were classified as confirmed GBS and eligible for enrollment in the case-control investigation.
We obtained numbers of suspected ZIKV infections in Salvador during January 1 –August 31, 2015, from CIEVS. We also obtained incidence for suspected and confirmed dengue and chikungunya infections from routine surveillance through the National Notifiable Disease Information System (Sistema de Informação de Agravo de Notificação). We juxtaposed these data to evaluate temporal relationships between dengue, chikungunya, and ZIKV infections compared with confirmed GBS cases.
For each GBS case-patient, we selected two neighborhood controls from the same general age grouping (12–19, 20–39, 40–59, 60+) as the case-patients. We did this to ensure a relatively equal age distribution between case-patients and controls given that age is a known risk factor for GBS [27]. To identify controls, we flipped a coin to determine the direction of travel from the case-patient’s house, and we used a random number generator to determine how many properties (1–20) to skip to choose the first house. We continued to move in the same direction until finding the first control, and we repeated the random number selection to find the second control, maintaining the direction of travel.
We interviewed all available case-patients and controls to obtain information about demographics, risk factors, and exposures (S1 Table) in the 2 months prior to onset of neurological symptoms of the GBS case-patients. We considered case-patients and controls as having suspected ZIKV disease if they had self-reported symptoms of rash with at least two other ZIKV-like symptoms: fever, conjunctivitis, arthralgia, myalgia, and peri-articular edema [28]. At the time of interview, to determine intermediate-term functional outcomes, we assessed for residual motor deficits using the Hughes GBS Disability Scale [29]. Following the interviews, serum samples were collected from case-patients and controls.
We tested serum samples by capture enzyme-linked immunosorbent assay for IgM antibodies against ZIKV (MAC-ELISA) and dengue viruses serotypes 1–4 (DENV Detect IgM Capture ELISA, InBios, Inc., Seattle, WA) [30]. We determined neutralizing antibody titers against ZIKV and dengue serotypes 1 and 2 using a 90% cutoff value for plaque-reduction neutralization tests (PRNT90) [31]. We defined a recent flavivirus infection as a positive or equivocal IgM test result for ZIKV or dengue. We discriminated between ZIKV and dengue infections if only one PRNT was positive (Table 1) [32].
We estimated the incidence of GBS using 2015 population estimates by the Brazilian Institute of Geography and Statistics (Instituto Brasileiro de Geografia e Estatística) [33]. To determine a possible association between GBS and a preceding ZIKV infection, we estimated that 37 case-patients and 74 controls would provide a power of 80% to detect a difference of 30% in ZIKV prevalence, with an alpha level of 5%. We performed logistic regression to calculate odds ratios and 95% confidence intervals for the association of GBS and demographics, known GBS risk factors, antecedent illness, and dengue and ZIKV infections. We also performed sensitivity analyses on combinations of symptoms and laboratory findings to evaluate consistency of results.
The human subjects review board at CDC and the Brazil Ministry of Health approved the investigation and determined it to be part of a public health response and not research [National Council of Ethics in Research (Conselho Nacional de Ética em Pesquisa) approval number 1.391.200]. All adult subjects provided informed written consent prior to interview participation and collection of specimens, and a parent or guardian of any child participant (under 18 years old) provided informed consent on their behalf.
During January 1 –August 31, 2015, 77 suspected GBS case-patients within the Salvador metropolitan area were reported to CIEVS. We reviewed all available medical records, and 50 (65%) patients had sufficient information to be classified as levels 1–3 of the Brighton Collaboration criteria for GBS: 7 (9%) met level 1 of diagnostic certainty, 43 (56%) met level 2, and none met level 3. Of the remaining 27 individuals, 24 (31%) did not have enough information for classification as GBS or had alternative diagnoses. The remaining 3 (4%) were excluded because of factors such as age below enrollment criterion or inaccurate address.
The majority (94%) of Brighton-confirmed case-patients based on medical chart review had neurologic illness onset during epidemiologic weeks 17–29 (April 26 –July 25) (Fig 1). During this period of peak GBS occurrence, the annualized incidence of GBS was 5.6 cases/100,000 population for the Salvador metropolitan area. Annualized age-group-specific incidence increased with age. The incidence was 1.5 cases/100,000 population for the 12–19 years age group, 3.9 for the 20–39 years age group, 7.3 for the 40–59 years age group, and 14.7 among persons ≥60 years of age. The median age of these cases was 47 years (range, 14–79 years). There was no significant difference in incidence between men and women (5.8 versus 5.5, p = 0.37).
Based on medical chart review of the 50 confirmed GBS case-patients, 44 were reported to have had a preceding illness (S2 Table); the median time between antecedent illness and neurologic symptom onset was 8 days (IQR 5–15). Prominent neurologic signs/symptoms of the GBS case-patients included leg and arm weakness, dysphagia, and facial weakness. Median time from onset of neurologic symptoms to nadir was 6 days (IQR 4–9). Nine case-patients had electrodiagnostic studies available for review. Of the available reports, 5 were interpreted as being consistent with the acute motor axonal neuropathy (AMAN) subtype of GBS, and the other 4 demonstrated patterns interpreted as the acute inflammatory demyelinating polyradiculoneuropathy (AIDP) subtype of GBS. All case-patients were hospitalized; 46 (92%) received intravenous immunoglobulin (IVIG), 17 (34%) required ICU-level care, 11 (22%) required mechanical ventilation, and 3 (6%) died (Table 2).
The reports of suspected symptomatic ZIKV infections occurred as a prominent clustering during April–June of 2015 while no apparent fluctuations were reported in either dengue or chikungunya infections throughout 2015. The outbreak of GBS peaked approximately 7 weeks after the peak of suspected ZIKV infections (Fig 1).
Of the 47 GBS case-patients who were alive at the time of the investigation, 2 declined participation and we could not locate 4, leaving 41 individuals that we could further evaluate through interview. The median age for these case-patients was 44 years (range 14–78), which was not significantly different from the controls with a median age of 50 (range 13–87). Additional demographics and exposure histories did not differ between case-patients and controls (S1 Table) with the exception that a higher proportion of GBS case-patients compared to controls reported an antecedent illness in the 2-month period prior to neurologic symptom onset of GBS case-patients (Table 3). Symptoms most frequently reported by GBS case-patients included rash, headache, fever, myalgias, and arthralgias.
At the time of our assessment, all living GBS case-patients were at least 5 months out from neurologic symptom onset (median 220 days, range 160–321 days). Thirty-five case-patients (85%) reported at least minor residual motor deficits; 17 (41%) had substantial residual motor deficits and could not walk without assistance.
Case-patients were found to meet criteria for suspected ZIKV disease significantly more often than controls (Table 3). There were no case-patients or controls who met laboratory criteria for recent or prior ZIKV infection or recent dengue infection. Recent flavivirus infections were equally prevalent between case-patients and controls. However, being a case-patient was significantly associated with evidence of recent flavivirus infection when combined with clinical criteria for suspected ZIKV disease. Only a small number of case-patients and controls had no evidence of prior ZIKV exposure based on negative PRNTs. Nearly all of the samples tested positive for dengue (serotypes 1 and 2) virus-neutralizing antibodies (100% of case-patients versus 96% of controls), leaving only 3 controls with no evidence of prior exposure to ZIKV, dengue, or other flaviviruses.
Our investigation demonstrated a high incidence of GBS geographically and temporally clustered in the setting of an ongoing large outbreak of ZIKV. GBS case-patients in this investigation were more likely than non-GBS controls to report symptoms suggestive of ZIKV illness in the 6–8 weeks prior to neurologic illness onset; in addition, based on both medical record review and self-report, the distribution of onsets of the antecedent illnesses clustered during the 2 weeks before the onset of neurologic symptoms. This temporal relationship supports an etiological association between these two illnesses. Despite this epidemiological evidence, the serologic data could not confirm this relationship. However, the serologic findings from specimens collected an average of 7 months after GBS onset showed a high prevalence of ZIKV-neutralizing antibodies in cases and controls, consistent with the ZIKV epidemic that was recognized in the community during the preceding months [22].
Baseline population estimates of GBS incidence in Brazil are limited but are reported at 0.05–0.6 cases/100,000 people per year, which are substantially lower than expected [34–36]. The incidence of GBS in North America and Europe is 0.81–1.89 with an expected worldwide incidence of 1.1–1.8 cases per 100,000 people per year [27, 37]. Applying a baseline estimate of 1.5 cases per 100,000 people per year results in a 3.7-times increased incidence of GBS for the outbreak period for persons at least 12 years of age.
The characteristics of GBS illness during this outbreak were largely similar to what would be expected for typical GBS disease patterns with some notable exceptions. The high attack rates in the older individuals reflects an unusually steep increase in GBS incidence during the outbreak. The incidence of GBS in this investigation is 10-times higher among the oldest age group compared with the youngest, in contrast to a 2–3-times increase in incidence reported elsewhere for the same population groups [27]. Indicators of GBS severity, such as need for intensive care monitoring and mechanical ventilation, were comparable to other reports of GBS [1]. Overall, the 6% mortality rate was similar to rates in North America and Europe and possibly reflected the high utilization of IVIG and supportive care of the GBS case-patients in this outbreak. Most of the case-patients regained lost function after their GBS illness; however, 41% of case-patients still required assistance with walking 6 months later, which is higher than the 20% that is more commonly expected [38]. Additionally, there was a more rapid progression to nadir than the 2–4 weeks usually observed for GBS, though this was similar to findings in French Polynesia [18, 39, 40].
Few studies exist that characterize the subtypes and electrophysiologic findings of GBS in Brazil [41]. Though there was a limited number of electrophysiologic studies performed for case-patients in this investigation, there appeared to be a relatively equal distribution of AMAN and AIDP subtypes. This contrasts with French Polynesia where AMAN was the predominant subtype and with Puerto Rico and Colombia where AIDP was most frequently reported [18, 20, 21]. Additional investigations are required to define the electrophysiologic features of ZIKV-associated GBS, which could contribute to understanding the underlying pathophysiologic mechanisms.
We noted a 7-week interval between the peaks of reported ZIKV infections and GBS, which is longer than would be expected if ZIKV was biologically associated with GBS. This finding was consistent with data reported elsewhere [25]. However, this differed from individual-level data, in which the median interval between onset of ZIKV-like illness and GBS was 8 days. Several factors may have contributed to this effect. One possibility suggested by Paploski, et al. [25] is that once the community perceived ZIKV infections as benign, persons may have stopped seeking care, artificially foreshortening the epidemiologic peak of ZIKV infections. Alternatively, there may be limitations in the surveillance data for ZIKV infections given that very few case-patients had laboratory confirmation. This is especially notable considering recent reports that more closely align occurrences of ZIKV infections and GBS in Bahia [6].
The preceding illnesses of the case-patients were most prominently characterized by rash, headache, fever, myalgias, and arthralgias—symptoms commonly reported with ZIKV infections [42]. The occurrence of acute illness among the case-patients 8 days prior to onset of GBS provides a biologically plausible argument for a causal association between the acute illness and GBS and has been similarly reported in other studies of Zika-associated GBS [20, 21]. We found that there was also a significant, but much less robust, association for gastrointestinal symptoms, such as nausea/vomiting and diarrhea. Gastrointestinal symptoms have been reported in the setting of other ZIKV outbreaks [20, 43], and it is possible that these symptoms are an under-recognized clinical feature of ZIKV illness, rather than manifestations of Campylobacter infection leading to the GBS cases observed in this investigation. However, it is probable that not all of the case-patients had the same antecedent etiology, which is supported by the fact that several of the case-patients did not report symptoms of Zika-like illness or have laboratory evidence of previous ZIKV infection.
The attack rate of ZIKV in northeastern Brazil is unknown. However, if it was as high as reported in French Polynesia, it could limit our ability to detect differences in seropositivity between case-patients and controls. Tests of association cannot discriminate between such high rates of infection without a very large sample size. Notwithstanding, the laboratory data can be used to demonstrate likely ZIKV exposure, which hence could be presumed to represent a relatively recent infection since the virus has only been identified in Brazil since early 2015. Additionally, given this finding of greater ZIKV-like symptoms in the case-patients, even if rates of ZIKV infection are not different between the case-patients and controls, it may indicate that individuals with symptomatic ZIKV infections are more predisposed to the development of GBS. It is not yet understood what factors lead to symptomatic versus asymptomatic infections, though possibilities may include prior infectious exposures causing potentiation, initial infectious viral load, or host immunologic response. The latter theory is supported by this observed correlation with GBS, which is itself a manifestation of an aberrant immune response. Further exploration of the immunopathogenesis of ZIKV may provide insight into the mechanism of Zika-associated GBS.
Because patients for whom GBS case status was not ascertained were not interviewed, we did not systematically collect demographic information on them, so no comparisons could be made with case-patients enrolled in the investigation. The major limitations in antecedent illness analysis are the non-specific nature of the symptoms, possible underreporting of acute illnesses by controls, and recall bias in reporting of symptoms by case-patients. Because the investigation was performed approximately 7 months after the acute illness onset, laboratory findings were limited to serology. Because there is no information on how long ZIKV-specific IgM persists, the finding of a negative IgM result so far out from the initial suspected ZIKV infection is difficult to interpret. Furthermore, the cross-reactivity of dengue and ZIKV antibodies makes accurate discrimination between these pathogens particularly challenging since all samples with ZIKV-specific neutralizing antibodies also had dengue virus-specific neutralizing antibodies [32, 44], though measuring ZIKV-specific neutralizing antibodies several months from the acute infection may be a more specific indicator [45]. Regardless, the laboratory findings more effectively document flavivirus exposure rather than being able to confirm Zika infection or exposure, and other possible infectious triggers cannot be ruled out; in particular, Campylobacter-specific surveillance data was not available and could not be retrospectively ascertained. Additionally, ascertainment of neurological impairment was based on retrospective abstraction of medical records, which may be limited by lapses in documentation.
The findings of this investigation, along with similar epidemiologic findings throughout Latin America and previously published data from French Polynesia [6, 18, 20, 21, 25], strongly suggest a ZIKV-associated GBS. The apparent etiological association between ZIKV and GBS and the observed higher ZIKV-related GBS attributable risk in older persons should be substantiated through additional prospective studies of GBS during ZIKV outbreaks. These studies should include molecular confirmation of infection to further define distinctive clinical and laboratory features of such cases. If this etiologic relationship is true, prevention of GBS may be enhanced by minimization of exposure to mosquitoes through personal protection and environmental control methods. It may also be prudent to target public health messaging about GBS to older adult populations during ZIKV outbreaks. Additionally, these findings may inform future preparedness efforts to build GBS-related diagnostic, treatment, and hospital capacity in areas at risk for ZIKV infection.
|
10.1371/journal.pgen.1001242 | Continuous and Periodic Expansion of CAG Repeats in Huntington's Disease R6/1 Mice | Huntington's disease (HD) is one of several neurodegenerative disorders caused by expansion of CAG repeats in a coding gene. Somatic CAG expansion rates in HD vary between organs, and the greatest instability is observed in the brain, correlating with neuropathology. The fundamental mechanisms of somatic CAG repeat instability are poorly understood, but locally formed secondary DNA structures generated during replication and/or repair are believed to underlie triplet repeat expansion. Recent studies in HD mice have demonstrated that mismatch repair (MMR) and base excision repair (BER) proteins are expansion inducing components in brain tissues. This study was designed to simultaneously investigate the rates and modes of expansion in different tissues of HD R6/1 mice in order to further understand the expansion mechanisms in vivo. We demonstrate continuous small expansions in most somatic tissues (exemplified by tail), which bear the signature of many short, probably single-repeat expansions and contractions occurring over time. In contrast, striatum and cortex display a dramatic—and apparently irreversible—periodic expansion. Expansion profiles displaying this kind of periodicity in the expansion process have not previously been reported. These in vivo findings imply that mechanistically distinct expansion processes occur in different tissues.
| Huntington's disease (HD) is a genetically determined neurodegenerative disorder identified by the presence of a mutation for a long series of CAG repeats (>36 repeats) in the Huntingtin (HTT) gene. Longer repeat sequences cause disease onset at a younger age. The mutation encodes an expanded glutamine tract within the huntingtin protein. This enlarged polyglutamine fragment in the protein leads to the formation of the huntingtin aggregates that are observed in HD brains. The stretch of CAG repeats expands with age in affected brain areas, increasing the length of the polyglutamine tract, and is believed to amplify the effect of the disease. Several HD mouse models display phenotypes relevant to the human disease. We have investigated the rate and modes of expansion in striatum, cortex, and tail in transgenic R6/1 mice. Tail was included as a stable tissue, however we observed a small continuous expansion of CAG repeats in tail tissues. In brain tissues, we identified a periodic expansion process consisting of predominantly seven repeat steps. Our findings point towards a very controlled molecular mechanism as the cause of expansion in the most severely affected tissues, which may provide useful targets that can be used to inhibit disease development.
| Huntington's disease (HD) is a genetically determined neurodegenerative disorder, the onset of which is known to depend upon the length of glutamine-encoding CAG-repeat sequences lying within the Huntingtin (HTT) gene [1]. Humans may develop the disease if they have more than 36 repeats and disease onset usually starts during mid-life. An inverse relationship has been shown between CAG repeat length and age of onset in HD [2]–[5]. Additionally, somatic instability in human cortex has recently been shown to be a good predictor of disease onset [6]. Children with 108–256 CAG repeats are reported to show disease onset from one and a half years to six years of age [7].
Trinucleotide repeat (TNR) instability varies between organs in a variety of neurodegenerative disorders which are caused by expansion of CAG repeats in a coding gene, with the greatest instability observed in the brain [8]–[11]. In HD, striatum tissue shows the most severe neuropathology, followed by cortex. CAG length expansion is correlated with neuropathology and probably precedes the onset of symptoms [12]. The CAG repeat length is unstable in most cell types of the brain, but neurons tend to show the greatest mutation lengths in both humans and mice [13]–[15]. Meanwhile, minimal expansion is considered to occur in many other somatic tissues.
TNR sequences may form slipped strands during replication or repair, creating loops or hairpins, which protrude from the DNA duplex [16].
In the earliest model for repeat expansion the DNA polymerase forms slip-outs on the nascent strand leading to small-scale repeat expansion in repetitive sequences [17]. Loops of repeat-containing DNA are believed to cause either expansions or contractions during replication, when the slip-out occurs in the nascent or template strand, respectively [18], [19]. Several models have been suggested to explain TNR expansion during replication, such as folding of the lagging strand template into a hairpin, stalled replication forks and the orientation of the TNR in the genome, as well as the location of the origin of replication, as shown in several experiments in bacteria, yeast and human cells (Reviewed in [20]). More recently, a pertinent role of DNA repair proteins in CAG repeat expansion has been demonstrated in vivo. In particular, deletion of the mismatch repair (MMR) proteins, Msh2 and Msh3 [21]–[24] has been shown to abolish age-dependent somatic CAG repeat expansion in mouse models for HD. MMR has also been shown to be involved in TNR expansion in mouse models of myotonic dystrophy (DM1) [25]–[27]. Furthermore, the age-dependent expansion of TNR sequences in somatic cells was shown to be modified by the base excision repair (BER) 8-oxoguanine DNA glycosylase (Ogg1) in the R6/1 mouse model, demonstrating that there may be a link between oxidative DNA damage and TNR instability [13]. The flap endonuclease 1 (FEN1), which removes 5′-flaps during replication [28] and is involved in long-patch BER [29] is also implicated in expansion. Secondary TNR structures have been shown to inhibit FEN1 activity [30]. In addition to flap endonuclease activity, the EXO [31] and GAP activities of FEN1 have been shown to contribute to the resolution of TNR secondary structures in vitro [32]. Recently, it was shown that the stoichiometry of BER proteins, such as Ogg1, polymerase β and FEN1, may contribute to the tissue-selectivity of somatic HD CAG repeat expansion [33].
Nevertheless, the processes causing this expansion remain poorly understood, particularly in mammalian systems, although the formation of secondary DNA structures within the repeat sequence is thought to underlie the process [34]. Here we present evidence for two distinct modes of somatic expansion identified by the analysis of CAG repeat fragments from 103 HD R6/1 mice; a continuous slight expansion in tail, lung, heart and spleen, and a dramatic periodic expansion in striatum, and cortex, which we also compare to the expansions observed in liver. The continuous expansion process is shown here to conform to a bi-directional, forward-biased model that represents the occurrence of multiple short – tending towards unitary – CAG repeat insertions and deletions, at random moments as the mouse ages. In contrast, the dramatic expansion seen in brain tissues demonstrates a periodicity centred around seven repeats, which correlates with the stochastic insertion of stable TNR segments of consistent length. Meanwhile liver tissue shows a comparable average increase in CAG repeat length to striatum, but with a much weaker inclination to exhibit periodicity, tending more towards a continuum. This suggests either a much less controlled insertion length when compared to expansions in striatum, or that liver tissue undergoes both types of expansion simultaneously. We also present discursive models for these two expansion mechanisms. Identification of these two independent modes of expansion, in particular the tight mechanistic control implicit in the expansions within neuropathologically relevant tissues, increases our understanding of the tissue-dependent progress of HD. This brings us a step closer to inferring the in vivo mechanisms of the molecular components involved, by showing that only a limited selection of the existing models for expansion are able to explain the age-dependent CAG repeat expansions we observe.
In order to understand the mechanisms underlying somatic CAG instability, 42 R6/1 HD exon 1 transgenic mice were sacrificed at either 10 or 21 weeks of age, whereupon tail, heart, lung, spleen, liver, cortex and striatum samples were taken for analysis of HD CAG repeat length. A tail biopsy at 3 weeks of age represents the reference level of CAG repeats present near birth in all tissues for each mouse [35] (Figure S3). Thus changes in the CAG composition of tissues in an individual mouse could be compared over a 7- or 18-week period. A slight expansion was observed in tail (Figure 1A), whereas cortex and striatum demonstrated a dramatic and periodic expansion process, with no significant difference between genders (Figure 1B and 1C). Liver demonstrated an equally rapid, but apparently more continuous expansion. Heart, lung and spleen displayed a slight expansion that was identical to tail (Figure S11). A parallel dataset from 61 hHD+/−Fen1E160D/E160D mutant mice, in which flap endonuclease activity of FEN1 is reduced to ∼20% [36] was included in the study. Reduced FEN1 endonuclease activity did not affect the rate of CAG repeat expansions measured in any tissues, implying that this mutation did not affect any role FEN1 plays in expansion. The two datasets were qualitatively and quantitatively identical with regard to the following analysis in all organs tested and were therefore combined, such that our analysis covers observations across two HD genotypes, reinforcing the ubiquity of the results.
Individual fragments from tail fit well to a normal distribution and are thus described by the mean (μ) and standard deviation (σ) of the curves fitted to raw data (Figure 2A, 2B; see Methods). Expansion within the population of 59 mice (10 week old mice excluded) is clearly shown (Figure 2C) by the relative difference in μ of the 21-week and 3-week groups. The median expansion found in 59 tails of 21-week mice was 1.97 CAG triplets (Figure 2E). In addition, σ of individual tail data sets is shown to increase from 1.98 triplets at 3-weeks to 2.87 triplets at 21-weeks (Figure 2D). The increasing σ is not an artefact of PCR errors, as is demonstrated in Figure 3A.
Having made these observations, it is necessary to consider them in the context of potential models for expansion, in order to fully investigate their implications and attempt to parameterize the processes involved. A continuous increase in both mean and standard deviation can be generally accounted for by multiple stochastic unitary (single CAG-repeats) extension and contraction events on the CAG tracts within the sample (Figure 3B). A full discussion of the potentially applicable models and considerations is presented as supporting information (Figure S8, Text S1, and Videos S1, S2, S3, S4, S5, S6, S7, S8) and we confine ourselves here to a simple application of the most probable hypothesis, yielding upper estimates for expansion and contraction rates. Assigning probabilities to non-simultaneous unitary expansions and contractions, and respectively, allows the measured temporal change in the mean () and variance () of tail samples to be defined by (1) and (2).(1)(2)Using a time interval () of one day, the measured expansion of 1.97 repeats and the concomitant increase in average standard deviation from 1.98 to 2.87, the values of and are calculated to be 0.026 and 0.010 respectively. This gives a maximum expectation of ∼0.036 (pe+pc) events per repeat tract per day.
In contrast to fragment data from tail, heart, lung and spleen, raw data from 10- and 21-week cortex and striatum tissue show a peak retained at the 3-week repeat level alongside an age-dependent number of periodically spaced subsequent peaks (Figure 4A, Figures S1, S2, and S10), to which a series of normal distributions were fitted (see Methods). Knowing that the relative areas of overlapping distributions define the proportion of each mean CAG repeat length present (demonstrated with mixed samples and serial dilutions of a 21-week striatum sample in Figures S4 and S5), we infer – on account of the regularity of the intervals between neighbouring peaks at both 10- and 21-weeks of age – that expansion involves a proportion of the brain tissue undergoing insertions of consistent-length CAG repeat fragments over time. If expansion events inserted CAG fragments of uncontrolled length, the clarity of subsequent peaks would be lost. Likewise if different cell-types within one sample expanded at different rates, one would expect a continuum of peak separations in the collected accumulated data from many mice, and would have little reason to expect a consistent periodicity between peaks at 10-weeks and 21-weeks. This argues for the stochastic step-wise insertion of CAG fragments with an average length μb−μa (Figure 4A (10wk and 21wk) and 4B) by a mechanism that may recur within the same cell. To measure the periodicity seen in brain tissues, we compiled histograms of all intervals between identifiable peaks, binned by size, from the individual cortex and striatum samples of mice aged 10- and 21-weeks (Figure 4C). The measured intervals are clearly shown to be distributed around a peak at 7 CAG repeats, with a mean length of 7.14 (σ = 1.78 with a cut-off for doublet measurements set at intervals ≥12). The median interval of 7 also confirms that this expansion process is centred around the insertion of 7-repeat fragments into the CAG tract, although the width of the starting distribution indicates insertion of 5 to 9 repeat-fragments (see Figure S9 and Videos S1, S2, S3, S4, S5, S6, S7, S8 for further discussion). The relative sizes of peaks in 21-week striatum (see Methods) imply that on average, a total of ∼10,000 7-repeat insertions occurred in each dramatically expanding striatal sample (on our timescale probably mainly within neurons; however, glial cells also undergo expansion and not all neuronal cells are guaranteed to show expansion [14], [15], giving a periodic expansion probability estimate of 0.018 events per repeat tract per day. The fact that efficiency of PCR amplification of longer CAG tracts is reduced, may result in some measure of underestimation with this value. This is ∼70% of the estimated probability for unitary expansion events in tail, however the 7-repeat average insertion size renders the resulting expansion more dramatic in striatum. In contrast, liver data while showing comparable levels of average expansion to brain tissues, lacks a clear signature of periodicity (see Figure 1D) tending towards bimodality, with a more continuously located second peak. This would imply a much less controlled insertion length during the expansion process, or possibly a combination of expansion mechanisms.
Having shown data and analysis to define these two distinct modes of expansion, we place our findings into the context of existing literature, in order to develop reasonable hypothetical models for these two types of expansion.
The continuous expansion we observe shows a progressive increase in CAG tract length, which is comparable with the expansions observed in fibroblasts derived from an adult HD mouse [37]. In the debate regarding the relative roles of replication and repair in TNR instability, these results present an interesting question, since the continuous expansion process occurs in organs containing dividing cells. A process occurring with the previously calculated expectation rate of ∼0.036 events per repeat tract per day on the ∼360 nucleotide CAG segment of the 2.5 gigabase mouse genome, would correlate to ∼250,000 genome-wide events per cell per day. This is well above the upper estimate for daily DNA damage events, making it unlikely that these are the sole initiator of expansions in tail tissue. Several replication-based models for TNR expansion in dividing cells [18], [19] have been proposed. The potential for replication to be entirely responsible for this expansion is considered in detail elsewhere (Figures S7, S8, and Text S1) and is considered to be unlikely, particularly in light of the fact that lymphoma tissues (with necessarily higher replication rates) isolated from several mice showed no increase in TNR instability (Figure S7). Therefore we infer that other mechanisms must also prevail, and propose that slipped strand structures [16], [38] generated by out of register rehybridization of CAG repeats [27] during transcription, or genome maintenance, may spontaneously form unstable loops or cruciforms which may subsequently stabilize by migrating apart. Similar small loops can also be formed by polymerase slippage during replication [17]. We propose that two separate pathways may repair these loops, leading to single repeat expansions or contractions (Figure S8). Further research is needed to resolve the specific origin of this mode of expansion.
A hypothesis for periodic expansion is also presented (Figure S9). The previously calculated periodic expansion probability of 0.018 events per repeat tract per day would correlate to ∼100,000 genome-wide events per cell per day. This lies in the vicinity of upper estimates for accumulated oxidative DNA damage [39]. It is therefore possible that DNA damage contributes as catalyst for periodic expansions in brain tissues. However, oxidative damage is not sufficient to trigger somatic instability [33]. Of particular interest here, are the potential molecular components that could repeatedly generate a regularly sized repeat insertion that is dominantly seven repeats in length. We have therefore chosen to briefly review the relevant literature in search of further insight.
Theory suggests that CAG flaps ranging from 4 to 16 triplet repeats in length can form thermodynamically unstable hairpin structures with an even number of repeats [40]. However, under physiological conditions, 6 triplet repeats have been shown to form hairpins irreversibly [41], [42]. This implies that a progressively generated triplet repeat flap can stabilise into a 6-repeat hairpin at the free end that could be cleaved by FEN1 causing no net expansion [43] (Figure S9). However, the presence of metastable intermediates during flap generation may allow the flap length to increase beyond 6 repeats before a stable structure is formed, thereby producing a hairpin with an overhanging CAG at the 5′-end. This free CAG repeat can hybridise back to the DNA duplex, and it has been shown by Liu et al. [31] that such hybridisation facilitates bubble formation followed by ligation and expansion. In a few instances, two or even three free CAG repeats in the 5′-end of the hairpin would produce a periodicity of eight or nine CAG repeat steps. In addition, the size of the hairpin could vary with a few repeats. However, this occurs more rarely, as observed in Figure 4C. After gap filling and ligation, a loop of excess CAG repeats in one strand would be produced (Figure S9) that can be recognized by the MMR complex Msh2-Msh3 with high affinity [44]. This binding might further stabilize the CAG loop and additionally explain why Msh2-Msh3 is necessary for large expansions to occur in striatum [21]–[24], although the role of MMR in causing TNR expansion is not understood. Moreover, Msh2-Msh3 function ceases due to impaired ATPase activity on loops exceeding 16 nucleotides in lengths [45] - probably due to the presence of A•A mispaired bases in the loop [22] - which may explain why the MMR system fails to repair longer extraneous CAG loops. Subsequently, a nick generated on the strand opposite to the loop could result in faulty repair of the CTG strand along the CAG slip-out, causing expansion by a repair process independent of MMR [46] (Figure S9). Thus, oxidative damage on the CTG strand could result in base excision repair (BER) and Ogg1 moderated expansion, which is also in concordance with the proposed ‘toxic oxidation cycle’ by Kovtun et al., 2007 [13]. We therefore propose that a coincidental cooperation can occur between MMR and BER in cases where a long CAG flap is able to stabilize itself, which can form the basis of the consistent periodic insertion we observe.
It should be pointed out that it is possible that other stabilized structures, such as loop-outs or a stabilizing interaction with one of the many proteins and complexes that are in contact with the DNA may also serve to cause the observed periodicity. The majority of the potentially applicable models are covered in the literature [13], [24], [27], [34]. Further work is necessary to resolve this completely.
Cell proliferation in neurons and glial cells has been observed in the subependymal layer adjacent to the caudate nucleus in human HD brains [47]. However, polymerase slippage usually forms small expansions [17], and the repetitive uniformity of the periodic expansion makes it unlikely that polymerase slippage is responsible for the dramatic expansions seen in cortex and striatum. Furthermore, the lack of periodically spaced peaks containing fewer repeats than the mice were born with – as would be expected from a bi-directional process - means that 7-repeat hairpin-based contraction events occur either at a negligible rate in comparison to expansions, or do not occur at all. During the 18 week period, tail-type expansion events are not evident in striatum since they would obfuscate latter periodic peaks (Figure S6). Therefore, the two expansion mechanisms seem to be either entirely independent – not occurring simultaneously in the same cell – or that if they do share common elements, they progress along mutually exclusive pathways. However, it is important to notice that both expansion mechanisms must be dependent on proteins from the MMR complex, since expansions are eliminated in all tissues in either Msh2 or Msh3 nulls in HD [21]–[24], and also in DM1 [25], [26]. To date, we have not managed to define the individual cell types that are specific to these expansion mechanisms. It will be of great interest to compare the expansions observed in animal models to those in cultured HD mouse fibroblasts, as a way of identifying the cell specificity of these modes of expansion.
The liver may be particularly interesting in this regard, since this tissue potentially exhibits a mix of both types of expansion mechanisms. This could be attributed to different modes of expansion occurring in different cell types. Indeed, instability of the DM1 CTG•CAG repeat is known to occur in liver hepatocytes with high DNA ploidy [48].
The question also arises as to why FEN1 did not influence CAG repeat expansion in the organs tested. One might expect a difference, since a recent in vitro study has shown that FEN1, together with long-patch BER of long repeat sequences by polymerase β, promotes expansion by facilitating the ligation of hairpins formed by strand slippage [49]. However, FEN1 flap activity has shown to be circumstantial, with much lower activity in the striatum than in the cerebellum of R6/1 mice [33]. In yeast the capture of flap structures by FEN1, rather than the endonuclease activity, is the most important function of FEN1 in preventing TNR expansion [50]. EXO- and GAP activity of FEN1 are also involved in in vitro triplet repeat expansion in yeast [32], and these activities are probably not influenced by the Fen1E160D/E160D mutation. Therefore, it seems that the 20% endonuclease activity [36] of the Fen1E160D/E160D mutation does not affect the rate of CAG repeat expansions. In concordance with our finding Fen1 did not control instability of (CTG)n*(CAG)n repeats in a knock-in mouse model for DM1 [51].
It is worth considering briefly why this periodicity has not been described before, since there are numerous potential reasons. One possibility could be that the mice used in the present study exhibit more instability due to environmental factors [52] or genetic background [53]. Perhaps the periodic signal becomes more disperse in older mice used elsewhere; degrading the quality of the data, and that the age-range, as well as the relatively long starting CAG lengths, of our samples is optimal for observing this periodicity. Another possibility is simply that later versions of the GeneMapper system are more sensitive, in comparison to the GeneScan method applied in some older studies, allowing us to see more detail. While periodicity has not featured in other studies of similar tissues and disease models, it is difficult to state unequivocally that it was unobservable in their data. The small volume of data presented in articles, uncertainty over the precise PCR conditions used and the simple fact that periodicity was not the focus of these investigations can be sufficient cause for this phenomenon to have been previously overlooked. There is some variability among replicate striatum samples as shown in Figure S10. This could be explained by sampling error or polymerase slippage in early PCR cycles. Sampling error is however unlikely to be the reason behind the periodicity as explained in Figure S12 and Text S2.
So far, we have only studied periodicity in the R6/1 mouse model and without specific studies of other HD CAG mice models the generality of our data is unknown. Yet, the R6/1 transgenic mouse is a widely accepted and commonly used model for human HD that exhibits a progressive neurological phenotype that exhibits many of the features of HD [54]. HD CAG repeat instability has shown to be similar in humans and mice, with the longest expansion lengths occurring in striatum, followed by cortex, and little expansion in cerebellum and most other tissues [8], [14], [35]. In addition, the HD CAG repeat length appears to be expand most in neuronal cells rather than glial cells in both species [14], [15]. Due to the long starting CAG repeat length in the transgenic mouse, the model may be most relevant as a model for juvenile HD. Given the stated similarities, there are grounds to suspect that the mechanisms of expansion are identical in mice and humans, only occurring at an accelerated pace in the mouse on account of the long repeat tract. In this case, the mouse model would function as a good model for human cases with mid-life age of onset as well. However, periodicity has not previously been reported in HD patients. A possible explanation could be that the repeat length in the R6/1 mouse is longer than the repeat length that has been analyzed in any HD human tissue with regard to somatic instability. The genomic localization of the randomly integrated HD gene fragment in R6/1 mice might modulate the CAG stability. Furthermore, CAG instability in HD patient brain cannot be analyzed at an early time-point that would allow for a direct comparison to the R6/1 data.
Importantly, CAG repeat expansion in human cortex is associated with an earlier age of disease onset, in addition to the role of the constitutional CAG repeat length [6]. This implies that there are disease modifiers that influence somatic instability, and conversely, factors that determine somatic instability which may modify disease pathogenesis. It is therefore critical to understand the mechanism of the different expansion processes and the factors involved.
To summarize, we present two different mechanisms of somatic CAG repeat expansion; a continuous bi-directional expansion in tail, lung, heart and spleen tissues, and a dramatic periodic expansion centred around 7-repeat insertions in striatum and cortex. Further experiments are needed to determine whether the 7-repeat step-size is independent of CAG tract length and species and it remains to be shown whether these two models can explain the expansions observed in other brain tissues and organs, as well as in humans. Nevertheless, these results provide significant new insights into in vivo expansion mechanisms, which may also be relevant to other triplet repeat disorders.
All animal experiments have been approved by the local and national animal - and are carried out according to the regulation by FELASA (Federation of European Laboratory Animal Science Associations).
B6CBA-Tg(HDexon1)61pb/J mice of the R6/1 line [54] with ∼119 CAG repeats in exon 1 of the HTT gene, were purchased from The Jackson Laboratories and either interbred or crossed with C57BL/6J Fen1E160D/E160D mice [36]. The mice were fed Rat and Mouse No.3 breeding diet (Special Diet Services) and tap water ad libitum. At 3 weeks of age the mice were anesthetized by i.p. injection of a combination of Midazolam (Dormicum “Roche”) and Fentanyl/Fluanisone (Hypnorm) solutions, and tail biopsies were taken. DNA from tail biopsies of the first two generations were lysed as described [36] and DNA isolated using standard NaCl precipitation or phenol/chloroform procedure. At 10 and 21 weeks of age the mice were sacrificed by cervical dislocation. The organs were harvested, frozen on dry ice and stored at −70°C. During dissection of striatum we lost 9 samples. DNA from all tissues and tails from the F2 and F3 generation was isolated according to the DNeasy Blood & Tissue kit (Qiagen GmbH, Germany).
DNA from tail biopsies of 3-week old mice were used for genotyping. HD mice were genotyped with forward 5′-cggctgaggcagcagcggctgt-3′ and reverse 5′-gcagcagcagcagcaacagccgccaccgcc-3′ PCR primers [54] according to the Advantage GC 2 PCR Kit & Polymerase Mix (Clontech, CA). The Fen1+/+ and/or Fen1E160D/E160D knock-in allele was genotyped as described [36].
CAG repeats were sized by PCR with primers 5′-FAM-atgaaggccttcgagtccctcaagtccttc-3′ and 5′-ggcggctgaggaagctgagga-3′ according to [54] with slight modifications. Approximately 75ng of genomic DNA (this approximates to DNA extracts from ∼10,000 cells) was amplified with AmpliTaq Gold DNA polymerase with PCR Buffer II, 1.25 mM MgCl2 (Applied Biosystems, CA), and 2.5 mM dNTPs (GE Healthcare). The cycling conditions were 94°C for 10 min, 35 cycles of 94°C for 30 sec, 64°C for 30 sec, 72°C for 2 min, and a final extension at 72°C for 10 min. The FAM-labeled PCR products were mixed with GeneScan - 600 LIZ Size Standard and HiDi Formamide (Applied Biosystems) and run on an ABI 3730 Genetic Analyzer (Applied Biosystems). Sizing of the PCR fragments was performed by using the GeneMapper Software Version 3.7 (Applied Biosystems).
All raw data was processed through a masked Nelder-Mead simplex fitting method, optimising free parameters of standard deviation, mean and amplitude to fit consecutive normal distributions sufficient to account for ≥98% of the total area of the raw data set. In the case of tail data, only a single normal distribution was required. These optimised parameters were returned as the means (μ), and standard deviations (σ), which were used to define the TNR lengths present in each data set. CAG repeat tracts were flanked by sequences 86 bp in length as verified by sequencing. Thus, the mean number of CAG triplets (μt) present in a fragment analysis sample with a measured mean (μm) is defined by μt = (μm−86)/3. When analysing the periodicity present in striatum and cortex data, standard frequency analysis methods are not suitable, therefore our peak fitting method was used to fit consecutive normal distributions to the raw data (Figure 4A). We were unable to perform fitting analysis on 25 striatum samples due to the quality of the PCR product. The means (μa, μb etc) and relative areas (Aa, Ab etc) of each peak were calculated (Figure 4B). The intervals between neighbouring means (S1, S2 etc) were also recorded at both 10 and 21 weeks. The area (A) of each peak was used to estimate the number of cells containing triplet repeats that had expanded with a step size defined by the separation interval (S). The average area of the first peak in all 21-week data (Aa from Figure 4B) was used to estimate the proportion of non-expanding cells in the striatum as 54% (σ = 15.9), implying that approximately 45% of each striatum sample underwent periodic expansions within the first 21 weeks. By comparison, the proportion of dramatic expanding tissue observed in cortex samples was ∼20%. Previous work has shown that the dramatic expansion observed in the striatum of adult mouse brain tissue largely occur in the neuronal cells [15], [14], although slower expansion can be observed in glial cells and we used this to approximate the level of expansion in neuronal cells in combination with an estimate of the average number of expansion events that were measured in 21-week mice.
|
10.1371/journal.pgen.1004728 | An AGEF-1/Arf GTPase/AP-1 Ensemble Antagonizes LET-23 EGFR Basolateral Localization and Signaling during C. elegans Vulva Induction | LET-23 Epidermal Growth Factor Receptor (EGFR) signaling specifies the vulval cell fates during C. elegans larval development. LET-23 EGFR localization on the basolateral membrane of the vulval precursor cells (VPCs) is required to engage the LIN-3 EGF-like inductive signal. The LIN-2 Cask/LIN-7 Veli/LIN-10 Mint (LIN-2/7/10) complex binds LET-23 EGFR, is required for its basolateral membrane localization, and therefore, vulva induction. Besides the LIN-2/7/10 complex, the trafficking pathways that regulate LET-23 EGFR localization have not been defined. Here we identify vh4, a hypomorphic allele of agef-1, as a strong suppressor of the lin-2 mutant Vulvaless (Vul) phenotype. AGEF-1 is homologous to the mammalian BIG1 and BIG2 Arf GTPase guanine nucleotide exchange factors (GEFs), which regulate secretory traffic between the Trans-Golgi network, endosomes and the plasma membrane via activation of Arf GTPases and recruitment of the AP-1 clathrin adaptor complex. Consistent with a role in trafficking we show that AGEF-1 is required for protein secretion and that AGEF-1 and the AP-1 complex regulate endosome size in coelomocytes. The AP-1 complex has previously been implicated in negative regulation of LET-23 EGFR, however the mechanism was not known. Our genetic data indicate that AGEF-1 is a strong negative regulator of LET-23 EGFR signaling that functions in the VPCs at the level of the receptor. In line with AGEF-1 being an Arf GEF, we identify the ARF-1.2 and ARF-3 GTPases as also negatively regulating signaling. We find that the agef-1(vh4) mutation results in increased LET-23 EGFR on the basolateral membrane in both wild-type and lin-2 mutant animals. Furthermore, unc-101(RNAi), a component of the AP-1 complex, increased LET-23 EGFR on the basolateral membrane in lin-2 and agef-1(vh4); lin-2 mutant animals. Thus, an AGEF-1/Arf GTPase/AP-1 ensemble functions opposite the LIN-2/7/10 complex to antagonize LET-23 EGFR basolateral membrane localization and signaling.
| In the nematode, Caenorhabditis elegans, an evolutionarily conserved Epidermal Growth Factor Receptor (EGFR) signaling pathway is required to induce three epithelial cells to initiate a program of vulva development. EGFR on the basolateral membrane is essential to engage and transmit this signal. Here we demonstrate that AGEF-1 and the AP-1 clathrin adaptor complex function with two Arf GTPases to regulate EGFR localization and signaling. In humans, EGFR also localizes to the basolateral membrane of epithelial cells, and excessive EGFR signaling is a major driver of cancer. In C. elegans, we show that loss of AGEF-1 results in an increase in basolateral EGFR localization in the vulva precursor cells, and in sensitized genetic backgrounds, a corresponding increase in vulva induction. While the human AGEF-1 proteins, BIG1 and BIG2, have not been previously implicated in EGFR signaling and cancer, mutations in BIG2 are causal of periventricular heterotopia, a condition whereby neurons fail to migrate to the cerebral cortex during brain development. As migrating neurons require polarized protein localization, BIG2 and AGEF-1 may have similar functions in these polarized cell types.
| C. elegans vulval cell induction requires a highly conserved Epidermal Growth Factor Receptor (EGFR)/Ras GTPase/Mitogen Activated Protein Kinase (MAPK) signaling pathway providing an in vivo model in which to study signaling in a polarized epithelia [1], [2]. During larval development, an equivalence group of six vulval precursor cells (VPCs), P3.p-P8.p, have the potential to be induced to generate the vulva. The anchor cell in the overlying gonad secretes the LIN-3 EGF-like ligand, inducing the closest VPC, P6.p, to adopt the primary vulval fate, and a combination of graded LIN-3 EGF signal and lateral signaling by LIN-12 Notch specifies the neighboring VPCs, P5.p and P7.p, to adopt the secondary vulval fate. Together P5.p-P7.p generate the 22 nuclei of the mature vulva, eight cells from the primary cell and seven from each of the secondary cells. The remaining VPCs, P3.p, P4.p, and P8.p, divide once and fuse with the surrounding hypodermal syncytium (50% of the time P3.p fuses prior to dividing) and thus adopt a tertiary non-vulval fate. Inhibition of LET-23 EGFR signaling causes a Vulvaless (Vul) phenotype in which less than the normal three VPCs are induced. Conversely, increased LET-23 EGFR signaling causes a Multivulva (Muv) phenotype in which greater than three VPCs are induced.
LET-23 EGFR localizes to both the apical and basolateral membranes of the VPCs, though, it is the basolateral localization that is thought to engage LIN-3 EGF and induce vulva induction [3], [4], [5]. A tripartite complex of proteins, LIN-2 Cask, LIN-7 Veli, and LIN-10 Mint (LIN-2/7/10), interacts with the C-terminal tail of LET-23 EGFR and is required for its basolateral localization [3], [4]. Mutations in any component of the complex, or the let-23(sy1) mutation, which deletes the last six amino acids of LET-23 EGFR that are required for its interaction with LIN-7, result in LET-23 EGFR localizing only to the apical membrane and a strong Vul phenotype [3], [4], [6], [7], [8]. The Vul phenotype of lin-2/7/10 mutants or the let-23(sy1) mutant are easily suppressed to a wild-type or even a Muv phenotype by loss of negative regulators of LET-23 EGFR signaling such as sli-1 Cbl, gap-1 RasGAP, rab-7 GTPase, and unc-101 AP-1μ [5], [9], [10], [11], [12]. Thus far, no suppressors of the lin-2/7/10 mutant Vul phenotype have been shown to restore LET-23 EGFR to the basolateral membrane.
UNC-101 and APM-1 are two μ1 subunits for the AP-1 adaptor protein complex, which function redundantly to antagonize vulva cell induction [12], [13]. In mammals, AP-1 localizes to the trans-Golgi network (TGN) and endosomes, promotes formation of clathrin-coated vesicles, and is involved in regulated secretion from the TGN. [14], [15], [16]. In epithelial cells, AP-1 sorts cargo, including EGFR, to the basolateral membrane, which would be inconsistent with AP-1 antagonizing signaling [17], [18]. The small GTPase, Arf1, recruits AP-1 to the TGN and thus facilitates the formation of clathrin-coated vesicles [15], [19], [20]. BIG1 and BIG2 are Sec7 domain containing guanine nucleotide exchange factors (GEFs) for class I Arf GTPases [21], [22], and are required for recruitment of AP-1 to the TGN and endosomes [23], [24]. To date, neither Arf1 nor the BIG1/2 GEFs have been implicated in EGFR/Ras/MAPK signaling.
Here we identify C. elegans AGEF-1, a homolog of yeast Sec7p and the mammalian BIG1 and BIG2 Arf GEFs, as negatively regulating EGFR/Ras/MAPK-mediated vulva induction. We show that AGEF-1 regulates protein secretion in multiple tissues, regulates polarized localization of the SID-2 transmembrane protein in the intestine, and regulates the size of late endosomes/lysosomes with the AP-1 complex in the macrophage/scavenger cell-like coelomocytes. Genetic epistasis places AGEF-1 upstream or in parallel to LET-23 EGFR. We find that the ARF-1.2 and ARF-3 GTPases also negatively regulate LET-23 EGFR signaling. Moreover, our genetics are consistent with AGEF-1 BIG1/2, ARF-1.2 Arf1 and UNC-101 AP-1μ1 functioning together in preventing ectopic vulva induction. It has been 20 years since UNC-101 was identified as a negative regulator of LET-23 EGFR signaling, however its mechanism of action has remained an enigma [12]. Contrary to the role of AP-1 in basolateral sorting in mammalian cells, we demonstrate that AGEF-1 BIG1/2 and UNC-101 AP-1μ1 antagonize the basolateral membrane localization of LET-23 EGFR in the VPCs. Thus, the AGEF-1/Arf GTPase/AP-1 ensemble antagonizes LET-23 EGFR-mediated vulva induction via regulation of LET-23 EGFR membrane localization.
We previously reported that rab-7(ok511), a maternal effect embryonic lethal mutant, strongly suppresses the lin-2(e1309) Vul phenotype [11]. To identify new candidate regulators of LET-23 EGFR trafficking and signaling we conducted a clonal screen for essential suppressors of lin-2(e1309) (see Materials and Methods). In this screen we identified vh4 as a strong suppressor of the lin-2(e1309) Vul phenotype (Figure 1A–D; Table 1, lines 1–4). The vh4 mutation can suppress the 100% Vul phenotype of lin-2(e1309) to 20% Vul, and 30% Muv. In a lin-2(+) background, however, vh4 mutant animals have 100% wild-type vulva induction. Consistent with a potential role in vesicular trafficking, the coelomocytes (macrophage-like scavenger cells) of vh4 mutants accumulate abnormally large vesicular structures (Figure 1E, F). Additionally, vh4 mutants have a dumpy body morphology, uncoordinated movement and ∼50 percent embryonic lethality (Figure 1G, H).
To determine the molecular identity of vh4, we used a single nucleotide polymorphism (SNP) mapping strategy [25]. Genome-wide mapping located vh4 to the right arm of chromosome I, and interval mapping placed vh4 in a 2.75 map unit region between SNPs haw14137 and pkP1071 at positions 20.65 and 23.4 map units, respectively (Figure 2A). We further refined the genomic interval by complementation with chromosomal deficiencies dxDf2 and eDf3. vh4 failed to complement the large deficiency dxDf2, but complemented the small eDf3, indicating that vh4 lies in a 0.9 map unit region (20.65–21.51) containing 27 genes. We found that one obvious candidate, vps-28, was an RNAi suppressor of the lin-2 Vul phenotype [11]. However, vh4 complemented the vps-28(tm3767) deletion allele and no lesion in the vps-28 coding sequence of vh4 animals was detected by DNA sequencing suggesting that vh4 is not an allele of vps-28. Whole genome sequencing revealed a homozygous G to A transition at position 3082 in exon 11 of the agef-1 gene (AAA TTT TTG GAA AAG GGA GAA CTT CCG AAT TTC CGA TTT) that corresponds to a glutamate to lysine substitution in a conserved region of the predicted AGEF-1 protein (Figure 2B). Consistent with vh4 being a mutation in agef-1, we find that agef-1(RNAi) suppresses the severity of the lin-2 Vul phenotype (Table 1, lines 3 and 7) and agef-1(vh4) mutant oocytes have defects in CAV-1 body formation as previously seen with agef-1(RNAi) (Figure S1J–M) [26]. Finally, agef-1(vh4) fails to complement two deletion alleles, agef-1(ok1736) and agef-1(tm1693), resulting in a strong embryonic lethal phenotype. These data indicate that vh4 is a hypomorphic allele of agef-1.
To determine the identity of the large vesicles in agef-1(vh4) coelomocytes, we used GFP tagged endosomal and Golgi proteins. Since the vesicles are presumably in flux, we measured the diameter of the largest GFP-positive vesicle per coelomocyte. We found a modest, but significant increase in the size of vesicles positive for the early endosomal 2×FYVE::GFP and the pan-endosomal RME-8::GFP markers in agef-1(vh4) animals as compared to wild-type (Figure 3A–C and Figure S2A–C). However, the large vesicles in agef-1(vh4) coelomocytes visible by DIC optics correspond to LMP-1::GFP, a marker for late endosomes/lysosomes (Figure 3D–F). This finding corroborates a concurrent study identifying large LMP-1 positive vesicles in the coelomocytes of agef-1(RNAi) animals [27]. LMP-1 is a transmembrane protein, whose mammalian homolog, Lamp1, can transit from Golgi via the plasma membrane and endosomes to the lysosome [28]. Therefore, we assessed the morphology of the Golgi in agef-1(vh4) mutants using a mannosidase II::GFP marker. While there might be a slight increase in the size of the Golgi mini-stacks, they were distinct from the large LMP-1::GFP positive vesicles seen by DIC (Figure S2D–F). Of note, the mannosidase II::GFP strands that appear to interconnect the Golgi mini-stacks in wild-type (30/37 coelomocytes) were largely absent in agef-1(vh4) mutants (6/42 coelomocytes are interconnected). These data show that agef-1(vh4) disrupts endosome and Golgi morphology and possibly trafficking.
To test if agef-1(vh4) coelomocytes have an endocytosis defect we analyzed the internalization of a signal secreted GFP (ssGFP) that is expressed in body wall muscle cells, secreted into the pseudocoelom, and endocytosed by the coelomocytes [29]. We found less ssGFP in the coelomocytes of agef-1(vh4) animals as compared to wild-type (Figure 4A–E). However, we did not detect a significant accumulation of ssGFP in the pseudocoelom of agef-1(vh4) mutants as would be expected for an endocytosis defect. Rather there was a clear accumulation of ssGFP in the body wall muscle cells of agef-1(vh4) animals as compared to wild-type (Figure 4F–J). While this does not rule out a potential endocytosis defect in agef-1(vh4) coelomocytes, it does indicate that agef-1(vh4) mutants have a secretion defect in the body wall muscle cells.
We also analyzed a Yolk::GFP fusion (YP170::GFP) that is secreted from the intestine, and internalized by maturing oocytes [30]. We did not detect a difference in the uptake of YP170::GFP by oocytes (Figure S1A–D), however YP170::GFP levels in the intestine were higher in agef-1(vh4) animals than in wild-type (Figure S1E–I). An independent study also found impaired secretion of yolk in agef-1(RNAi) animals [31]. Thus, agef-1(vh4) mutants have impaired protein secretion from both body wall muscle and intestinal cells.
To understand the role of AGEF-1 in the LET-23 EGFR/LET-60 Ras signaling pathway, we made double mutants with agef-1 and several mutations in core components of the pathway. A gain of function mutation in let-60 ras (n1046) causes a Muv phenotype that can be enhanced by loss of a negative regulator of the pathway [11], [32], [33], [34]. agef-1(vh4) significantly enhances the Muv phenotype of let-60(n1046), consistent with AGEF-1 being a negative regulator of signaling (Table 1, lines 10–11). We performed epistasis analysis to determine at which step of the pathway AGEF-1 functions. We found that agef-1(vh4) strongly suppresses the Vul phenotype of the let-23(sy1) mutant (Table 1, lines 12–13). The sy1 allele truncates the last six amino acids of LET-23 EGFR that are required for its interaction with the LIN-2/7/10 complex, and thus behaves identical to mutations in components of this complex [3], [8]. However, agef-1(vh4) fails to suppress the Vul phenotype of the let-23(sy97) allele that results in a more severe truncation of LET-23 EGFR that blocks signaling to the LET-60 Ras (Table 1, lines 14–15) [7]. We next tested if agef-1(vh4) can suppress the Vul phenotype of lin-3(e1417), a strong hypomorphic allele of lin-3 EGF [35]. We found that agef-1(vh4) partially suppressed the lin-3(e1417) Vul phenotype (Table 1, lines 16–17). These data are consistent with AGEF-1 antagonizing signaling upstream or in parallel to LET-23 EGFR.
SLI-1 Cbl, a putative E3-ubiquitin ligase, and UNC-101 AP-1μ are negative regulators of LET-23 EGFR signaling that also function at the level of LET-23 EGFR [9], [12], [36]. Like agef-1(vh4), mutations in sli-1 Cbl and unc-101 AP-1μ do not cause a vulval phenotype alone, but double mutants cause a synergistic Muv phenotype. Therefore, we tested if agef-1(vh4) is Muv in combination with strong loss-of-function alleles of sli-1 Cbl and unc-101 AP-1μ. We found that agef-1(vh4); sli-1(sy143) animals are strongly Muv, suggesting that AGEF-1 might function in parallel to SLI-1 Cbl (Table 1, lines 18–19). We were unable to identify unc-101(sy108) agef-1(vh4) double mutants segregating from unc-101(sy108) agef-1(vh4)/unc-101(sy108) mothers suggesting that they are zygotic lethal. Thus, we fed L1 larvae RNAi and found that both unc-101(RNAi) agef-1(vh4) and unc-101(sy108) agef-1(RNAi) animals have a strong Muv phenotype (Table 1, lines 20–22). Since there are two AP-1µ genes, unc-101 and apm-1, that are functionally redundant, we cannot conclude whether AGEF-1 functions in parallel to UNC-101, or whether they function together; we favor the later, see below. However, the strong genetic interactions of agef-1(vh4) and mutations in sli-1 and unc-101 further support AGEF-1 functioning at the level of LET-23 EGFR to negatively regulate signaling.
The identification of a putative Arf GEF as a negative regulator of LET-23 EGFR signaling suggests that one or more of the four C. elegans Arf GTPases might also regulate LET-23 EGFR signaling. The mammalian Arf GTPases have been placed in three classes based on homology [37]. To gain a better understanding of the relationship of the C. elegans and human Arf GTPases we undertook a phylogenetic analysis (Figure 2C). From this we conclude that C. elegans ARF-1.2 is homologous to Class I Arfs; ARF-3 is related to both Class I and II Arfs, but clusters with the Class II; ARF-6 is a homolog of the Class III Arf, whereas ARF-1.1 appears to be a Caenorhabiditis specific Arf GTPase that is distinct from the Arf-like Arl GTPases (Figure 2C). We used RNAi and deletion mutants to test each arf gene for suppression of the lin-2(e1309) Vul phenotype. RNAi of either arf-1.2 or arf-3 partially suppressed of the lin-2(e1309) Vul phenotype (Table 2, lines 1–3). The arf-1.2(ok796) deletion mutant was a much more potent suppressor of the lin-2(e1309) Vul phenotype consistent with RNAi being less effective in the VPCs [11; this study] (Table 2, lines 4–5). The arf-3(tm1877) deletion is zygotic lethal and did not permit analysis. However, arf-3(RNAi) into arf-1.2(ok796); lin-2(e1309) animals led to an even stronger suppression of the Vul phenotype comparable to that of agef-1(vh4); lin-2(e1309) (Table 2, line 8). Neither the arf-6(tm1447) nor the arf-1.1(ok1840) deletions were able to suppress the lin-2(e1309) Vul phenotype (Table 2, lines 9–12). These data suggest that ARF-1.2 and ARF-3 function in a partly redundant manner, possibly with AGEF-1, to antagonize LET-23 EGFR signaling.
To test if arf-1.2 was required in the VPCs we generated two transgenic extrachromosomal arrays, vhEx7 and vhEx8, expressing ARF-1.2::GFP under the control of the VPC-specific promoter, lin-31 [38]. Both transgenic lines were able to strongly rescue arf-1.2(ok796) suppression of the lin-2(e1309) Vul phenotype (Table 2, lines 6–7). Thus, ARF-1.2 functions in the VPCs to negatively regulate LET-23 EGFR signaling. We next tested whether VPC-specific overexpression of ARF-1.2::GFP can revert the suppression of Vul phenotype in agef-1(vh4); lin-2(e1309). Both lines, vhEx7 and vhEx8, led to a more severe Vul phenotype when expressed in agef-1(vh4); lin-2(e1309) animals (Table 1, lines 5–6). This suggests that AGEF-1 functions in the VPCs through ARF-1.2 to antagonize LET-23 EGFR signaling.
Having observed a strong Muv phenotype in unc-101(RNAi) agef-1(vh4) and unc-101(sy108) agef-1(RNAi) doubles (Table 1, lines 21–22), we hypothesized that arf-1.2(ok796) would have similar interactions. Indeed, both unc-101(sy108); arf-1.2(ok796) and agef-1(vh4); arf-1.2(ok796) animals have a strong Muv phenotype (Table 1, lines 24–25). Given that agef-1(vh4) is a weak hypomorphic allele and ARF-1.2 and UNC-101 AP-1μ are each functionally redundant with ARF-3 and APM-1 AP-1μ, respectively; these data are consistent with AGEF-1, ARF-1.2 and UNC-101 AP-1μ functioning together to negatively regulate LET-23 EGFR signaling.
If AGEF-1, the ARF GTPases and the AP-1 complex function together, we expect that they will have shared phenotypes. We tested whether the ARFs and the AP-1 complex regulate the size of vesicles in coelomocytes as does AGEF-1. While unc-101(sy108) mutants do not have large vesicles, further depletion of the AP-1 complex by RNAi of apm-1 AP-1μ or apg-1 AP-1γ in the unc-101(sy108) background resulted in enlarged LMP-1::GFP vesicles in the coelomocytes (Figure 3G–I). Consistent with previous studies, we found no evidence for arf-1.2 or arf-3 in regulating the size of vesicles in the coelomocytes [27], nor do deletions in arf-1.1 or arf-6. The complement of ARF GTPases that function with AGEF-1 and AP-1 in coelomocytes remains to be determined.
The AP-1 complex has recently been shown to restrict both apical and basolateral membrane protein localization in the C. elegans intestine [39], [40]. Similarly, we found that the apically localized SID-2 transmembrane protein [41], was mislocalized to the cytoplasm and basolateral membranes in agef-1(vh4) mutants (Figure S3A–D), suggesting that AGEF-1 and AP-1 might function together to regulate polarized localization of membrane proteins in the intestine.
The role of AGEF-1 in restricting SID-2::GFP on the apical membrane suggests that AGEF-1, the ARF GTPases and the AP-1 complex might restrict LET-23 EGFR to the apical membrane in the VPCs. To test this hypothesis we made use of two transgenic strains expressing a LET-23 EGFR GFP fusion (zhIs035 and zhIs038) that mimic the localization of endogenous LET-23 EGFR as seen by antibody staining [5]. In wild-type animals, LET-23::GFP localizes to both the apical and basolateral membranes of P6.p and in the lin-2(e1309) animals LET-23::GFP localizes strictly to the apical membrane (Figure 5A, C and Figure S4A, C). At the Pn.px stage, some basolateral, or lateral only localization is seen in lin-2(e1309) animals. Despite the lack of basolateral localization at the Pn.p stage, we find that the LET-23::GFP transgenes fully rescue the lin-2(e1309) Vul phenotype (Table 1, lines 8 and 9), suggesting that the levels of LET-23 EGFR at the basolateral membrane required for VPC induction are below the level of detection. Similarly, the gaIs27 LET-23::GFP transgene, that is only detectable by immunostaining with anti-GFP antibody, suppressed the lin-2(e1309) egg-laying defective phenotype [11].
To determine if AGEF-1 regulates LET-23 EGFR localization we compared the ratio of basolateral versus apical localization of LET-23::GFP in the P6.p cell of wild-type and agef-1(vh4) animals (Table 3). In wild-type, the average basal/apical intensity of LET-23::GFP in P6.p was 0.49 for zhIs035 and 0.65 for zhIs038. In agef-1(vh4) animals, the average basal/apical intensity of LET-23::GFP in the P6.p cell is 0.79 for zhIs035 and 0.93 for zhIs038 reflecting a decrease in apical intensity and an increase in basolateral intensity. We also found LET-23::GFP is present on the basolateral membrane of the intestinal cells in agef-1(vh4) animals whereas we did not see this in wild-type by confocal microscopy (Figure S3E–H). Therefore, AGEF-1 represses basolateral localization of LET-23::GFP in the VPCs and intestinal cells.
We next tested if agef-1(vh4) could restore the basolateral localization of LET-23::GFP in lin-2(1309) animals. We found that ∼40% of agef-1(vh4); lin-2(e1309) animals with zhIs035 have weak basolateral membrane localization of LET-23::GFP in P6.p compared to 9% in lin-2(e1309) animals (Figure 5C–D, G). Similarly, at the P6.px stage, we see an increase in the number of animals with basolateral localization of LET-23::GFP in agef-1(vh4); lin-2(e1309) as compared to lin-2(e1309) single mutants (Figure 5C′–D′, G′). No basolateral LET-23::GFP was seen with agef-1(vh4); lin-2(e1309) animals with the lower expressing zhIs038 (Figure S4C′–D′, F′). Since agef-1(vh4) is a weak hypomorphic mutation, we tested if knocking down the AP-1 complex via unc-101(RNAi) can further restore basolateral localization of LET-23 EGFR in agef-1(vh4); lin-2(e1309) mutants. We found that unc-101(RNAi) agef-1(vh4); lin-2(e1309) animals with either zhIs035 or zhIs038 had an increase in basolateral membrane localization of LET-23::GFP in both the intensity and the number of animals (Figure 5F–G′ and Figure S4E, F). unc-101(RNAi); lin-2(e1309) animals only showed mild restoration of LET-23::GFP using the zhIs035 transgene (Figure 5E, E′, G and G′). The restoration of LET-23 EGFR on the basolateral membrane in agef-1(vh4); lin-2(e1309), unc-101(RNAi); lin-2(e1309) and unc-101(RNAi) agef-1(vh4); lin-2(e1309) animals suggests that AGEF-1 and UNC-101 AP-1μ negatively regulate LET-23 EGFR signaling by limiting basolateral membrane localization.
Regulators of LET-23 EGFR trafficking are likely required for viability, as is the case for the RAB-7 GTPase [11]. In a screen for essential negative regulators of LET-23 EGFR-mediated vulva induction we identified a hypomorphic allele in the agef-1 gene. AGEF-1 is the C. elegans homolog of the yeast Sec7p and human BIG1 and BIG2 Arf GEFs, which function with class I Arf GTPases and the AP-1 complex to regulate cargo sorting and trafficking from the TGN [42]. We demonstrate that AGEF-1 regulates protein secretion, polarized protein localization, and late endosome/lysosome morphology. We show that AGEF-1 antagonizes signaling in the VPCs, upstream or in parallel to LET-23 EGFR, and that the class I/II Arf GTPases, ARF-1.2 and ARF-3, also negatively regulate signaling. Our genetic and phenotypic data are consistent with AGEF-1, the ARF-1.2 and ARF-3 GTPases, and the AP-1 complex together preventing ectopic vulva induction. The AGEF-1/Arf GTPase/AP-1 ensemble antagonizes the basolateral membrane localization of LET-23 EGFR in the VPCs; and hence, LET-23 EGFR-mediated vulva induction.
The clonal screen for suppressors of the lin-2(e1309) Vul phenotype was initially aimed at identifying maternal effect lethal mutants, like rab-7(ok511). Instead, we identified two strong suppressors of lin-2(e1309) with partial embryonic lethal phenotypes, agef-1(vh4) and vh22 (J. Meng, O.S. and C.E.R., unpublished data); which we currently believe function independently of each other and rab-7. The agef-1 deletion alleles are zygotic lethal and RNAi in the VPCs with agef-1, arf-1.2 and rab-7 has proven less effective than their corresponding genetic mutations [11; this study]. Therefore, the agef-1(vh4) mutation, being a recessive partial loss-of-function allele, provides a unique tool to study the function of agef-1, particularly in tissues refractory to RNAi such as the VPCs and neurons.
The agef-1(vh4) lesion changes a conserved negatively charged Glutamic Acid in the HDS2 domain to a positively charged Lysine. Collectively, the HDS2, HDS3, and HDS4 domains of yeast Sec7p have been shown to have an autoinhibitory function [43]. However, the specific function of the HDS2 domain is not known. Given the recessive nature of the agef-1(vh4) allele, it suggests that the HDS2 domain has a positive role in promoting AGEF-1 function.
Consistent with yeast Sec7p and human BIG1/BIG2 functioning in the secretory pathway, we found that agef-1(vh4) animals had defects in secretion of ssGFP from body wall muscles and Yolk::GFP from the intestine. Similar yolk secretion defects were recently reported for agef-1(RNAi) [31]. We found that agef-1(vh4) coelomocytes accumulated enlarged LMP-1::GFP positive late endosomes/lysosomes. Independently, Tang et al. [27] found that agef-1(RNAi) also caused enlargement of LMP-1::GFP vesicles and proposed a role for AGEF-1 in late endosome to lysosome trafficking, however, they did not find a defect in lysosome acidification or protein degradation. We do not know the reason for the enlarged late endosomes/lysosomes, but it could reflect a defect in retrograde transport from late endosomes to the Golgi as has been shown for knockdown of BIG1 and BIG2 or the AP-1 complex in mammalian cells [23]. Consistent with this idea, we found that knockdown of the AP-1 complex also induced enlarged LMP-1::GFP vesicles. Tang et al. [27] also reported that ssGFP accumulated in the pseudocoelom suggesting an uptake defect in the coelomocytes. We found that agef-1(vh4) mutants accumulated ssGFP in the body wall muscles rather than the pseudocoelom; thus the reduced ssGFP in coelomocytes could be explained by reduced secretion from the body wall muscles. However, we cannot rule out an uptake defect in the coelomocytes as well. Perhaps these discrepancies reflect a difference in reducing the levels of agef-1 by RNAi versus the vh4 missense mutation.
Our genetic analysis with agef-1(vh4) indicate that AGEF-1 is a potent negative regulator of LET-23 EGFR-mediated vulva induction. Similar to other negative regulators, agef-1 enhanced the Muv phenotype of the gain-of-function Ras mutant, let-60(n1046), and was a potent suppressor the Vul phenotypes of lin-2(e1309) and let-23(sy1) mutations, restoring vulva induction and even inducing a Muv phenotype. However, agef-1(vh4) failed to suppress a strong let-23(sy97) allele similar to sli-1 and unc-101 mutations and consistent with a role for AGEF-1 upstream or in parallel to LET-23 EGFR. In accordance with AGEF-1 being an Arf GEF, we found that ARF-1.2 and ARF-3, Class I/II Arf GTPases, also negatively regulate LET-23 EGFR signaling. The arf-1.2(ok796) deletion allele was a less potent suppressor of the lin-2(e1309) Vul phenotype as compared to the agef-1(vh4) mutant. However, arf-3(RNAi) in arf-1.2(ok796); lin-2(e1309) doubles showed suppression comparable to that in agef-1(vh4); lin-2(e1309) mutants. Therefore, ARF-1.2 and ARF-3 appear to function in a partly redundant manner during vulva development. Furthermore, expression of an ARF-1.2::GFP fusion in the VPCs rescued the suppressed Vul phenotype of both arf-1.2(ok796); lin-2(e1309) and agef-1(vh4); lin-2(e1309) animals indicating that ARF-1.2 antagonizes signaling in the VPCs likely downstream of AGEF-1.
In mammalian cells, the BIG1/BIG2 proteins and Arf1 recruit the AP-1 adaptor protein complex to the TGN and endosomes [15], [19], [20]. Both of the C. elegans AP-1μ subunits, unc-101 and apm-1, negatively regulate LET-23 EGFR mediated vulva development [12], [13]. In fact, apm-1(RNAi) unc-101(sy108) animals had a Muv phenotype, indicating that UNC-101 and APM-1 are functionally redundant during vulva induction, thus revealing a role for the AP-1 complex in inhibiting ectopic vulva induction [13]. Our findings that various double-mutant combinations between agef-1(vh4), arf-1.2(ok796) and unc-101(sy108) AP-1μ result in a synergistic Muv phenotype are consistent with AGEF-1, the Arfs and AP-1 functioning together to inhibit ectopic vulva induction. However, we cannot conclude whether they function in parallel pathways or in a common pathway due to the fact that agef-1(vh4) is not a null allele and the unc-101(sy108) and arf-1.2(ok796) mutations, while severe loss-of-function or null alleles, function in a partly redundant manner with apm-1 and arf-3, respectively. We favor a model whereby AGEF-1, the Arfs, and AP-1 function in a common pathway since this is most consistent with data from yeast and mammals, and that loss of AGEF-1 and components of the AP-1 complex have similar phenotypes in coelomocytes and the intestine. While synergistic genetic interactions are typically more indicative of genes in parallel pathways, we interpret that no single mutation in the AGEF-1/Arf/AP-1 pathway is sufficient to increase LET-23 EGFR signaling above a threshold necessary for ectopic induction. It is only when the activity of the AGEF-1/Arf/AP-1 pathway is further compromised by two mutations that LET-23 EGFR signaling increases above a threshold to induce a synergistic Muv phenotype. It is important to note that the AGEF-1/Arf/AP-1 pathway is essential, and only animals that survive to the fourth larval stage can be scored for vulva induction phenotypes. Thus, LET-23 EGFR signaling and localization phenotypes would likely be more severe if we were able to assess true null mutations in the VPCs only.
In polarized epithelial cells, the AP-1 complex mediates sorting and polarized distribution of transmembrane proteins, including EGFR, and thus the AGEF-1/Arf GTPase/AP-1 ensemble could regulate signaling via LET-23 EGFR localization. In the P6.p cell, we showed that the localization of LET-23 EGFR is altered in agef-1(vh4) animals using two transgenic lines (zhIs035 and zhIs038) expressing LET-23::GFP [5]. In wild-type animals, LET-23::GFP is present on both the apical and basolateral domains, however the average levels of LET-23:GFP on the apical membrane are double (zhIs035) or close to double (zhIs038) that on the basolateral membrane (Figure 6A). In agef-1(vh4) animals there was a redistribution of LET-23::GFP from apical to basolateral membrane bringing the average intensities closer to equal, suggesting that AGEF-1 either promotes apical localization or antagonizes basolateral localization of the receptor (Figure 6A, B). In the lin-2(e1309) background, LET-23::GFP is apical only (Figure 6C). In the more highly expressed line, zhIs035, we see some lateral only or faint basolateral in the P6.p descendants, P6.pa and P6.pp of lin-2(e1309) larvae. In the zhIs035 line, agef-1(vh4) partially restores LET-23::GFP on the basolateral membrane in lin-2(e1309) larvae. RNAi of unc-101 also partially restores basolateral localization and enhances the effect of agef-1(vh4) such that we see increased levels of LET-23::GFP, with both lines, in lin-2(e1309) larvae (Figure 6D). Therefore, AGEF-1 and UNC-101 AP-1μ cooperate to antagonize LET-23 EGFR basolateral localization and thus provide a mechanism by which these genes/proteins antagonize LET-23 EGFR signaling. Despite the lack of basolateral localization of LET-23::GFP in lin-2 mutant animals, the two LET-23::GFP transgenes used in this study rescued the lin-2(e1309) Vul phenotype, suggesting that the levels of receptor required for VPC induction are below detection. Therefore, the modest amount of LET-23::GFP restored to the basolateral membrane in agef-1(vh4); lin-2(e1309) or unc-101(RNAi); lin-2(e1309) could be more than sufficient to explain the strong restoration of VPC induction in these double mutants.
Our findings that an AGEF-1/Arf GTPase/AP-1 ensemble antagonizes the basolateral localization of LET-23 EGFR is contradictory to the established role of the mammalian AP-1A and AP-1B complexes in sorting transmembrane proteins to the basolateral membrane through the specific binding of basolateral sorting motifs in the cytoplasmic tail [18]. In fact, the AP-1B complex promotes the basolateral localization of EGFR in MDCK cells [17]. LET-23 EGFR does have several putative AP-1 sorting motifs, and thus could be a direct target for AP-1 regulation, but this would imply that AP-1 is impeding basolateral localization. A precedent for AP-1 having an antagonistic role in protein sorting or secretion has been found with the yeast Chs3p and Fus1p proteins, which rely on the exomer for secretion [44], [45]. In the absence of exomer, Chs3p and Fus1p are retained internally in an AP-1 dependent manner [45], [46], [47]. An analogous situation whereby the LIN-2/7/10 complex sorts/maintains LET-23 EGFR localization on the basolateral membrane and the AGEF-1/Arf/AP-1 pathway plays an antagonistic role could exist.
Recent studies in C. elegans and mice have shown that both basolateral and apical membrane cargos are mislocalized in the absence of the AP-1 complex [39], [40], [48], suggesting that the AP-1 complex is required to maintain the polarity of the epithelial cells [reviewed in 18]. Similarly, we find that agef-1(vh4) mutants mislocalized the SID-2 protein to the basolateral membranes, which is strictly apical in wild-type animals. Therefore, AGEF-1 might function with AP-1 to maintain polarity in the intestinal epithelia and by extension the AGEF-1/Arf GTPase/AP-1 ensemble could indirectly regulate LET-23 EGFR localization via maintenance of VPC polarity.
In summary, an AGEF-1/Arf GTPase/AP-1 ensemble functions opposite the LIN-2/7/10 complex to regulate apical versus basolateral localization of LET-23 EGFR in the VPCs, thus explaining how it negatively regulates LET-23 EGFR-mediated vulva induction. We don't yet know whether the AGEF-1/Arf GTPase/AP-1 ensemble directly regulates LET-23 EGFR sorting and localization or whether it is indirect via maintenance of VPC polarity. Further studies will be required to sort out the mechanisms by which the AGEF-1/ARF GTPase/AP-1 ensemble regulates LET-23 EGFR localization.
General methods for the handling and culturing of C. elegans were as previously described [49]. C. elegans Bristol strain N2 is the wild-type parent for all the strains used in this study; E. coli stain HB101 was used as a food source. The Hawaiian strain CB4856 was used for SNP mapping. All experiments were performed at 20°C. Information on the genes and alleles used in this work can be found on WormBase (www.wormbase.org) and are available through Caenorhabditis Genetics Center (www.cbs.umn.edu/cgc) unless otherwise noted in the strain list (Table S1).
lin-2(e1309) L4 hermaphrodites were mutagenized as previously described [49]. F1 progeny (m/+; lin-2) were transferred to individual plates. Due to the strong Vul phenotype of lin-2(e1309) animals, the self-progeny hatch internally [6]. F2 progeny were screened at the adult stage in order to identify plates that had a large number of eggs and egg-layers, additional preference was given to plates that had Muv, embryonic lethality, or dumpy phenotypes, similar to the rab-7(ok511) mutant. Progeny of a total of 2430 F1 animals (4860 haploid genomes) were screened and two lin-2 (e1309) suppressor mutants that are dumpy and partly embryonic lethal were identified, vh4 and vh22.
Single nucleotide polymorphism (SNP) mapping was used to place vh4 to the right arm of chromosome I [25]. Chromosome mapping showed linkage of vh4 to SNPs at 13 (F58D5), 14 (T06G6) and 26 (Y105E8B) map units (m.u.). Interval mapping using two sets of recombinants, 141 animals in total, was conducted using the following SNPs: pkP1133 at 17.4 m.u (A/T Bristol/CB4856, RFLP DraI); pkP1134 at 18.95 m.u. (T/C Bristol/CB4856, RFLP AflIII); haw14061 at 19.51 m.u. (T/C Bristol/CB4856, sequencing); haw14137 at 20.65 m.u. (T/A Bristol/CB4856, sequencing); haw14164 at 21.04 m.u. (C/T Bristol/CB4856, sequencing); CE1-248 at 21.97 m.u. (T/A Bristol/CB4856, sequencing); CE1-220 at 23.16 m.u. (A/G Bristol/CB4856, sequencing); pkP1071 at 23.4 m.u. (C/T Bristol/CB4856, RFLP EcoRI). In the course of interval mapping the following predicted sequencing SNPs were confirmed: haw14061 and haw14137 as T/C and T/A Bristol/CB4856, respectively. Genomic DNA from vh4 and vh22 was isolated and submitted to Genome Quebec for Illumina sequencing. Within the defined map region, the agef-1 gene was the only gene carrying a non-synonymous mutation unique to the vh4 strain.
RNAi feeding was performed essentially as previously described [50] using the unc-101 (I-6G20), agef-1 (I-6L22), arf-1.2 (III-3A13), and arf-3 (IV-4E13) clones from Ahringer RNAi library (Geneservice, Cambridge, United Kingdom). Clones were verified by DNA sequencing. To avoid embryonic and larval lethal phenotypes, synchronized L1 larvae were placed on RNAi plates and scored for vulva induction when the animals reached L4 stage 36–48 hours later.
General methods for live animal imaging using Nomarski differential interference contrast (DIC) microscopy were as previously described [51]. Animals were analyzed on an Axio Zeiss A1 Imager compound microscope (Zeiss, Oberkochen, Germany) and images were captured using an Axio Cam MRm camera and AxioVision software (Zeiss, Oberkochen, Germany). Muv and Vul phenotypes were scored by counting the numbers of vulval and non-vulval descendants of P3.p–P8.p in L4 stage larvae as described previously [11]. Fisher's exact test (www.graphpad.com/quickcalcs) was used for statistical analysis of the vulval phenotypes. Comparison of GFP intensities wild-type and agef-1(vh4) was performed using identical exposure times for conditions being compared. Fiji image processing tool was used to measure intensities in raw images; any adjustments to contrast/brightness were for presentation purposes and were performed after analysis [52]. Tissue/organ of interest was outlined using free hand selection tool followed by measurement of the average pixel intensity. Images selected for figures are representative of the mean value for average pixel intensity for the group. Statistical analysis and graphing was done using Prism 5 (GraphPad Software, Inc., La Jolla, CA).
Confocal analysis was performed using a Zeiss LSM-510 Meta laser scanning microscope with 63× oil immersion lens (Zeiss, Oberkochen, Germany) in a single-track mode using a 488 nm excitation for GFP. Images were captured using ZEN 2009 Image software (Zeiss, Oberkochen, Germany). Animals at the L4 larval stage were selected for visualization of endocytic/secretory compartments in the coelomocytes. Images selected for figures are representative of the mean value for the largest vesicle diameter for the group. Statistical analysis and graphing was done using Prism 5 (GraphPad Software, Inc., La Jolla, CA). Confocal analysis of zhIs038 transgene-carrying animals was performed at early L3 larval stage using the Zeiss LSM-510 Meta laser scanning microscope. Confocal analysis of zhIs035 was performed using the Zeiss Axio Observer Z1 LSM-780 laser scanning microscope with 63× oil immersion lens (Zeiss, Oberkochen, Germany) in a single-track mode using an Argon multiline laser with 488 nm excitation for GFP. Images were captured using ZEN 2010 Image software (Zeiss, Oberkochen, Germany). The apical and basal LET-23::GFP intensities were measured using Fiji by drawing a line through the center of the nucleus in the DIC channel and transferring the selection into the GFP channel to prevent bias.
arf-1.2 was amplified by PCR from wild-type cDNA using the primers 5′-CATAAGAATAGTCGACATGGGAAACGTGTTCGGCAGC-3′ (forward) and 5′-GATTCTGATTACCGGTTCAGATCTATTCTTGAGCT-3′ (reverse) containing SalI and AgeI cut sites, respectively. The PCR product was cloned into pEGFP-N1 plasmid using SalI 639 and AgeI 666 sites. arf-1.2::GFP was digested using SalI and NotI and subcloned into the p255 lin-31 promoter plasmid. Transgenic animals were generated by DNA microinjection [53] of the Plin-31::ARF-1.2::GFP plasmid and a marker plasmid Pttx-3::GFP at a concentration of 50 ng/µl of each into N2 animals using maxiprep quality DNA. Two of three lines were used for this study, vhEx7 and vhEx8. Rescue of agef-1(vh4) and arf-1.2(ok796) mediated suppression of the lin-2(e1309) Vul phenotype was scored in animals expressing ARF-1.2::GFP in the VPCs.
Analysis of the Arf GTPases was performed using MAFFT version 7 multiple alignment program for amino acid or nucleotide sequences online (http://mafft.cbrc.jp) [54]. Input sequences were human NP_001649.1 (Arf1), NP_001650.1 (Arf3), NP_001651.1 (Arf4), NP_001653.1 (Arf5), AAV38671.1 (Arf6), NP_001168.1 (Arl1) and C. elegans NP_501242.1 (ARF-1.1), NP_498235.1 (ARF-1.2), NP_501336.1 (ARF-3), NP_503011.1 (ARF-6), NP_495816.1 (ARL-1). Phylogenetic tree was constructed and visualized using Archaeopteryx [55], [56].
|
10.1371/journal.ppat.1004389 | Glutamate Secretion and Metabotropic Glutamate Receptor 1 Expression during Kaposi's Sarcoma-Associated Herpesvirus Infection Promotes Cell Proliferation | Kaposi's sarcoma associated herpesvirus (KSHV) is etiologically associated with endothelial Kaposi's sarcoma (KS) and B-cell proliferative primary effusion lymphoma (PEL), common malignancies seen in immunocompromised HIV-1 infected patients. The progression of these cancers occurs by the proliferation of cells latently infected with KSHV, which is highly dependent on autocrine and paracrine factors secreted from the infected cells. Glutamate and glutamate receptors have emerged as key regulators of intracellular signaling pathways and cell proliferation. However, whether they play any role in the pathological changes associated with virus induced oncogenesis is not known. Here, we report the first systematic study of the role of glutamate and its metabotropic glutamate receptor 1 (mGluR1) in KSHV infected cell proliferation. Our studies show increased glutamate secretion and glutaminase expression during de novo KSHV infection of endothelial cells as well as in KSHV latently infected endothelial and B-cells. Increased mGluR1 expression was detected in KSHV infected KS and PEL tissue sections. Increased c-Myc and glutaminase expression in the infected cells was mediated by KSHV latency associated nuclear antigen 1 (LANA-1). In addition, mGluR1 expression regulating host RE-1 silencing transcription factor/neuron restrictive silencer factor (REST/NRSF) was retained in the cytoplasm of infected cells. KSHV latent protein Kaposin A was also involved in the over expression of mGluR1 by interacting with REST in the cytoplasm of infected cells and by regulating the phosphorylation of REST and interaction with β-TRCP for ubiquitination. Colocalization of Kaposin A with REST was also observed in KS and PEL tissue samples. KSHV infected cell proliferation was significantly inhibited by glutamate release inhibitor and mGluR1 antagonists. These studies demonstrated that elevated glutamate secretion and mGluR1 expression play a role in KSHV induced cell proliferation and suggest that targeting glutamate and mGluR1 is an attractive therapeutic strategy to effectively control the KSHV associated malignancies.
| Kaposi's sarcoma associated herpesvirus (KSHV), prevalent in immunosuppressed HIV infected individuals and transplant recipients, is etiologically associated with cancers such as endothelial Kaposi's sarcoma (KS) and B-cell primary effusion lymphoma (PEL). Both KS and PEL develop from the unlimited proliferation of KSHV infected cells. Increased secretion of various host cytokines and growth factors, and the activation of their corresponding receptors, are shown to be contributing to the proliferation of KSHV latently infected cells. Glutamate, a neurotransmitter, is also involved in several cellular events including cell proliferation. In the present study, we report that KSHV-infected latent cells induce the secretion of glutamate and activation of metabotropic glutamate receptor 1 (mGluR1), and KSHV latency associated LANA-1 and Kaposin A proteins are involved in glutaminase and mGluR1 expression. Our functional analysis showed that elevated secretion of glutamate and mGluR1 activation is linked to increased proliferation of KSHV infected cells and glutamate release inhibitor and glutamate receptor antagonists blocked the proliferation of KSHV infected cells. These studies show that proliferation of cancer cells latently infected with KSHV in part depends upon glutamate and glutamate receptor and therefore could potentially be used as therapeutic targets for the control and elimination of KSHV associated cancers.
| Kaposi's sarcoma-associated herpesvirus or human herpesvirus-8 (KSHV/HHV-8) infection is etiologically associated with Kaposi's sarcoma (KS), a vascular endothelial tumor, and two B-cell lymphoproliferative diseases, primary effusion lymphoma (PEL) or body-cavity based lymphoma (BCBL) and multicentric Castleman's disease [1], [2], [3]. These cancers occur more frequently in the setting of immunosuppression, including HIV-1 infected patients, and develop from cells latently infected with KSHV. In vivo KSHV has a broad tropism and viral genome and transcripts are detected in a variety of cells such as B cells, endothelial cells, monocytes, keratinocytes, and epithelial cells [4], [5]. Latent KSHV DNA is present in vascular endothelial and spindle cells of KS lesions, associated with expression of latency associated ORF73 (LANA-1), ORF72 (v-cyclin D), K13 (v-FLIP), and K12 (Kaposin) genes and microRNAs [5]. Cell lines with B cell characteristics, such as BC-1, BC-3, BCBL-1, HBL-6 and JSC have been established from PEL tumors [4], [5]. In PEL cells, in addition to the above set of latent genes, the K10.5 (LANA-2) gene is also expressed [4], [5]. About 1–3% of PEL cells spontaneously enter the lytic cycle and virus induced from these cells by chemicals serve as the source of virus. Multiple genome copies of both KSHV and EBV exist in latent form in BC-1, HBL-6 and JSC cells while BCBL-1 and BC-3 cells carry only the KSHV genome [4], [5]. KSHV infects a wide variety of human cell types in vitro, including fibroblasts, keratinocytes, B cells, endothelial, and epithelial cells [4], [5], [6]. Following infection, KSHV establishes latency within the target cells and the expression of the viral latent ORF71, ORF72, ORF73, and K13 genes continues to maintain latency [7]. In addition, host genes required for regulating apoptosis, signal induction, cell cycle regulation, inflammatory response, and angiogenesis are also highly upregulated in the latently infected cells [8].
Studies have linked the expression of KSHV latency genes ORF71 (v-FLIP), -72 (v-Cyclin), -73 (LANA-1) and K12 (Kaposin A) to the oncogenic activity of latently infected cells [9], [10], [11]. These genes induce the oncogenic potential of KSHV by increasing proliferative potential, growth and chromosome instability as well as by preventing apoptosis of the infected cells [9], [10], [11]. A hallmark of KSHV associated cancers is the excessive secretion of cytokines and growth factors [12], [13], [14]. Modulation by viral proteins and virally induced cellular proteins promote the secretion of autocrine and paracrine cytokines and growth factors leading into the proliferation, survival, and growth of the latently infected cells [12], [13], [15], [16], [17], [18]. However, the mechanism behind KSHV induced cancer progression is not completely understood.
Glutamate is a major excitatory neurotransmitter in the mammalian brain. It also plays a central role in several cellular functions, including cell survival and death by its interaction with receptors [19]. Glutamate released into the extracellular space binds and activates two classes of cell surface receptors, ionotropic (iGluRs) and G protein-coupled metabotropic glutamate receptors (mGluRs). There are three groups of mGluRs, and group I mGluRs have been extensively studied in relation to cell survival and death. Group I consists of mGluR1 and mGluR5 subtypes (mGluR1/5) which are coupled to Gαq/11 proteins. Agonist stimulation of group I mGluRs activate PLC, which results in the activation of PKC, PKC dependent pathways, and ERK1/2 [20], [21], [22]. The group I mGluRs are distributed in a variety of non-neuronal cells including human B and microvascular dermal endothelial cells, the natural target cells of KSHV [23], [24], [25]. However, the involvement of excess glutamate secretion and glutamate receptor expression in cell proliferation is an unexplored area of research in the KSHV oncogenesis field. This current study was undertaken with a rationale that identifying and defining the role of glutamate in KSHV biology will lead into targeted specific treatments for KSHV-associated malignancies.
Our studies demonstrate that KSHV infection induces the secretion of glutamate and expression of mGluR1 receptor, and increased mGluR1 expression was detected in KS and PEL tissue sections. Most notably, glutamate secretion and mGluR1 activation in KSHV latently infected cells occurred through two independent pathways regulated by two individual viral latent proteins, LANA-1 and Kaposin A. Our data highlight how KSHV LANA-1 and Kaposin A proteins contribute to the generation of glutamate, activation of mGluR1, and strongly suggest the possibility of exploiting the glutamatergic system for the therapeutic intervention of KSHV dependent cancers.
To determine the role of glutamate in KSHV infection, we first evaluated the secretion of glutamate during de novo KSHV infection of primary human microvascular dermal endothelial cells (HMVEC-d). Kinetics of glutamate secretion showed that KSHV infection robustly increased glutamate release as early as 8 h post-infection (p.i.) which continued to increase throughout the 5 d p.i. observation period (Figure 1A). In contrast, when the cells were infected with replication defective UV treated KSHV for 5d, there was no significant difference in glutamate secretion between uninfected and UV-KSHV infected cells, (Fig. 1A) suggesting that viral gene expression is required for the increased secretion of glutamate. To determine whether the secretion of glutamate is specifically induced by KSHV, cells were infected with KSHV pre-incubated with heparin (Hep-KSHV), which is known to block the binding and entry of KSHV to the target cells [26]. In contrast to the untreated virus, heparin treated virus (Hep-KSHV) considerably reduced the secretion of glutamate (Figure 1A). This suggested that KSHV entry and infection is required for the increased secretion of glutamate.
KS is an endothelial tumor, whereas PEL is of B-cell origin [2], [27], [28]. The telomerase immortalized endothelial cell line (TIVE) latently infected with KSHV (TIVE-LTC), and the PEL derived B-cell line BCBL-1 are well-established in vitro models to study KS and PEL, respectively [2], [29]. In addition, BJAB-KSHV, a Burkitt's lymphoma B-cell line carrying latent KSHV DNA, has also been used as an additional model for studying KSHV pathogenesis [30]. To test whether the process of glutamate generation is relevant to KS, we measured the secretion of glutamate in KSHV TIVE-LTC cells as well as in uninfected control TIVE cells. Similar to de novo KSHV infection, higher levels of glutamate release were observed in KSHV(+) TIVE-LTC cells than in KSHV(−) TIVE cells (Figure 1B). When the association of glutamate to PEL was assessed, high levels of glutamate release were observed in KSHV (+) BCBL-1 and BJAB-KSHV cells compared to the KSHV (−) B-cell line BJAB (Figure 1C).
To elucidate the mechanisms of glutamate generation in the infected cells, we next determined the expression of glutaminase, the major enzyme responsible for glutamate production [31]. Compared to the uninfected cells, a time dependent increase in glutaminase expression was observed during 8 h, 24 h, 48 h and 5 d of de novo infection of primary HMVEC-d cells by KSHV (Figure 1D, lanes 1–6). In contrast, at 5 d p.i. with UV-KSHV, no significant difference in glutaminase expression from uninfected cells was observed (Figure 1D, lanes 1 and 7). These results demonstrated that the increased glutamate secretion is linked with increased glutaminase expression in the infected cells. This link was further confirmed by the detection of a higher level of glutaminase expression in the latently infected TIVE-LTC (2.8 fold), BJAB-KSHV (2.5 fold), and BCBL-1 cells (3.4 fold) than in their respective uninfected control TIVE and BJAB cells (Figure 1E, lanes 1–5).
To further investigate the role of glutaminase in glutamate secretion, we used a glutaminase specific inhibitor, L-DON (6-diazo-5-oxo-norleucine) [32]. Cells were treated with L-DON at a concentration of 500 µM and 1 mM and the supernatants were analyzed for glutamate release. We found that 500 µM of L-DON inhibited glutamate secretion by >50% and 1 mM of L-DON by >65%. Dose dependent inhibition of glutamate secretion in L-DON treated cells strongly suggested that glutaminase is the major enzyme that contributes to the generation of excess glutamate in KSHV infected cells (Figure S1A). L-DON had no significant cytotoxicity on BJAB cells at 500 µM and 1 mM concentrations (data not shown).
KSHV latency-associated ORF73 gene product LANA-1 has been shown to induce c-Myc expression [33]. Since c-Myc has also been shown to activate the expression of glutaminase [34], we hypothesized that the increased glutamate secretion observed in KSHV infected cells could be mediated by LANA-1 through its c-Myc activation, which in turn stimulates the expression of glutaminase. To test this hypothesis, when BJAB cells were transduced with lentivirus constructs of LANA-1, we observed increased secretion of glutamate in LANA-1 transduced cells compared to vector alone (Figure 2A). We also observed ∼2-fold increase in c-Myc and glutaminase protein expression in LANA-1 transduced cells (Figure 2B).
To support our finding that LANA-1 mediated c-Myc activation is directly involved in glutaminase expression and glutamate secretion, we used lentiviruses encoding shRNAs to knock down c-Myc in BJAB cells over expressing LANA-1. As shown in figure S1B, LANA-1 over expression induced the secretion of glutamate, and this induction was abolished by the knockdown of c-Myc (Figure S1B). Since no considerable increase in glutamate release was observed in the absence of c-Myc in LANA-1 expressing cells (Figure S1B), these results suggested that LANA-1 mediated c-Myc activation is required for glutamate release.
To confirm the functional relationship of c-Myc expression with glutaminase expression in KSHV infected cells, we transduced TIVE-LTC cells and BCBL-1 cells with c-Myc and control shRNA lentiviral vectors. A significant reduction in glutaminase expression was observed in c-Myc knockdown TIVE-LTC cells (61%) and BCBL-1 cells (67%) compared to control shRNA transduced cells (Figures 2C and D). These results suggested that LANA-1 mediated c-Myc activation plays a crucial role in the expression of glutaminase and glutamate secretion in cells latently infected with KSHV.
Among the several types of glutamate receptors, mGluR1 is considered an oncogenic protein due to its ability to regulate the functions related to cancer cell proliferation [35], [36]. Hence, we theorized that the biological effect of glutamate in latent KSHV induced oncogenesis may be mediated through the expression of mGluR1 receptors. To test this, we first determined mGluR1 expression by RT-PCR in primary endothelial cells infected for 5 d with KSHV and UV-KSHV. Compared to uninfected cells, KSHV infection increased the expression of mGluR1 (Figure 3A). In contrast, UV treated virus had no significant effect on mGluR1 expression (Figure 3A) and suggested that sustained mGluR1 receptor expression probably depended upon KSHV gene expression. When the relative expression levels for the mGluR1 receptor in KSHV latent TIVE-LTC, BJAB-KSHV and BCBL-1 cells as well as control BJAB and TIVE cells were determined by RT-PCR, upregulation of mGluR1 in both KSHV (+) TIVE-LTC cells and BCBL-1 cells was observed compared to uninfected TIVE and BJAB cells (Figure 3B). Western blot (Figure 3C) and immunoprecipitation analysis (Figure S2A) confirmed the higher levels of mGluR1 protein in de novo KSHV infected primary cells compared to the uninfected and UV-KSHV infected cells. Similarly, high levels of mGluR1 expression were also observed in BJAB-KSHV, BCBL-1 and TIVE-LTC cell lines by Western blots (Figure 3D) and by immunoprecipitation analysis (Figure S2B).
The expression of mGluR1 in KSHV infected primary cells and latent cells was also examined by immunofluorescence assay (IFA). Increased mGluR1 staining was detected in LANA-1 expressing spindle shaped HMVEC-d cells infected with KSHV (Figure 3E), as well as in TIVE-LTC, BJAB-KSHV and BCBL-1 cells compared to their respective uninfected controls (Figures 3F and G). These results clearly demonstrated that KSHV infection results in the increased mGluR1 expression in latently infected cells.
To verify the pathological association of mGluR1 in KSHV associated cancers, we immunostained normal as well as KSHV infected KS and PEL tissues by dual labeled IFA for mGluR1 and KSHV LANA-1 as a marker for infection. Strong positive immunostaining for both mGluR1 and LANA-1 were detected in the spindle shaped endothelial cells of KS tissue (Figure 4A), and in the stomach PEL (Figure 4B) samples. In contrast, only a basal level of mGluR1 was detected in control normal skin and stomach samples (Figures 4A and B). These results clearly demonstrated the in vivo association of increased mGluR1 expression with KSHV infection.
The expression of mGluR1 in non-neuronal cells is regulated by RE-1 silencing transcription factor/neuron restrictive silencer factor (REST/NRSF) [37]. Binding of REST to a DNA recognition sequence called the neuron restrictive silencer elements (NRSE or RE-1) repress the expression of neuronal genes such as mGluR1 in non-neuronal cells [37], [38], [39]. To analyze whether REST expression plays any role in mGluR1 expression in KSHV infected cells, we determined the expression of REST mRNA and protein. real-time RT-PCR analysis of REST revealed similar levels of REST expression in both uninfected and KSHV infected latent cells (Figures S3A and B). However, REST protein expression determination by Western blots showed 55%, 72% and 42% reduction in BJAB-KSHV, BCBL-1 and TIVE-LTC cells, respectively, compared to the respective controls (Figures S3C and D). This suggested that REST expression in KSHV infected cells is probably modulated at the post-transcriptional level.
To decipher the mechanism regulating REST expression at the post-translational level, we first determined the subcellular localization of REST in TIVE and TIVE-LTC cells by IFA. In the uninfected TIVE cells, REST was highly expressed and was predominantly localized in the nucleus (Figure 5A). In contrast, REST distribution was markedly decreased in the nucleus of TIVE-LTC cells and was predominantly localized to the cytoplasm (Figure 5A). A similar cytoplasmic relocalization of REST was observed in almost all KSHV infected BJAB-KSHV and BCBL-1 cells compared to the uninfected BJAB cells where it was exclusively localized in the nucleus (Figure 5B).
Western blot analysis of cytoplasmic and nuclear fractions of the KSHV positive cell lines confirmed that REST localization is significantly decreased in the nucleus (Figures 5C and D) with a concomitant increase in the cytoplasm of infected cells, whereas it was undetectable in the cytoplasm of uninfected TIVE and BJAB cells (Figures 5C and D, lane 1). Interestingly, analyses of REST in the cytoplasmic fractions from the infected cells showed a small shift in molecular weight in both TIVE-LTC (Figure 5C, lane 2) and BCBL-1 cells (Figure 5D, lane 3). We reasoned that this small shift in band size could be due to phosphorylation of REST, which is known to result in migration differences on SDS-PAGE. To determine whether the shifted band detected in the cytoplasm is indeed the phosphorylated form of REST, we first treated the cytoplasmic extracts from TIVE-LTC cells with lambda phosphatase or with lambda phosphatase and phosphatase inhibitor, and then the extracts were Western blotted. Treatment with lambda phosphatase resulted in the disappearance of the modified band, suggesting that the shift in band size was due to phosphorylation (Figure S4A).
It has been reported that serine phosphorylation of REST in the conserved phosphodegron motif promotes recognition by the E3 ubiquitin ligase β-TRCP and ubiquitination [40], [41]. As our data suggested an unexpected decrease of REST in the infected cells, we next asked whether the phosphorylation of REST in the cytoplasm was followed by its phosphorylation-dependent ubiquitination. To examine this regulatory role, we first verified the serine phosphorylation of REST in the cytoplasm of infected cells by immunoprecipitating with phosphoserine antibody and Western blotting with REST antibody. Consistent with the Western blot results (Figures 5C and D), a significant level of serine phosphorylation of REST was detected in KSHV-infected TIVE-LTC, BJAB-KSHV and BCBL-1 cells compared to a very low level of phosphorylation in KSHV-negative TIVE and BJAB cells (Figure 6A).
We next determined whether β-TRCP could be associated with phosphorylated REST in the cytoplasm of infected cells. Immunoprecipitation of cytoplasmic extracts of BJAB, BJAB-KSHV and BCBL-1 cells with REST and Western blots with anti-β-TRCP antibodies showed increased interaction of REST with β-TRCP in the infected cells, whereas it was barely detectable in uninfected cells (Figure 6B). We next determined whether REST degradation occurs in the cytoplasm of KSHV-infected cells. Analysis of cytosolic fractions from TIVE-LTC, BJAB-KSHV and BCBL-1 cells by immunoprecipitation with REST and Western blots for polyubiquitin revealed a higher level of ubiquitination in TIVE-LTC, BJAB-KSHV and BCBL-1 cells compared with TIVE and BJAB cells displaying lower levels of ubiquitination (Figure 6C). Thus, the ubiquitination levels of REST correlated with REST phosphorylation and the association of REST with β-TRCP in the cytoplasm.
In order to confirm that the ubiquitin proteasome system is involved in the degradation of REST in KSHV infected cells, BCBL-1 and TIVE-LTC cells were treated with the proteasome inhibitor MG132, and the cell lysates were Western blotted for REST. As shown in figure 6D, compared to the untreated cells (lanes 3, 5, and 7), MG132 treatment increased the protein level of REST in the infected BJAB-KSHV, BCBL-1, and TIVE-LTC cells (lanes 4, 6, and 8). However, MG132 treatment had no significant effect on the REST protein level in uninfected BJAB cells (lanes 1 and 2). This result further supported our finding that the degradation of REST observed in the infected cells was probably mediated by the ubiquitin proteasome pathway.
Since REST was more localized in the cytoplasm of latently infected cells, we hypothesized that latent KSHV protein(s) in the infected cells binds and sequesters REST in the cytoplasm, which in turn leads to overexpression of the mGluR1 gene. To determine the identity of the KSHV latent protein responsible for this, BJAB cells were transduced with the lentiviral constructs of KSHV latent ORF71, -72, -73, and Kaposin A genes, expression levels assessed by real-time PCR (Figure S4B), and mGluR1 level analyzed by Western blot. ORF K12 or Kaposin A transduction led to a robust increase in mGluR1 expression in BJAB cells, indicating the involvement of Kaposin A in the regulation of mGluR1 expression, whereas the other latent genes did not significantly induce the expression of mGluR1 (Figure 7A). mGluR1 expression in Kaposin A transduced BJAB cells was further confirmed by immunoprecipitation experiments (Figure S4C). Transduction efficiencies were determined by control lentiviral GFP expression (Figure S4D). We also observed higher levels of mGluR1 protein expression in primary HMVEC-d cells transduced with Kaposin A which further demonstrated the Kaposin A dependency of mGluR1 expression (Figure 7B).
To determine whether Kaposin A is responsible for the observed cytoplasmic relocalization of REST in the infected cells, we transduced HMVEC-d cells with a lentiviral Kaposin A construct (ORF K12) and localization was determined by IFA using anti-Kaposin A antibodies. This analysis revealed that a major portion of endogenous REST was translocated into the cytoplasm and colocalized with Kaposin A in the transduced cells (Figure 7C).
To verify that REST binds to Kaposin A in the cytoplasm of KSHV infected cells, we immunoprecipitated REST from cytoplasmic fractions of both uninfected BJAB and KSHV-infected BCBL-1 cells and then Western blotted with anti-Kaposin A antibodies, which detected specific bands of Kaposin A at approximately 16–18 kDa (Figure 8A). BCBL-1 cell lysates used as positive control also identified 16–18-kDa immunoreactive bands of Kaposin A in the infected cells (Figure 8A). The predicted molecular weight of Kaposin A is 6-kDa; however, WB analyses often detect specific bands of about 16–18-kDa and above [42], [43], [44]. Similar immunoprecipitation analysis using TIVE and TIVE-LTC cells also revealed that REST interacts with Kaposin A in the infected cell cytoplasm (Figure 8B). We also observed the colocalization of REST and Kaposin A in the cytoplasm of TIVE-LTC (Figure 8C) and BCBL-1 cells (Figure 8D).
To further verify the physical interaction of REST with Kaposin A, we co-transduced 293T cells with Kaposin A and retroviral FLAG tagged REST and the cytoplasmic and nuclear lysates were immunoprecipitated with Kaposin A and Western blotted with anti-FLAG antibodies. This co-immunoprecipitation experiment demonstrated the ability of Kaposin A to interact with REST in the cytoplasm, but not in the nucleus (Figure 8E).
We next examined the staining pattern and colocalization of REST and Kaposin A in KS and PEL patient samples and in normal tissues by immunofluorescence analysis. As shown in Figures 9A and B, strong nuclear staining of REST was observed in normal skin tissues as well as in normal stomach tissues. In contrast, cytoplasmic localization of REST and notable colocalization with Kaposin A were observed in the endothelial cells of KS as well as in the cells of PEL tissues, presumably the B cells. Together, these results suggested that Kaposin A expression regulates mGluR1 expression through interaction with REST in the cytoplasm of KSHV infected cells.
As shown in figure 6B and C, phosphorylated REST interacts with β-TRCP and promotes the ubiquitination and degradation of REST in the cytoplasm of infected cells. It has previously been reported that REST has a degron motif and the phosphorylation of REST at serine 1024, 1027, and 1030 of the degron motif is required for the interaction of REST with β-TRCP during oncogenic transformation [41]. Because Kaposin A is a protein involved in transformation of infected cells [45], [46], we postulated that Kaposin A binding with REST phosphorylates REST at the 1024, 1027, and 1030 residues, leading to the interaction with β-TRCP and ubiquitination of REST.
To investigate this, we co-transduced 293T cells with Kaposin A and FLAG REST-WT or FLAG-REST triple mutant (where all three phosphodegron residues are mutated-FLAG-REST-S1024/1027/1030A), cytoplasmic fractions immunoprecipitated with anti-phosphoserine antibodies and Western blotted with anti-FLAG antibodies. A significant level of serine phosphorylated REST was detected in Kaposin A and FLAG REST-WT transduced cells (Figure 10A, upper panel). In contrast, the serine phosphorylation of REST was severely impaired in Kaposin A and FLAG-REST triple mutant transduced cells suggesting that Kaposin A mediates REST phosphorylation in its phosphodegron sites. As Kaposin A is involved in REST phosphorylation in the conserved degron sites, we next determined whether phosphorylated REST binds to endogenous β-TRCP. As shown in Figure 10A, middle panel, immunoprecipitation with FLAG and Western blot with β-TRCP showed a markedly increased interaction of REST with endogenous β-TRCP in Kaposin A and FLAG REST-WT transduced cells, whereas no interaction was observed in Kaposin A and FLAG-REST triple mutant transduced cells. These data demonstrated that blocking Kaposin A mediated phosphorylation of REST weakens its association with endogenous β-TRCP.
To further investigate which specific degron site is phosphorylated by Kaposin A, we transiently transduced 293T cells with lentiviral Kaposin A and 48 h after transduction, the cells were transfected with pCMV-FLAG-REST WT plasmid or pCMV-FLAG-REST individually mutated at serine 1024, (pCMV-FLAG-REST-S1024A), 1027 (pCMV-FLAG-REST-S1027A), or 1030 (pCMV-FLAG-REST-S1030A). Cell lysates were immunoprecipitated using anti-phosphoserine antibodies followed by Western blotting with anti-FLAG antibodies or vice versa. As shown in Figure 10B, the serine 1027 mutant completely abolished the capacity for phosphorylation (Figure 10B, lane 5, first and second panel). However, the serine 1024 and 1030 mutants had no effect on phosphorylation compared to wild type REST (Figure 10B, lane 4 and 6, first and second panel), indicating that the phosphorylation of these two sites are not directly mediated by Kaposin A. The defect in REST phosphorylation in mutant serine 1027 suggested that Kaposin A initially phosphorylates REST on serine 1027. Phosphorylation on 1027 may provide the signal to phosphorylate the other degron residues.
In order to determine whether the serine 1027 induced phosphorylation is responsible for REST ubiquitination, the cell lysates immunoprecipitated with anti-FLAG antibody were analyzed by Western blotting with an anti-polyubiquitin antibody. Consistent with the increased phosphorylation, the ubiquitination was markedly increased in wild type REST, as well as in serine 1024 and 1030 mutants transfected cells (Figure 10B, lanes 3, 4, and 6, third panel). In contrast, the phosphorylation defective mutant 1027 failed to induce ubiquitination (Figure 10B, lane 5, third panel). These results suggest that serine 1027 mediated phosphorylation is required for the ubiquitination of REST. We also observed that the phosphorylation defective mutant FLAG-REST-S1027A stabilized REST (Figure 10B, lane 5, fourth panel). The observed reduction of REST in FLAG-REST WT and FLAG-REST-S1024A and -S1030A (Figure 10B, lanes 3, 4 and 6, fourth panel), after Kaposin A stimulation may be due to the degradation of phosphorylated REST at the 1027 residue. Taken together, our studies demonstrated that Kaposin A regulates REST phosphorylation in the conserved phosphodegron motif which enhances the ubiquitination of REST and thus reduces the level of REST.
To further verify the role of Kaposin A in REST phosphorylation, 293T cells transduced with vector alone or Kaposin A were transfected with FLAG-REST S1027 or FLAG-REST WT first and then transfected with control or Kaposin A specific siRNA. After 48 h post transfection, levels of REST phosphorylation were assessed by immunoprecipitating with anti-FLAG antibody followed by Western blotting with anti-phosphoserine antibody. We observed that compared to control siRNA transfected cells, Kaposin A siRNA transfected cells abolished the phosphorylation and degradation of REST in REST WT transfected cells (Figure 10C, lane 4 and 5, first and second panel). Kaposin A specific siRNA efficiently knocked down the expression of Kaposin A in the transduced cells (Figure 10C, lane 5, third panel). As expected, cells transfected with phosphorylation defective mutant REST-S1027 had no effect on phosphorylation (Figure 10C, lane 3, first panel). These data confirm that Kaposin A is essential for the phosphorylation of REST.
We next focused on the biological response of glutamate release and binding to its receptors. We postulated that the glutamate released by infected cells binds to mGluR1 permitting cellular signaling and the proliferation of glutamate secreting infected cells. To determine the effects of glutamate and mGluR1 on cell proliferation, primary HMVEC-d cells infected with KSHV for 3 d were cultured for 2 d in the presence or absence of glutamate release inhibitor riluzole, and mGluR1 antagonists A841720 and Bay 36-7620, pulsed with BrdU for 2 h and BrdU incorporation determined by IFA. As shown in Figure 11A, HMVEC-d cells infected with KSHV for 5 d showed a much higher rate of proliferation than the uninfected cells. This increased proliferation of HMVEC-d cells was significantly reduced by exposure to riluzole, A841720 and Bay 36-7620 (Figure 11A). These results were also confirmed by BrdU cell proliferation ELISA (Figures S5A and B).
We further tested the involvement of riluzole, A841720, and Bay 36-7620 in TIVE and TIVE-LTC, BJAB and BCBL-1 cell proliferation by BrdU cell proliferation ELISA. No treatment and vehicle treatment were used as controls. As shown in Figure 11B, treatment with riluzole, A841720 and Bay 36-7620 showed a concentration dependent decrease in the proliferation of both TIVE-LTC and BCBL-1 cells (Figures 11B and C). Due to the absence or low level of expression of mGluR1 receptors, only a minimal effect was observed in the proliferation of uninfected TIVE and BJAB cells (Figures S5C and D). To further confirm the effect of inhibitors on cell proliferation, cells treated with riluzole, A841720 or Bay 36-7620 were monitored using a vibrant MTT cell proliferation assay kit. Riluzole, A841720, and Bay 36-7620 caused a 60–70% decrease in cell growth compared to the untreated control (Fig. 11 D and E). Next, we confirmed the role of mGluR1 on the proliferation of infected cells by using mGluR1 shRNA. TIVE- LTC cells transduced with mGluR1 shRNA or control shRNA were assayed for BrdU incorporation. Compared to control shRNA cells, mGluR1-shRNA significantly reduced the proliferation of TIVE-LTC cells (Fig. 11G), indicating that mGluR1 plays a key role in the proliferation of KSHV infected cells.
Collectively, these results suggested that riluzole and mGluR1 antagonists suppressed the binding of glutamate to the receptors of infected cells and thereby arresting the activation of receptors by glutamate leading into the proliferation of KSHV infected cells.
Glutamate release along with autocrine and paracrine glutamate receptor signaling has been demonstrated to accelerate cell proliferation and tumor progression [47], [48]. During the latent phase of KSHV infection, the cytokines and growth factors released into the extracellular milieu play significant roles in the long term proliferation, survival, and maintenance of the infected cells which probably results in KSHV associated malignancies [8], [12], [13], [15], [16], [17], [18], [49]. Our comprehensive studies demonstrating the increased secretion of glutamate into the cytokine milieu in response to KSHV infection suggest that glutamate could be acting as an autocrine and paracrine growth factor during KSHV induced oncogenesis. Secretion of glutamate occurs in uninfected and infected cells, with comparatively low levels in uninfected cells. We have demonstrated that KSHV infection and appropriate viral gene expression are critical for the generation and release of glutamate in the infected cells. As the viral genome persists in a latent state in the infected cells, the expression level of the latent genes may affect glutamate secretion. Our current study clearly suggested a mechanism whereby the latent ORF73 gene expression affect the stability of c-Myc activation and the depletion of which resulted in reduced glutaminase expression and glutamate secretion. This implies that the level of infection and consistent expression of viral genes are required for the continued secretion of glutamate.
Our studies also show that KSHV infected cells induced the highest levels of glutaminase expression and caused a moderate increase in glutamate release. This difference could be attributable to glutamate transporters and the uptake of glutamate into cells. The glutamate taken up by the cells is converted into glutamine via the glutamine synthetase pathway [50]. Since there are several evidences to indicate that glutamate uptake and its enzymatic conversion are significant steps to maintain extracellular glutamate concentration [50], [51], [52], it is possible that expression or functional impairment of glutamate transporters may also be involved in the maintenance of extracellular glutamate levels in the infected cells.
c-Myc has numerous significant effects on cancer cell metabolism by modifying expression of proteins involved in metabolic pathways [53]. It is known to stimulate increased expression of its target proteins and glutaminase expression by transcriptional repression of mir23a/b in cancer cells [34]. The increased c-Myc activity may also significantly alter the metabolism of glutamine in the infected cells. These changes in glutamine metabolism may profoundly influence the synthesis of molecules involved in growth and survival of infected cells. Although increased glutaminolysis is a supplementary source of energy and may provide significant benefits in terms of the survival of the infected cells, they require additional factors for the induction of cell proliferation or transformation. Thus, while the role of increased metabolism and the components involved in metabolism remains to be determined, it is clear from our study that the secreted glutamate is being used to activate mGluR1 which contribute to the proliferation of infected cells.
Interestingly, we report that KSHV infected cells also upregulate the expression of glutamate receptor mGluR1, which in turn results in increased proliferation as a result of glutamate binding to mGluR1 in the infected cells. Enhanced expression of mGluR1, and the intracellular signaling pathway activated by mGluR1, has the ability to induce cell proliferation and oncogenic transformation [35], [36]. Our data provide evidence that mGluR1 is upregulated in in vitro latently infected cells and in vivo patient samples. Mechanistically, mGluR1 overexpression involves relocalization of REST from the nucleus to the cytoplasm and loss of REST expression in the infected cells. Decreased REST expression, relocalization of REST, and degradation of REST are possible adaptations to antagonize REST-mediated effects to accomplish the overexpression of mGluR1 [39], [54], [55]. A remarkable difference in the pattern of REST localization observed in the infected cells indicates that mGluR1 expression may be regulated via the relocalization of REST. Translocation of REST to the cytoplasm relieves the NRSE or RE1 mediated transcriptional repression in the promoter regions of mGluR1 and upregulates its transcription (Figure 12). Another one of our major findings is that the KSHV latent protein Kaposin A is responsible for cytoplasmic relocalization of REST and mGluR1 activation (Figure 12). Kaposin A mediated oncogenesis has been demonstrated in vitro in Rat3 fibroblasts and in nude mice [45], [46]. Previous studies have suggested that Kaposin A regulates oncogenesis by influencing the phosphorylation of signaling molecules involved in cellular processes, such as cell proliferation and gene transcription [42], [46]. Our findings suggest that sequestration of REST in the cytoplasm by Kaposin A modulates phosphorylation-dependent ubiquitination of REST by altering the phosphorylation status of REST (Figure 12). Kaposin A regulates REST phosphorylation at the specific degron sites which are essential for binding to β-TRCP and degradation of REST during oncogenic transformation. Thus, the downregulation of REST, which is seen in actively proliferating cancer cells [56], [57], might be involved in the regulation of mGluR1 and in cellular transformation during KSHV induced cancer development.
Since Kaposin A does not have a known protein kinase domain, how Kaposin A binding to REST induces phosphorylation of REST needs to be elucidated. Several mechanisms are possible to account for the phosphorylation of REST by Kaposin A. Kaposin A has been reported to phosphorylate a number of kinases involved in cell proliferation [46]. Therefore, it is possible that Kaposin A may couple through one of these kinases for the activation of REST and recruitment of β-TRCP. It is also possible that the interaction of Kaposin A with REST may induce the phosphorylation of REST by allowing a conformational change. These modifications would create a favorable molecular environment for the cross talk between REST and β-TRCP.
In addition to Kaposin A mediated mGluR1 expression, we observed that over expression of LANA-1 also lead to a slight increase in mGluR1 expression. This increase in mGluR1 expression may be a result of an alteration in the N-terminal repressor domain of REST on the mGluR1 promoter. The N-terminal repression domain of REST represses target gene expression by recruiting the transcriptional corepressor mSin3 and then forming a complex with histone deacetylase (HDAC) [58], [59]. Since LANA-1 has already been shown to be associated with mSin3 co-repressor as well as with HDAC [60], it is expected that LANA-1 may be able to bypass REST mediated repression by sequestration of the mSin3/HDAC complex which results in the expression of mGluR1 genes. It is also known that mSin3/HDAC regulated repression is not sufficient for complete transcriptional repression of REST target genes [58], [59], [61]. Therefore, the slight induction of mGluR1 expression in LANA-1 expressing cells could be due to the partial derepression of REST target genes by LANA-1.
Glutamate receptor antagonists and glutamate release inhibitors were shown to be effective in suppressing the proliferation of non-neuronal cancer cells [62], [63]. Identification of the activity of glutamate and mGluR1 in glioma and melanoma development has been the rational approach for testing glutamate release inhibitors talampanel and riluzole in clinical trials for the treatment of glioma and melanoma, respectively [64], [65]. Our functional data shows that the increased production of glutamate and expression of mGluR1 in response to KSHV infection promotes the proliferation of infected cells. Several studies have demonstrated that the signaling pathways activated by mGluR1 contribute to the proliferation and survival of cancer cells [22], [66]. Further studies are essential to determine the role of glutamate and mGluR1 activity in signal induction, viral gene expression, and viral genome maintenance in cells latently infected with KSHV. The blocking effect of riluzole, and the mGluR1 antagonists on proliferation of KSHV infected cells suggests that these molecules could potentially be used for the treatment of KSHV associated malignancies by directly targeting the glutamatergic system in the infected cells.
Primary human dermal microvascular endothelial cells (HMVEC-d cells CC-2543) were purchased from Clonetics, Walkersville, MD. KSHV negative B-lymphoma cell line BJAB, and the KSHV latently infected B-cell line BCBL-1, were obtained from ATCC. BJAB-KSHV (KSHV–GFP recombinant virus in BJAB) was a gift from Dr. Blossom Damania (University of North Carolina, Chapel Hill). TIVE (telomerase-immortalized vein endothelial cell line) and TIVE LTC cells (TIVE cells carrying KSHV in a latent state) were a gift from Dr. Rolf Renne (University of Florida). These cell lines were maintained as described previously [67].
Induction of the KSHV lytic cycle with TPA in BCBL-1 cells, and KSHV purification procedures have been previously described [68]. UV-treated replication-defective KSHV was prepared by exposing the purified virus stock to UV light (365 nm) for 20 min at a 10-cm distance. KSHV DNA was extracted from live KSHV and UV-treated KSHV, and the copies were quantitated by real-time DNA PCR using primers amplifying the KSHV ORF73 gene as described previously [7]. Unless stated otherwise, primary cells were infected with KSHV at 50 MOI (multiplicity of infection) per cell at 37°C.
Rabbit anti-mGluR1 and β-TRCP antibodies as well as mouse anti-BrdU antibodies were from Cell Signaling, Beverly, MA. Mouse anti-glutaminase and rabbit anti-mGluR1 and -TATA binding protein (TBP) antibodies were from Abcam, Cambridge, MA. Mouse anti-tubulin and β-actin antibodies were from Sigma, St. Louis, MA. Mouse anti-c-Myc (9E10) and REST antibodies were from Santa Cruz, Santa Cruz, CA. Rat anti-Kaposin A/C and mouse anti-polyubiquitin antibodies were from Millipore, Temecula, CA. Mouse anti-ORF73 antibodies were generated in Dr. Chandran's laboratory. Anti-rabbit and anti-mouse antibodies linked to horseradish peroxidase were from KPL Inc., Gaithersburg, Md. Alexa 488 and 594 conjugated secondary antibodies were from Invitrogen. Protein A and G–Sepharose CL-4B beads were from Amersham Pharmacia Biotech, Piscataway, NJ. Lambda phosphatase (λPPase), and L-DON (6-diazo-5-oxo-norleucine) were from Santa Cruz. Riluzole, A841720 and Bay 36-7620 were from Tocris Bioscience, Minneapolis, MN.
Plasmids encoding FLAG-tagged human REST wild-type and site-specific REST mutant plasmids (pCMV-FLAG-REST-S1024A, pCMV-FLAG-REST-S1027A, pCMV-FLAG-REST-S1030A), wild type FLAG-REST and triple mutant FLAG-REST-S1024/1027/1030A cloned into retroviral vector pQCXIN were provided by Dr. Stephen Elledge [41] (Harvard Medical School). Lentiviral constructs of KSHV ORF71 (vFLIP), ORF72 (vCyclinD), ORF73 (LANA-1) and ORFK12 (Kaposin A) were obtained from Dr. Chris Boshoff at the UCL Cancer Institute [69]. A plasmid encoding c-Myc shRNA sequence (plasmid #29435) was from Addgene. Transfection was performed using 5 µg of plasmid DNA and lipofectamine 2000 (Invitrogen) as per the manufacturer's instructions.
Lentivirus was produced by transfection with a four-plasmid system, as previously described [70]. Briefly, 293T cells were transiently transfected with lentiviral constructs and the plasmid packaging system (Gag-Pol, Rev and VSV-G), the supernatants were collected, and filtered. Infections were carried out by incubating the virus preparation with cells in the presence of polybrene. The infection efficiency was estimated by analyzing GFP-expressing lentiviral vectors as positive controls. The expression levels of transduced viral genes were assessed by real-time PCR.
For mGluR1 knockdown, lentiviruses encoding mGluR1 shRNA or control shRNA were purchased from Santa Cruz Biotechnology. TIVE-LTC cells were transduced with control lentivirus shRNA and mGluR1 lentivirus shRNA according to the manufacturer's instructions and selected by puromycin hydrochloride.
An equal number of uninfected and infected cells were used for the experiments. Supernatants harvested at different times were centrifuged and glutamate levels were determined in 96-well plates by using a glutamate assay kit as per the manufacturer (Biovision, Mountain View, CA). The concentration of glutamate was determined by measuring the absorbance at 450 nm with a microplate reader.
Total RNA was isolated with TRIzol Reagent (Invitrogen) and treated with DNase I (Ambion) at 37°C for 30 min. Reverse transcription was performed using a High-Capacity cDNA reverse transcription kit (Applied Biosystems). Regular PCR for mGluR1 was performed using 5 µl of the synthesized cDNA using appropriate forward and reverse primers as described by Choi et al [71]. PCR primers were as follows: mGluR1 5′-GTGGTTTGATGAGAAAGGAG-3′ (forward) and 5′-GTTGCTCCACTCAAGATAGC-3 (reverse). β-actin 5′-GCTCACCATGGATGATGATATCGCC-3′ (forward) and 5′GGATGCCTCTCTTGCTCTGGGCCTC-3′ (reverse).
Quantitative real time-PCR was performed with SYBR Green and an ABI prism 7000 sequence detection system (Applied Biosystems, Foster City, CA). The comparative Ct method was used to quantitate gene expression relative to the uninfected control. The following primer set was used: REST (forward 5′-GAGGAGGAGGGCTGTTTACC-3′; reverse 5′-TCACAGCAGCTGCCATTTAC-3′).
Primers used for qRT-PCR of viral genes: ORF71 (forward 5′-AGGTTAACGTTTCCCCTGTTAGC-3′; reverse, 5′-AGCAGGTCGCGCAAGAG-3′), ORF72 (forward-5′-AGCTGCGCCACGAAGCAGTCA-3′; reverse, 5′-CAGGTTCTCCCATCGACGA-3′), ORF73 (forward 5′-CGCGAATACCGCTATGTACTCA-3′; reverse 5′-GGAACGCGCCTCATACGA-3′), Kaposin A (forward 5′ GGATAGAGGCTTAACGGTGTTT-3′; reverse 5′-CAGACAAACGAGTGGTGGTATC-3′).
A pool of two siRNAs synthesized by Integrated DNA technologies (IDT) were used to knockdown Kaposin A. siRNA sequences were as follows: siRNA1- 5′-r(UUGCAACUCGUGUCCUGAAUGCUACGG)-3′, siRNA2-5′- r(CCACAAACACCGUUAAGCCUCUAUCCA)-3′. Cells were transfected with siRNA at 100 pmol (50 pmol each) using siLentFect (Biorad) according to the manufacturer's instructions. Cell lysates were collected at 48 h post-siRNA transfection for immunoprecipitation and Western blot analysis.
Formalin-fixed, paraffin-embedded tissue samples from healthy subjects and patients with KS and primary effusion lymphoma were obtained from the ACSR (AIDS and Cancer Specimen Resource). Sections were deparaffinized with HistoChoice clearing reagent and rehydrated through ethanol to water. For antigen retrieval, the sections were microwaved in 1 mmol/l EDTA (pH 8.0) for 15 min, permeabilized with 0.5% Triton X-100 for 5 min, and then blocked with blocking solution (Image-iT FX signal enhancer-Invitrogen) for 20′ at RT. Immunostaining was performed using anti-mGluR1 and anti-mouse LANA-1 antibodies, followed by Alexa-488 and Alexa-594 conjugated secondary antibodies. Nuclei were stained with 4′,6-diamidino-2-phenylindole (DAPI) (Molecular Probes, Invitrogen), and stained cells were viewed under a fluorescence microscope with a 20× objective and the Nikon Metamorph digital imaging system.
Cells grown on 8 well chamber slides (Nalge Nunc International) were fixed with 4% paraformaldehyde for 15 min, permeabilized with 0.2% Triton X-100, and blocked with Image-iT FX signal enhancer (Invitrogen) for 20 min. Cells were then incubated with primary antibodies against the specific proteins and subsequently stained with Alexa 488 or 594 conjugated secondary antibodies. Cells were mounted in mounting medium containing DAPI. Images were acquired using a Nikon 80i fluorescent microscope equipped with a Metamorph digital imaging system.
Cells were lysed in RIPA buffer containing 15 mM NaCl, 1 mM MgCl2, 1 mM MnCl2, 2 mM CaCl2, 2 mM phenylmethylsulfonyl fluoride, and protease inhibitor mixture (Sigma). The cell lysates were centrifuged at 13,000× g for 20 min at 4°C. Samples mixed with sample buffer containing β-mercaptoethanol, heated at 95°C for 5 min, and separated by SDS PAGE. The protein samples were then Western blotted with the indicated primary antibodies followed by incubation with species-specific HRP-conjugated secondary antibodies. Immunoreactive bands were visualized by enhanced chemiluminescence (Pierce, Rockford, IL) according to the manufacturer's instructions. To determine the fold change, blots were scanned, and quantified by densitometric analysis (Alpha Innotech Corporation, San Leonardo, CA) and normalized with respect to the amount of β-actin.
For immunoprecipitations, 300–500 µg of cell lysates prepared in RIPA buffer or in NP-40 buffer were incubated with the appropriate primary antibody for 4–8 h with end-over-end rotation at 4°C, and the precipitated proteins captured by Protein A or G-Sepharose. The samples were Western blotted with specific primary and secondary antibodies.
Statistical significance was calculated using a two tailed Student's t-test. P<0.05 was considered significant.
|
10.1371/journal.pgen.1004035 | ESCRT-I Mediates FLS2 Endosomal Sorting and Plant Immunity | The plant immune receptor FLAGELLIN SENSING 2 (FLS2) is present at the plasma membrane and is internalized following activation of its ligand flagellin (flg22). We show that ENDOSOMAL SORTING COMPLEX REQUIRED FOR TRANSPORT (ESCRT)-I subunits play roles in FLS2 endocytosis in Arabidopsis. VPS37-1 co-localizes with FLS2 at endosomes and immunoprecipitates with the receptor upon flg22 elicitation. Vps37-1 mutants are reduced in flg22-induced FLS2 endosomes but not in endosomes labeled by Rab5 GTPases suggesting a defect in FLS2 trafficking rather than formation of endosomes. FLS2 localizes to the lumen of multivesicular bodies, but this is altered in vps37-1 mutants indicating compromised endosomal sorting of FLS2 by ESCRT-I loss-of-function. VPS37-1 and VPS28-2 are critical for immunity against bacterial infection through a role in stomatal closure. Our findings identify that VPS37-1, and likewise VPS28-2, regulate late FLS2 endosomal sorting and reveals that ESCRT-I is critical for flg22-activated stomatal defenses involved in plant immunity.
| Plants deploy plasma membrane immune receptors to survey their environment for potential threats. One of these receptors, FLAGELIN SENSING 2 (FLS2) recognizes bacterial flagellin (flg22) and thereby triggers a multitude of defense responses, enhancing immunity against infectious pathogens. Regulation of the subcellular localization of FLS2 is therefore an important aspect in plant disease resistance. FLS2 is known to shuttle between the plasma membrane and endosomal compartments but enters the late endosomal trafficking pathway upon ligand-dependent activation. A key question is the regulation of activated FLS2 in late endosomal trafficking. Here, we show that FLS2 is internalized into the lumen of multivesicular bodies and discovered by genetic inhibition that this step is regulated by components of the ENDOSOMAL SORTING COMPLEXES REQUIRED FOR TRANSPORT-I (ESCRT-I). Furthermore, we reveal that these ESCRT-I components play crucial roles in plant immunity impacting the flg22-triggered closure of stomata, prominent entry points of pathogenic bacteria, which occurred downstream of the known flg22 responses. These findings highlight the roles of endosomal trafficking in regulating FLS2 subcellular localization and plant immunity.
| The metazoan and plant immune systems deploy pattern recognition receptors (PRRs) at the cell surface to sense a wide range of potentially pathogenic microbes through the presence of distinct pathogen-associated molecular patterns (PAMPs), conserved molecules displayed by microbes [1]. In plants, engagement of PRRs leads to the activation of signaling pathways that include mitogen-activated kinase (MAPK) cascades and a series of defense responses ranging from a rapid burst of reactive oxygen species (ROS) to deposition of callose [1]. FLAGELLIN SENSING 2 (FLS2) encodes the PRR that perceives the bacterial PAMP flagellin (flg22) and is required for immunity against bacteria [1]. Upon binding of flg22 to the receptor, FLS2 signaling pathways are activated by complex formation and phosphorylation between FLS2 and BRASSINOSTEROID INSENSITIVE 1 (BRI1)-ASSOCIATED KINASE 1 (BAK1) [2]. Activated FLS2 is internalized via the endocytic pathway raising the possibility that the pool of signaling FLS2 receptors at the plasma membrane is under tight regulation.
Following uptake from the plasma membrane, endocytosed FLS2 arrives at the SYP61-positive trans-Golgi network (TGN)/early endosomal (EE) compartment and the activated receptor is delivered to late endosomal compartments/multivesicular bodies (LE/MVB), from where it can be sorted for degradation [3], [4]. Endosomal sorting of vacuolar cargo involves the delivery of cargo to the LE/MVBs, and more precisely to the luminal vesicles of these compartments. This has been demonstrated in plants for only few plasma membrane proteins: PINFORMED 1 (PIN1), REQUIRES HIGH BORON 1 (BOR1), and BRI1 [5]–[10].
Ubiquitination of the cytosolic domains of plasma membrane proteins has emerged as a key signal for the delivery of these proteins to the LE/MVBs, and more precisely to the luminal vesicles of these compartments [5]–[8] Upon flg22 elicitation, two E3 ligases, PUB12/13, are recruited to FLS2 in a BAK1-dependent manner, and this promotes ubiquitination of FLS2 [11]. Posttranslational modification with ubiquitin targets proteins for MVB luminal sorting, which allows for the hypothesis that ubiquitination facilitates receptor internalization. This is supported by findings that loss-of-function mutations in BAK1 and application of proteasome inhibitors block FLS2 endocytosis as well as several FLS2-mediated responses [2], [12].
The molecular machinery responsible for sorting ubiquitinated cargo to LE/MVBs is the ENDOSOMAL SORTING COMPLEX REQUIRED FOR TRANSPORT (ESCRT)-0, -I, -II, and –III [13]. The subunits of the ESCRTs are referred to as VACUOLAR PROTEIN SORTING (VPS), and with the exception of ESCRT-0, are highly conserved in plants [14]–[20]. The Arabidopsis VPS4 subunit homologue SKD1 (SUPPRESSOR OF K+ TRANSPORT GROWTH DEFECT 1) was reported to mediate vacuolar sorting of ubiquitinated cargo from the plasma membrane [21], [22], and the SKD1-interacting ESCRT-III related proteins CHARGED MULTIVESICULAR BODY PROTEIN (CHMP) 1A and B are involved in correct vacuolar sorting of PIN1, PIN2 and AUXIN RESISTANT 1 (AUX1) [9], [21]. However, surprisingly little is known about ESCRT-I-mediated cargo sorting and their function in plant processes.
Here, we found that endocytosed FLS2 co-localizes and co-purifies with the ESCRT-I subunit VPS37-1. In vps37-1 knock-out plants, the endocytic pathway was normal but flg22-induced FLS2 endocytosis was reduced. We found that vps37-1 mutants were affected not only in FLS2 internalization but also in the FLS2 localization to the lumen of MVBs indicating compromised vacuolar sorting of FLS2. Mutant vps37-1 plants, and likewise vps28-2, supported enhanced growth of Pseudomonas syringae pv. tomato DC3000 (Pto DC3000), which was associated with compromised flg22-triggered stomatal closure. Neither VPS37-1 nor VPS28-2 is involved flg22-induced ROS burst, MPK activation or callose deposition linking the late endocytic trafficking of FLS2 specifically with defense-associated stomatal closure. Altogether, our findings provide novel aspects of FLS2 endocytosis and plant immunity mediated by ESCRT-I.
To dissect post-internalization trafficking of FLS2, we examined the localization pattern of FLS2-containing ARA6/RabF1-RFP and RFP-ARA7/RabF2b-labelled endosomes. Transient expression of RFP-ARA7/RabF2b by particle bombardment in Arabidopsis leaves and prolonged Wortmannin-treated ARA6/RabF1-RFP transgenic plants revealed different types of endosomal compartments labeled by these Rab5 GTPases; these treatments identified different populations of normal sized and enlarged, ring-like structures (Figure S1). Enlarged structures have been previously identified as a result of homotypic fusion of MVBs and this induced structure allows optical resolution of the outer membrane from the luminal cargo [23]–[26]. We measured RFP fluorescent intensity of transections across these endosomal compartments. These data showed that both Rab5 GTPases localize to the outer membrane of the ring-like structures (Figure S1; 1) indicating both ARA6/RabF1 and ARA7/RabF2b are located at the outer membranes of MVBs, consistent with previous reports [21], [24], [25].
To determine FLS2-GFP location at MVBs, we triggered FLS2 endocytosis and examined sub-compartment localization at enlarged MVBs. We first examined FLS2-GFP endosomes in transiently expressing RFP-ARA7/RabF2b leaf cells (Figure 1A). FLS2 and ARA7/RabF2b derived GFP and RFP signals were overlapping and showed similar single peak GFP/RFP fluorescent intensity curves when measured from normal sized endosomal compartments (Figure 1B). At enlarged MVBs, endosomal FLS2 and ARA7/RabF2b showed different localizations. The FLS2-GFP signal was primarily present in the lumen as a filled circle within the ring-like structure marked by RFP-ARA7/RabF2b signal (Figure 1C). This is also illustrated by the GFP/RFP fluorescent intensity curves of transections across these endosomal compartments revealing GFP peaking between two RFP peaks. When we performed prolonged Wortmannin treatment of FLS2-GFP×ARA6/RabF1-RFP plants, the FLS2-GFP signal was similarly concentrated within the ARA6/RabF1-RFP signal in the lumen of these ring-like structures (Figure 1D, 1E). Thus, as with BRI1, BOR1 and other plasma membrane vacuolar cargo, FLS2 endocytosis via the late endosomal pathway involves trafficking to the lumen of MVBs [9], [10], [24].
FLS2 is ubiquitinated in response to flg22 elicitation and sorting of ubiquitinated plasma membrane proteins into luminal vesicles of MVBs is a process mediated by the ESCRT machinery. Therefore, we sought to determine whether ESCRT components play a role in FLS2 endocytosis and trafficking. ESCRT-I is a heterotrimeric complex composed of the subunits VPS23/ELC, VPS28 and VPS37 [19]. In Arabidopsis, the VPS28-1 subunit was recently shown to localize to the TGN/EE, from which MVB maturation could be observed [8]. As mutants in VPS23/ELC were in Ws-0 background, which lacks a functional FLS2 gene [19], [27], we focused on VPS37-1 and VPS28-2 [18]. To determine if VPS37-1 and VPS28-2 are involved in FLS2 endosomal trafficking, we performed in planta co-localization experiments following flg22 elicitation. Endosomal FLS2-GFP partially co-localized with RFP-VPS37-1 and RFP-VPS28-2, respectively (Figure 2A, S2A), a similar pattern compared to the partial co-localization of FLS2-GFP with ARA7/Rab2Fb- and ARA6/RabF1-positive endosomes at early time points after flg22 treatment in Arabidopsis [3]. This suggests that FLS2 traffics via ESCRT-I-positive compartments along its endocytic route.
To investigate whether co-localization is indicative of an interaction between ESCRT-I components and FLS2, we performed co-immuniprecipitation analysis with VPS37-1. In the absence of flg22, FLS2-GFP was detected only in minor amounts of immunoprecipitated RFP-VPS37-1 (Figure 2B). By contrast, the levels of FLS2-GFP were significantly increased in immunoprecipitated RFP-VPS37-1 upon flg22 elicitation. These results indicate that FLS2 forms an inducible complex with the ESCRT-I subunit VPS37-1 coinciding with their shared endosomal localization.
The observation that activated FLS2 co-localizes and forms a complex with VPS37-1 suggest that this ESCRT component plays a critical role in FLS2 endocytosis and trafficking. To investigate this role, we crossed FLS2-GFP into vps37-1 mutants and examined flg22-induced FLS2 endocytosis. Steady-state expression of FLS2-GFP and localization at the plasma membrane was wild type-like (Figure 3A, 3B, S3A), but following flg22 treatment we observed lower numbers of FLS2-GFP endosomes when compared to wild type plants (Figure 3A). Quantification of flg22-induced endosomes by high-throughput imaging demonstrated that despite a general increase in endosome numbers over time, the total numbers of FLS2-GFP endosomes in the 55–100 minutes following flg22 elicitation was significantly lower in the mutants compared to wild type plants (Figure 3C). FLS2-GFP endosomes present in vps37-1 mutants showed co-localization with RFP-ARA7/RabF2b indicating that, while fewer in number, these vesicles are endosomal compartments of the late FLS2 endocytic trafficking route (Figure S3C; [3]).
Reduced flg22-induced internalization of FLS2 was also observed in vps28-2 mutants (Figure S2E). However, we cannot rule out the possibility that this is caused by reduced FLS2-GFP protein levels in this background, because in several attempts of crosses and transformation we were unable to obtain vps28-2 lines expressing FLS2-GFP at similar levels than wild type, though steady-state expression of endogenous FLS2 was unaltered (Figure S2F). Nevertheless, our experiments collectively show that FLS2-positive endosomes are decreased in ESCRT-I mutant plants implying a defect in trafficking.
To ensure that the reduced number of FLS2 endosomes observed in vps37-1 mutants is not a result of global changes to the endosomal populations, we compared ARA7/RabF2b- and ARA6/RabF1-positive endosomes in mutant lines and wild type. We crossed RFP-ARA7/RabF2b and ARA6/RabF1-RFP into the vps37-1 background and endosome quantification by high throughput imaging indicated that there was no difference between the vps37-1 mutant and wild type (Figure 3D, 3E). This data suggest that steady-state endosomal numbers, at least as revealed by these two EE/LE/MVB markers, are not affected in vps37-1 plants and that the reduction in FLS2 endosomes likely reflects a defect in specific endosomal trafficking of the receptor.
The observed effect on FLS2 endocytosis in vps37-1 mutants could result from inhibited FLS2 trafficking at the plasma membrane. To test this hypothesis, we measured the fluorescence intensities of plasma membrane FLS2-GFP. In contrast to wild type plants, the fluorescence intensity of plasma membrane-resident FLS2-GFP in vps37-1 mutants decreased to a much lesser extent after flg22 treatment (Figure 3F). Depending on its activation status, FLS2 is internalized from the plasma membrane into two distinct trafficking routes [3]. We therefore examined whether vps37-1 mutants were affected in endosomal recycling of the non-activated FLS2 receptor. Treatment with Brefeldin A (BFA) caused the accumulation of FLS2-GFP in so-called BFA-bodies, stained by the endocytic tracer FM4-64 (Figure S3D). When vps37-1 leaves were treated with both BFA and flg22, FLS2-GFP-positive endosomes were detected around the BFA-body (Figure S3D), as previously described in wild type [3]. These observations show that endosomal recycling of the non-activated receptor is not regulated by VPS37-1 in agreement with a role of ESCRT in the delivery of cargo for vacuolar degradation. Additionally, these observations might indicate that flg22-induced FLS2-positive endosomes maintain trafficking along the late endosomal pathway rather than entering recycling trafficking.
Our observation revealed that flg22-induced endocytosed FLS2 localizes to the lumen of MVBs. Since the primary role of the ESCRT machinery is associated with sorting vacuolar cargo at MVB compartments [13], [15], we tested whether FLS2 localization at MVBs is altered in vps37-1 mutants. Using particle bombardment, we transiently expressed RFP-ARA7/RabF2b in leaves of FLS2-GFP × vps37-1 plants and examined the RFP and GFP fluorescence signals at enlarged MVBs. We observed three different types of fluorescent patterns (Figure 4A). At type-1 enlarged MVBs, the FLS2-GFP signal was primarily present in the lumen as a filled circle within the RFP-ARA7/RabF2b-labeled ring-like structure. The FLS2-GFP signal at type-2 enlarged MVBs was present in the lumen but also partially co-localized with the RFP-ARA7/RabF2b signal at the outer membrane of the ring-like structure. By contrast, the FLS2-GFP signal primarily co-localized with the RFP-ARA7/RabF2b signal at the outer membrane of type-3 enlarged MVBs and was almost not present in the lumen of its ring-like structure. This is further evident from measurements of the GFP/RFP fluorescent intensity curves of transections across these three types of enlarged MVBs (Figure 4B). These three types of FLS2 localization at enlarged MVBs might represent different stages of sorting activated FLS2 from the outer MVB membrane to the inner lumen.
We quantified the occurrence of type-1, type-2 and type-3 enlarged MVBs in wild type and vps37-1 plants. Strikingly, type-3 enlarged MVBs were only detected in vps37-1 mutants (Figure 4C). We counted about 26% (n = 65) type-3 enlarged MVBs in this background, whereas we could not identify this type of enlarged MVB in the wild type (Figure 4C). Type-2 enlarged MVBs were observed in both wild type and vps37-1 plants at about 32%. Type-1 enlarged MVBs were also present in both genotypes, but a different numbers. Wild type plants showed about 68% type-1 enlarged MVBs, which was reduced to about 42% in vps37-1 mutants. This result shows that FLS2 localization at MVBs is altered in vps37-1 mutants and indicates that VPS37-1 impacts FLS2 endosomal trafficking associated with sorting processes from the outer membrane to the lumen of MVBs. However, overall endocytic trafficking from the plasma membrane to the vacuole was not impaired in vps37-1 plants, because we did not observe any significant difference in the timing of FM4-64 uptake and vacuolar staining in both genotypes (Figure S3E). Together with no obvious developmental phenotype of vps37-1 plants, this indicates that endocytic trafficking might affect a specific subset of vacuolar cargo in these ESCRT-I mutants.
Given that FLS2 endocytosis is dependent on ESCRT-I this provides an opportunity to dissect which flg22-triggered defense responses are associated with changes in FLS2 trafficking. To test this hypothesis, we examined growth of Pto DC3000 in two independent alleles of vps37-1 and vps28-2 T-DNA insertion lines. Following spray infection of virulent Pto DC3000, bacterial growth was significantly higher in these ESCRT-I mutants compared to Col-0 wild type (Figure 5A, S2B) confirming a role for ESCRT-I in bacterial defense.
A critical layer of immunity control against bacterial infection involves the closure of stomata triggered by PAMPs, an important component of FLS2-mediated immunity that restricts pathogen entry at the pre-invasive level [28]. To address stomatal responses in vps37-1 and vps28-2, we treated cotyledons with flg22 and measured the stomatal apertures. This was compared with aperture measurements under control conditions and upon incubation with abscisic acid (ABA), a well-described hormonal trigger of stomatal closure in drought stress [29], [30]. We observed that flg22-induced stomatal closure was significantly impaired in vps37-1 and vps28-2 plants (Figure 5B, S2C). Because ABA-triggered stomatal closure was wild type-like in all tested genotypes, these results indicate that flg22-triggered stomatal responses were specifically affected (Figure 5B, S2C).
The flg22-induced oxidative burst, an early PAMP response mediated by the NADPH oxidase RbohD, is required for stomatal closure [1], [31] and therefore we examined this response in the ESCRT-I mutants. Wild type, vps37-1 and vps28-2 plants displayed comparable ROS production upon flg22 elicitation (Figure 5C, S2D). Thus, in vps37-1 and vps28-2 plants, the flg22-induced stomatal closure is likely affected at a step downstream of the oxidative burst.
FLS2 signaling activates MPK3 and MPK6, both of which were recently reported to control stomatal closure triggered by flg22 [32]. Flg22-activation of MAP kinases was unaltered in vps37-1 mutants when analyzed in whole plant extracts (Figure S3A). We additionally tested flg22-elicited callose deposition, a late PAMP response, because callose plays roles in plant immunity and has also been implicated in the mechanism of stomatal closure [33]. No significant difference in callose deposition was observed between flg22-treated wild type and vps37-1 leaves (Figure S3B). Taken together, we conclude that VPS37-1 and VPS28-2 function is required for full immunity against bacterial infection while not broadly affecting known FLS2-mediated responses. These ESCRT-I subunits are required for flg22-triggered stomatal closure through a mechanism independent of known components of the FLS2 pathway.
Components of the ESCRT machinery are known to localize at the outer membrane of MVBs coupling the formation of MVB luminal vesicles and sorting cargoes for vacuolar degradation [13], [21], [34]. Despite their essential role in delivering endocytosed cargo to the vacuole, it is surprising that knowledge about the molecular interaction between ESCRT components and cargoes is limited. In plants, so far only the ESCRT-III-related CHMP1A and CHMP1B have been linked with sorting the in planta cargoes PIN1, PIN2, and AUX1 [9]. All of these cargoes are internalized by the recycling endosomal pathway and are mis-localized in chmp1a × chmp1b mutants. Here, we established that VPS37-1 and VPS28-2, two ESCRT-I components, play a role in flg22-induced late endosomal trafficking of FLS2. Our data suggest this is critical for sorting the receptor from the outer membrane to the lumen of MVBs but does not impact recycling endocytosis, ARA6/RabF1- and ARA7/RabF2b-positive endosome numbers nor FM4-64 trafficking to the vacuole.
TOLL and the TOLL-LIKE RECEPTORs (TLRs) are essential PRRs of the metazoen immune systems, and like FLS2, TOLL and TLRs localize to endosomes [35], [36]. Ligand-activated TLR4 is endocytosed for degradation and associates with the ESCRT-0 subunit HRS at MVBs [37]. Likewise, endosomal TOLL is found in a complex with HRS, and knocking down HRS inhibited the degradation of CACTUS downstream of TOLL suggesting that endocytosis contributes to proper TOLL signaling [35]. Our experiments indicate that MVB sorting is disrupted by the vps37-1 knock-out mutation. Indeed, in the absence of VPS37-1, FLS2-GFP was observed at the MVB outer membrane. However, despite this altered sorting of FLS2-GFP, we did not detect any FLS2-GFP at the tonoplast as reported for the PIN proteins in chmp1a × chmp1b mutants [9]. This could suggest that activated FLS2 is inefficiently sorted to the MVB lumen and this impaired or delayed process allowed the visualization of these type-3 endosomes. Inefficient and/or delayed FLS2 endosomal sorting is in agreement with the greater abundance of FLS2-GFP at the plasma membrane observed in vps37-1 mutants, likely resulting from reduced internalization, compared to wild type.
Mutations in ESCRT complex subunits and other related proteins often induce severe developmental defects [9], [19], [38]. However, neither vps28-2 nor vps37-1 showed any obvious developmental phenotype indicating the possibility of genetic redundancy with their respective closely homologous genes VPS28-1 and VPS37-2. Significantly, both mutants were compromised in immunity against a bacterial pathogen. This and our observations that VPS37-1 and FLS2 are found in the same protein complex after flg22 elicitation imply that VPS37-1, and likewise VPS28-2, are involved in a mechanism to control the vacuolar sorting of activated FLS2. It is however possible that ESCRT-I components regulate this process for a number of plasma membrane proteins through a similar mechanism.
Recent reports describe both ubiquitin- and ubiquitin-independent degradation of vacuolar cargo in plants [8], [24]. Ubiquitination of endocytosed cargo has been identified for plasma membrane proteins including FLS2, BOR1 and PIN2 [6], [39]. Both the proteasome inhibitor MG132 and a mutation of a putative PEST degradation signal motif in the FLS2 kinase domain inhibit flg22-induced internalization of FLS2 providing indirect evidence that ubiquitination acts as a signal of FLS2 endocytosis [12], [40]. Further studies will reveal whether ESCRT-mediated sorting of endosomal FLS2 depends on ubiquitination, and whether this requires the function of the PUB12/13 E3 ligases that ubiquitinate activated FLS2 via interacting with BAK1 [11]. However, pub12/13 mutants show increased flg22 responses and resistance to Pto DC3000 infection, in contrast to our findings of enhanced susceptibility in the vps28-2 and vps37-1 mutants. Therefore it is possible that regulation of FLS2 by PUB12/13 is involved at a different step of the endocytosis pathway.
Another purpose of endocytic trafficking of plasma membrane proteins is to control the activated signaling pathways. To date, the intersection between receptor-mediated endocytosis and signaling in plants has been best studied for BRI1 [10], [41], [42]. A recent report blocking BRI1 internalization by targeting clathrin demonstrates that BRI1 mostly signals from the plasma membrane contrasting an earlier study revealing BRI1 signaling from endosomes [41], [42]. There is accumulating evidence that effects on the downstream responses can differ depending at which level the inhibition of endocytic trafficking occurs. For example, blocking the internalization of the EPIDERMAL GROWTH FACTOR RECEPTOR (EGFR) at the plasma membrane resulted in an increase of the transcriptional response [43], [44], whereas inhibition of EGFR endosomal sorting by VPS4 knock-down did not increase or alter the overall pattern of the EGF transcriptional response but rather specifically affected a subset of signaling pathways [43], [45]. Relatedly, we found that interference with FLS2 sorting at the level of MVBs compromised specifically FLS2-mediated stomatal closure (Figure S4), which is consistent with the notion that FLS2 activates separate signaling branches [46], [47].
In this study we identified that VPS37-1 and VPS28-2 are required for flg22-induced stomatal closure but not for a range of other flg22-induced defense responses. This implicates post-internalization sorting of FLS2 specifically in PAMP-triggered stomatal closure and thus identifies a role for FLS2 endocytosis in bacterial immunity. Interestingly, immunity in vps37-1 and vps28-2 was not compromised when bacteria were inoculated into the leaf tissue providing further evidence for a prominent role of these ESCRT-I components in stomatal immunity (data not shown). ESCRT-I-mediated stomatal closure is not sensitive to ABA treatment confirming it is part of a FLS2-specific mechanism that involves the late endosomal pathway. Although there is substantial knowledge about ABA-mediated control of stomatal apertures, the exact pathways underlying PAMP-triggered stomatal closure have only been partially described [48]. Current knowledge suggests branching and conversion of separate pathways and some role for ABA in the regulation of stomatal immunity [28], [32]. Flg22-induced stomatal closure in the ESCRT-I mutants is affected downstream of RbohD, MPK3 and MPK6, suggesting regulation could occur at the level of ion channel activity. The K+ channel KAT1 is required for ABA-mediated stomatal closure and also undergoes recycling endocytosis in response to ABA [49]. However, ABA-triggered stomatal closure and endosomal recycling was not affected in the ESCRT-I mutants, thereby indicating that VPS37-1 and VPS28-2 might function in a yet unknown manner in biotic stomatal aperture control. This line with the notion that ABA is primarily sensed by cytosolic receptors while FLS2 is membrane-bound (27,29), and it will be interesting to unravel the molecular mechanism of how endosomal trafficking intersects with stomatal immunity in the future.
This study has revealed that ESCRT-I plays a critical role in late endocytic sorting of FLS2 at the MVB. ESCRT-I is essential for plant immunity to a bacterial pathogen, specifically via flg22-induced stomatal closure. This identifies a role for FLS2 late endocytic trafficking in the intitiation of specific defense responses and the existence of an independent mechanism for flg22-induced stomatal closure.
Arabidopsis thaliana plants were grown on general soil (Arabidopsis mix, John Innes Centre, Norwich), for infection assays on Jiffy pellets (Jiffy Products, Norway), or for sterile conditions on Murashige and Skoog medium under 10 hours or 16 hours of light at 20–22°C and 65% humidity. Col-0/FLS2-GFP, ARA6/RabF1-RFP, and RFP-ARA7/RabF2b lines and fls2, mutants have been described previously [3], [12], [19], [21], [27], [50]. Homozygous T-DNA insertion vps28-2, vps37-1 (vps37-1.1) and vps37-1.2 lines were obtained from SAIL, SALK and GABI-KAT populations (Figure S4). Homozygous FLS2-GFP, ARA6/RabF1-RFP, RFP-ARA7/RabF2b lines in vps37-1 and vps28-2 backgrounds were obtained by crosses. FLS2-GFP was stably transformed into vps28-2 mutants or crossed. RFP-VPS28-2 lines were generated in wild type and FLS2-GFP plants by stable transformation as described previously [50]. Constructs used in this study were obtained by PCR amplifying VPS37-1 (AT3G53120), VPS28-2 (AT4G05000) and ELC (AT3G12400) from Col-0 cDNA (Table S1) and cloned into pGWB binary vectors [51] using Gateway (Invitrogen). All constructs were confirmed by sequencing. Nicotiana benthamiana plants were grown under 16 hours of light at 24°C and 45–65% humidity.
FLS2-GFP transient transformation in N. benthamiana, and transient expression of RFP-ARA7/RabF2b by particle bombarding Arabidopsis leaves was done as described [3], [4]. FM4-64 staining in leaves was done as described previously [3]. Roots were incubated with FM4-64 for 30 min at 4°C to ensure simultaneous staining of the plasma membrane followed by 2 h at RT before imaging. Wortmannin and BFA treatments were done as reported before [3]. For combined treatments, FLS2-GFP × ARA6/RabF1-RFP plants were treated with 10 µM flg22 for 45 min to allow the internalization of the activated receptor into late endosomes, followed by 30 µM Wortmannin treatment for 2 h before imaging.
Standard confocal microscopy was performed using the Leica SP5 microscope (Leica, Germany) and high-throughput confocal imaging was performed using the Opera microscope (PerkinElmer, Germany) as described previously [52]. Fluorescence intensity measurements were done by using the Leica SP5 software. A line was defined as region of interest (ROI) that transsected endosomal compartments and the fluorescence intensity of was determined per pixel along the ROI.
Bacterial inoculation assays were performed as previously described [53]. Briefly, Pto DC3000 was sprayed onto leaf surface at 108 cfu/ml. Disease symptoms and bacterial numbers were scored at 1 and 4 days post inoculation. Surface-sterilized leaf disks from two leaves and at least four plants per genotype were excised and subjected for extraction to determine bacterial titers.
ROS assays were performed as described previously [31]. Briefly, 16 leaf discs were excised per genotype of four-week-old plants and triggered with 1 µm flg22. ROS was measured with a Varioskan multiplate reader (Thermo Fisher Scientific, USA) for 35 min. MAPK activation was detected by immunoblot analysis using anti-phospho p44/p42 MAPK as previously described [54]. Inhibition of seedling growth was measured as previously described [54]. For callose induction, flg22 were applied at 1 µM for 24 hrs, and callose deposits were stained with aniline blue and visualized as described before [55]. Images were taken with the Axiophot microscope (Zeiss, Germany) and quantification of callose deposits was done using the Acapella software [56]. Flg22-triggered stomatal closure was essentially done as described previously [31], images were taken with the Opera microscope (PerkinElmar, Germany), and stomatal apertures were measured as the width/length ratio using the Acapella software.
Immunoblot analysis with the indicated antibodies was performed as described before [53]. Pull-down experiments were carried out as previously reported [57] with the following modifications: Transient transformed N. benthamiana leaves were infiltrated with flg22 solution (43), and subjected to protein extraction in 50 mM Tris-HCl, pH 7.5; 150 mM NaCl; 10% glycerol, 2 mM EDTA, 5 mM DTT, 1% Triton ×100; 1% (vol/vol) protease inhibitor mixture (Sigma). Following filtration through Miracloth (Calbiochem) and centrifugation at 8000 rpm and 16.000 rpm each for 15 min, 5 µl per g fresh weight of GFP-Trap coupled to agarose beads (Chromotek) were added to the supernatant and incubated for 2 hours, washed four times, boiled for 5 min at 65°C in extraction buffer and subjected to immunoblot analysis.
vps37-1.1 (SAIL_97_H04), vps37-1.2 (SALK_042859), VPS37-1 (At3g53120), vps28-2 (SALK_040274), VPS28-2 (AT4G05000).
|
10.1371/journal.pntd.0004993 | Culex quinquefasciatus from Rio de Janeiro Is Not Competent to Transmit the Local Zika Virus | The Americas have suffered a dramatic epidemic of Zika since May in 2015, when Zika virus (ZIKV) was first detected in Brazil. Mosquitoes belonging to subgenus Stegomyia of Aedes, particularly Aedes aegypti, are considered the primary vectors of ZIKV. However, the rapid spread of the virus across the continent raised several concerns about the transmission dynamics, especially about potential mosquito vectors. The purpose of this work was to assess the vector competence of the house mosquito Culex quinquefasciatus from an epidemic Zika area, Rio de Janeiro, Brazil, for local circulating ZIKV isolates.
Culex quinquefasciatus and Ae. aegypti (positive control of ZIKV infection) from Rio de Janeiro were orally exposed to two ZIKV strains isolated from human cases from Rio de Janeiro (Rio-U1 and Rio-S1). Fully engorged mosquitoes were held in incubators at 26 ± 1°C, 12 h:12 h light:dark cycle and 70 ± 10% humidity. For each combination mosquito population—ZIKV strain, 30 specimens were examined for infection, dissemination and transmission rates, at 7, 14 and 21 days after virus exposure by analyzing body (thorax plus abdomen), head and saliva respectively. Infection rates were minimal to completely absent in all Cx. quinquefasciatus-virus combinations and were significantly high for Ae. aegypti. Moreover, dissemination and transmission were not detected in any Cx. quinquefasciatus mosquitoes whatever the incubation period and the ZIKV isolate. In contrast, Ae. aegypti ensured high viral dissemination and moderate to very high transmission.
The southern house mosquito Cx. quinquefasciatus from Rio de Janeiro was not competent to transmit local strains of ZIKV. Thus, there is no experimental evidence that Cx. quinquefasciatus likely plays a role in the ZIKV transmission. Consequently, at least in Rio, mosquito control to reduce ZIKV transmission should remain focused on Ae. aegypti.
| The pandemic Zika epidemic has affected nearly all American countries. The etiological agent is a mosquito-borne-virus originated from Africa that spread to Asia and more recently, to the Pacific region and the Americas. We experimentally demonstrated that the common house nightly biting mosquito Culex quinquefasciatus from Rio de Janeiro was not susceptible to locally circulating Zika virus (ZIKV) strains. Dissemination was not observed in Cx. quinquefasciatus regardless of the ZIKV isolate used and the incubation period after the ingestion of an infected blood meal. No infectious ZIKV particle was detected in the saliva of the four Cx. quinquefasciatus populations examined until 3 weeks after virus exposure. In contrast, we confirmed that local Aedes aegypti mosquitoes can be infected, disseminate ZIKV at significantly high rates, and assured moderate to very high viral transmission after day 14 of virus exposure. We concluded that Cx. quinquefasciatus is not competent to transmit local ZIKV. Our results support that mosquito control should focus on Ae. aegypti to reduce Zika transmission.
| A Zika virus (ZIKV) epidemic has rapidly spread throughout tropical and subtropical zones of the American continent since early 2015 [1]. Brazil was likely the starting point of the Zika pandemic in the Americas [2, 3]. The Zika virus pandemic has spread to North America too. By July 2016, 45 American countries or territories have already reported active ZIKV transmission (http://www.cdc.gov/zika/geo/active-countries.html).
ZIKV is a positive-sense, single-stranded RNA mosquito-borne-virus of 10,807 nucleotides belonging to family Flaviviridae, genus Flavivirus. It is composed of three major lineages: East African, West African, and Asian [4]. ZIKV was first isolated from a sentinel rhesus monkey in the Zika forest in Uganda in 1947 [5]. The second ZIKV isolations were obtained from 20 pools of the forest canopy feeder mosquito Aedes (Stegomyia) africanus captured in the same area [6].
Almost 70 years have passed and little is known about natural ZIKV vectors. Aedes mosquitoes are considered the primary vectors of ZIKV in Africa with reported viral isolations from several species, especially from Ae. africanus [1, 7–10]. ZIKV was also isolated from several other mosquito species belonging to genus Aedes (subgenera Stegomyia and Diceromyia), Mansonia and Culex, and horse flies from the wild in Uganda [8]. More recently, natural infections screened by molecular methods in sylvatic African mosquitoes were again predominantly found in Aedes belonging to subgenus Stegomyia, but also in other species of Aedes, Mansonia, Culex, Anopheles [9, 10]. Nevertheless, ZIKV transmission in the wild has remained poorly understood. Only two sylvatic species (Ae. vittatus and Ae. luteocephalus) proved to be able to transmit ZIKV in laboratory assays [11].
The domestic mosquito Ae. (Stegomyia) aegypti was early shown to be competent to experimentally transmit ZIKV [12]. Due to its high anthropophilic and domestic behaviors and virus detection in field caught specimens [13, 14], this mosquito has been incriminated as the urban and periurban vector in Africa and Asia [1,15].
ZIKV has only recently emerged outside of its natural distribution in Africa and Asia, and has caused a series of epidemics in urban and periurban sites on Pacific islands [16–20] before reaching the Americas, probably in 2013 [21]. The spreading virus belonged to the Asian genotype [21]. Despite multiple efforts, mosquito vectors involved in the ZIKV outbreaks across the Pacific Ocean in 2007–2015 were not identified. Ae. aegypti and other local members of subgenus Stegomyia (Ae. hensilli and Ae. polynesiensis) were thought to be potential vectors [16, 22, 23]. Ae. (Stegomyia) albopictus was found naturally infected with ZIKV in urban sites in Gabon in 2007 [24] and Mexico (http://www.paho.org/hq/index.php?option=com_docman&task=doc_view&Itemid=270&gid=34243&lang=en). Additionally, Ae. aegypti from Singapore were competent to transmit the African ZIKV genotype in the laboratory [25]. Thereafter, Ae. albopictus has been considered a potential vector of ZIKV throughout its geographical range, concomitantly or not with Ae. aegypti [1, 24, 26, 27].
With the arrival of the ZIKV Asian genotype in the Americas, the global number of suspected and confirmed ZIKV cases reached levels never seen previously [28, 29]. Besides, the rapid geographical spread, the increased incidence of severe congenital troubles, such as microcephaly, and Guillain-Barré syndrome associated with ZIKV in Brazil led the World Health Organization to declare the ZIKV epidemic a Public Health Emergency of International Concern [1, 30]. ZIKV proved to have a high potential for geographic expansion in regions wherever Ae. aegypti mosquitoes are present, concomitantly with Dengue viruses 1–4 and Chikungunya virus prone areas of transmission, as it has occurred in Brazil and other American tropical and subtropical countries [29, http://www.cdc.gov/zika/geo/active-countries.html]. American Ae. aegypti and Ae. albopictus populations showed to be competent to transmit the ZIKV belonging to the circulating genotype, but displayed heterogeneous infection, dissemination and transmission rates in laboratory assays [26]. However, Ae. aegypti and Ae. albopictus populations from Brazil and USA exhibited low transmission efficiency to ZIKV [26], which appeared inconsistent with the rapid Zika spread throughout the Americas. Two main hypotheses might explain this scenario: (1) The large number of humans susceptible to ZIKV combined with high densities of anthropophilic Aedes mosquitoes compensate their relatively low vector competence to ZIKV [26]. (2) Although the recent ZIKV pandemic has occurred only in Stegomyia-infested zones and Ae. aegypti has been suggested to be the main vector, other anthropophilic, domestic and usually abundant mosquitoes such as Culex species could contribute to ZIKV transmission [1, 31]. For example, Culex species belonging to the Pipiens Assemblage [32], such as Cx. quinquefasciatus, were likely candidate due their high human-biting frequency and distribution in urban epidemic centers (http://www.reuters.com/article/us-health-zika-brazil-idUSKCN0W52AW). There is no information whether Cx. quinquefasciatus can transmit the virus or the potential role of this mosquito in the ZIKV transmission in nature. We herein comparatively assess the vector competence of Cx. quinquefasciatus and Ae. aegypti populations from Rio de Janeiro for two local ZIKV isolates.
Cx. quinquefasciatus populations tested in this study were collected from four districts of Rio de Janeiro: Manguinhos (MAN, 22°52’20”S 43°14’46”W), Triagem (TRI, 22°53’56”S 43°14’44”W) Copacabana (COP, 22°58’8.3”S 43°11’21”W) and Jacarepaguá (JAC, 22°57’42”S 43°24’11”W). For comparison, we used two populations of Ae. aegypti from Rio de Janeiro, Brazil: Urca (URC, 22°56’45”S 43°09’43”W) and Paquetá (PAQ, 22°45’44”S 43°06’26”). The mosquitoes were concurrently collected as larvae or eggs using ovitraps from January to March 2016 to initiate laboratory colonies. Each colony was started with at least 200 field-collected individuals from more than five distinct collecting sites and traps. Field collected larvae and eggs were hatched and reared in insectaries (26 ± 1°C; 70 ± 10% RH; 12 h:12 h light:dark cycle). Larvae were reared in pans (~100 larvae/pan measuring 30 x 21 x 6 cm) containing 1 liter of dechlorinated tap water supplemented with yeast tablets. Adults were kept under the same insectary controlled conditions described above, and supplied with a 10% sucrose solution. All experimental oral infections were performed with mosquitoes of the F1 generation, except for TRI (laboratory colony) and PAQ (F2).
Mosquitoes were challenged with two ZIKV strains of the Asian genotype, named Rio-U1 and Rio-S1, respectively isolated from urine and saliva of two patients in January 2016, living in distinct districts in Rio de Janeiro [33]. The viral samples were isolated, kept anonymized and provided by Bonaldo et al. [33], whose the institutional review board at Fundação Oswaldo Cruz has previously approved their study protocol. Viral stocks were obtained after two passages of the isolates onto Vero cells maintained with Earle’s 199 medium supplemented with 5% fetal bovine serum (FBS), under an atmosphere containing 5% CO2, and incubated at 37°C. Viral titer in supernatants were estimated by plaque-forming unit (PFU) assays of serial dilutions on Vero cells maintained at 37°C for 7 days and expressed in PFU/mL. Samples were kept at -80°C until use. The comparison of genomic sequences of ZIKV strains Rio-U1 (KU926309) with Rio-S1 (KU92630) yielded 99.6% identity, displaying six amino acid variations in the viral proteins. Phylogenetic analysis showed 99.7% identity of Rio-U1 and Rio-S1 strains with ZIKV isolates from Guatemala and other Brazil regions, including a Zika-associated microcephaly case. They all cluster (bootstrap score = 97%) within the Asian genotype and share a common ancestor with the ZIKV strain that circulated in French Polynesia in November 2013 [33].
Five to seven day-old females were isolated in feeding boxes and starved for 24 h and 48 h for Aedes and Culex mosquitoes, respectively. All mosquitoes were exposed to the infectious blood-meal containing a final viral titer of 106 PFU/mL which consists of a mixture of two parts of washed rabbit erythrocytes and one part of the viral suspension added with a phagostimulant (0.5 mM ATP). Females were fed through a pig-gut membrane covering the base of glass feeders containing the infectious blood-meal maintained at 37°C. Mosquito feeding was limited to 60 min. Only fully engorged females were incubated at 26°C constant temperature, 70 ± 10% RH and 12 h:12 h light:dark cycle, with daily access to 10% sucrose solution. When available, samples of 30 mosquitoes of each population were examined at 7, 14 and 21 days after virus exposure, hereinafter abbreviated as “dpi”.
Mosquitoes were individually processed as follows: abdomen and thorax (herein after referred to as body) were examined to estimate viral infection rate, head for dissemination and saliva for transmission. Each female was handled at a time, by using disposable and disinfected supplies to avoid contamination between individuals and between tissues of the same mosquito as previously described [34]. For the determination of viral infection and dissemination rates, each mosquito body and head were respectively ground in 500 μL and 250 μL of medium supplemented with 4% FBS, and centrifuged at 10,000 x g for 5 min at +4°C before titration. Body and head homogenates were serially diluted and inoculated onto monolayers of Vero cells in 96-well plates. After 1 h incubation of homogenates at 37° C, 150 μL of 2.4% CMC (carboxymethyl cellulose) in Earle’s 199 medium was added per well. After 7 days incubation at 37° C, cells were fixed with 10% formaldehyde, washed, and stained with 0.4% crystal violet. Presence of viral particles was assessed by detection of viral plaques. Additionally, body and head homogenates were individually submitted to specific ZIKV RNA detection and quantification through RT-qPCR, using the SuperScript III Platinum one-step RT-qPCR (Invitrogen) in QuantStudio 6 Flex Real-Time PCR System (Applied Biosystems). For each reaction, we used 600 nM forward primer (5’-CTTGGAGTGCTTGTGATT-3’, genome position 3451–3468), 600 nM reverse primer (5’-CTCCTCCAGTGTTCATTT-3’, genome position 3637–3620) and 800 nM probe (5’FAM- AGAAGAGAATGACCACAAAGATCA-3’TAMRA, genome position 3494–3517). The sequences of this primer set were provided by Isabelle Lepark-Goffart (French National Reference Centre for Arboviruses, IRBA, Marseille, France). The reverse transcription was performed at 45° C for 15 min. The qPCR conditions were 95° C for 2 minutes, followed by 40 amplification cycles of 95° C for 15 sec, 58° C for 5 sec and 60° C for 30 sec. For each run, numbers of ZIKV RNA copies were calculated by absolute quantitation using a standard curve, whose construction details are described elsewhere [33].
In order to assess the transmission rate (TR) and transmission efficiency (TE), mosquito saliva was collected in individual pipette tips containing 5 μL FBS and processed by PFU assays, as previously described [26]. Accordingly, mosquito saliva was inoculated onto Vero Cell monolayer in 6-well plates incubated 7 days at 37° C, under 3 mL with 2.4% CMC in Earle’s 199 medium overlay, and stained as described above. Viral titers of saliva were expressed as PFU/saliva.
Infection rate (IR) refers to the proportion of mosquitoes with infected body (abdomen and thorax) among tested mosquitoes. Disseminated infection rate (DIR) corresponds to the proportion of mosquitoes with infected head among tested mosquitoes (i.e.; abdomen/thorax positive). Transmission efficiency (TE) represents the proportion of mosquitoes with infectious saliva among the initial number of mosquitoes tested. Transmission rate (TR) represents the proportion of mosquitoes with infectious saliva among mosquitoes with disseminated infection.
To compare the viral load, the Wilcoxon signed rank test was adopted to analyze pairwise comparison at 7, 14 and 21 dpi for each mosquito population and tested virus strain. Significant difference was established when p-values were lower than 0.05. Data analyses were conducted with PRISM 5.0 software (GraphPad Software, San Diego-CA, USA, 2007).
This study was approved by the Institutional Ethics Committee on Animal Use (CEUA-IOC license LW-34/14) at the Instituto Oswaldo Cruz. No specific permits were required for performing mosquito collection in the districts in Rio de Janeiro.
We comparatively evaluated the susceptibility to infection of Cx. quinquefasciatus and Ae. aegypti from Rio de Janeiro to two ZIKV strains locally isolated. Infection rates (IR) were negligible to null in Cx. quinquefasciatus, whereas they remained very high for Ae. aegypti, (Fig 1A). With few exceptions, the IRs were of 100% in the two tested Ae. aegypti populations (URC and PAQ) at 14 and 21 dpi, for both virus isolates. In addition, when examining Ae. aegypti from URC, 80% have already been infected by 7 dpi (Fig 1A). In contrast, none of the four Cx. quinquefasciatus populations was likely to become infected except for 1 of 30 TRI Cx. quinquefasciatus challenged with ZIKV Rio-U1, at 14 dpi (viral load: 1,814 RNA copies/ml; 7.0 PFU/ml) (Fig 1A). ZIKV RNA copies (1,453 RNA copies/ml) were detected in 1 of 16 MAN Cx. quinquefasciatus at 14 dpi challenged with the same ZIKV strain. However infective viral particles were not detected in the homogenate of this specimen in repeated PFU assays. Viral load estimated in bodies of Ae. aegypti tended to increase with incubation time (Fig 2), and the lowest values being detected at 7 dpi (median: 1.1 x 106 RNA copies/ml, mean ± SE: 2.3 x 106 ± 2.4 x 106 RNA copies/ml) and the highest at 21 dpi (median: 1.5 x 109 RNA copies/ml, mean ± SE: 1.3 x 109 ± 8.3 x 108 RNA copies/ml). Accordingly, viral load was significantly higher at 21 dpi than at 7 (p = 0.0098) and 14 dpi (p = 0.009). Viral loads at 14 dpi in bodies of Ae. aegypti from PAQ [IR: 100%, Fig 1; viral load: 1.6 x 108 RNA copies/mL (median); 2.6 x 108 ± 2.8 x 108 RNA copies/mL (mean ± SE), Fig 2] were significantly higher than for URC [IR: 90.9%, Fig 1, viral load: 2.1 x 107 RNA copies/mL (median); 2.6 x 108 ± 4.3 x 108 RNA copies/mL (mean ± SE), Fig 2] when challenged with the same ZIKV isolate (Rio-U1).
Cx. quinquefasciatus did not showed viral dissemination regardless of the incubation period whereas dissemination infection rates (DIR) were consistently high (~85–97%) in Ae. aegypti at 14 and 21 dpi irrespective the ZIKV strain (Fig 1B). Accordingly, transmission determined by detecting infective viral particles in mosquito saliva was not observed in any pair of Cx. quinquefasciatus population-ZIKV strain regardless the time point of examination (Fig 1C). In contrast, significantly high transmission rates (TR: 71.6–96.5%) and transmission efficiency (TE: 60.6–93.3%) were observed in local Ae. aegypti (PAC and URC) at 14 dpi (Fig 1C and 1D).
At 14 dpi, viral load in the head of Ae. aegypti from URC infected with ZIKV Rio-S1 (Fig 2) were significantly higher (median: 1.2 x 107 RNA copies/mL; mean ± SE: 1.4 x 107 ± 9.5 x 106 RNA copies/mL) compared to ZIKV Rio U1 (median: 3.6 x 106 RNA copies/mL mean ± SE: 6.3 x 106 ± 7.8 x 106 RNA copies/mL, Fig 2) (p = 0.0003). When challenged with the same ZIKV isolate (Rio-U1), viral load in heads at 14 dpi was significantly higher in Ae. aegypti from PAQ (median: 1.8 x 107 RNA copies/mL, mean ± SE: 3.7 x 107 ± 5.0 x 107 RNA copies/mL, Fig 2) than URC (p = 0.0018). As expected, DIR was lower (DIR = 40%) in Ae. aegypti (URC) at 7 dpi, and no transmission was observed at this time point (Fig 1B–1D). TRs and TEs at 14 dpi were higher for PAQ compared to URC Ae. aegypti challenged with the same ZIKV isolate (Rio-U1) (Fig 1C and 1D), although viral load did not differ (p = 0.4203) between mosquito populations (Fig 3). Also, comparisons of viral loads in saliva of URC Ae. aegypti challenged with different ZIKV isolates did not show any difference (40.3 ± 64.5 PFU Rio-S1/saliva versus 34.2 ± 69.0 PFU Rio-U1/saliva; p = 0.3388) (Fig 3). No significant difference was apparent (p = 0.2212) in viral load in saliva between 14 and 21 dpi (Fig 3).
The Zika epidemics has affected nearly all American countries with ca. 445,000cumulative suspected cases, with 91,962 confirmed infections and 9 deaths due to ZIKV as of August 5, 2016 (http://ais.paho.org/phip/viz/ed_zika_cases.asp). South American countries had nearly 74% of the continental Zika suspected cases, with ca. 5% (165,932 suspected cases) from Brazil. The incidence rate in Brazil is 81.2/100,000 inhabitants Zika suspected cases, with 1,749 cases of microcephaly associated to ZIVK infection diagnosed by clinical, epidemiological and/or laboratory criteria as of May 2016 (http://www.paho.org/hq/index.php?option=com_content&view=article&id=11599&Itemid=41691). Rio de Janeiro is one of the highest risk areas in Brazil, with an incidence of 278.1/100,000 suspected Zika cases as of July 2016 (http://portalsaude.saude.gov.br/images/pdf/2016/julho/15/2016-boletim-epi-n28-dengue-chik-zika-se23.pdf).
To face such a severe health crisis, efficient and effective mosquito control strategies are essential. However, it depends on the definition of primary and/or potential local mosquito vectors. Other ZIKV transmission mechanisms besides Ae. aegypti have been observed. For instance, sexual ZIKV transmission between humans has been observed [35]. Natural ZIKV infections detected in several mosquito genera and even in horse flies would suggest that ZIKV could potentially infect a large range of mosquito species and even other hematophagous flies [31, 33, 36]. However, there is no evidence regarding the role of other mosquitoes or flies besides Aedes (Stegomyia) species in the ZIKV transmission in nature in the Americas. Indeed, there are no data whether other anthropophilic and domestic mosquitoes besides Ae. albopictus, and notably Ae. aegypti can transmit ZIKV.
In this work, we demonstrate for the first time, under laboratory conditions, that Cx. quinquefasciatus are not competent to transmit two ZIKV strains circulating in Brazil. Four tested populations were minimally infected with ZIKV and were unable to transmit this virus. In contrast, two Ae. aegypti populations were highly susceptible to ZIKV infection and dissemination, and competent to transmit the same virus strains. This is consistent with Ae. aegypti being more likely to sustain the current ZIKV outbreak in Rio de Janeiro and probably in other tropical American zones.
The Zika control program in Brazil, as well as in all epidemic American countries, consists essentially in intensifying and reinforcing the current strategies to control dengue for decades, which focuses in reducing Ae. aegypti density and longevity through eliminating or treating potential larval habitats and insecticide spraying (http://www.who.int/tdr/publications/documents/dengue-diagnosis.pdf). However, the traditional vector control strategies have usually failed to efficiently reduce dengue transmission and spread, even when properly adopted [38]. Several reasons have been identified to explain these failures, among which are insufficient community engagement and management and high insecticide resistance in the target species, the mosquito Ae. aegypti [39–41]. Intensifying Ae. aegypti control activities has also been unsuccessful in stemming the rapid spread of ZIKV [1]. Therefore, new technologies are urgently needed to adequately and better mitigate ZIKV transmission, likely requiring combinations of several approaches. For instance, it has been recently demonstrated that Wolbachia-infected Ae. aegypti from Brazil blocks ZIKV transmission [42]. In addition, local control programs should design specific control strategies against the potential vector Ae. albopictus, since it has been shown to transmit ZIKV in laboratory [25, 26, 37], with ZIKV detections in field-collected specimens [24, http://www.paho.org/hq/index.php?option=com_docman&task=doc_view&Itemid=270&gid=34243&lang=en].
The first determination of vector competence to ZIKV in American Ae. aegypti populations was conducted with a viral isolate from New Caledonia, as at the time of that evaluation, no local ZIKV strain were available. Nonetheless, the sequence of NS5 gene of ZIKV from New Caledonia displayed 99.4% identity with ZIKV from Brazil [26]. One Brazilian Ae. aegypti population, from Tubiacanga, Rio de Janeiro were challenged with the ZIKV New Caledonia. High susceptibility to infection and moderate dissemination rate, but with low transmission were found, suggesting unexpectedly low competence of local Ae. aegypti for ZIKV [26]. Our newly data with two Ae. aegypti populations from Rio de Janeiro (URC and PAC) orally challenged with two locally circulating ZIKV isolates (Rio-U1 and Rio-S1) revealed very high dissemination and moderate to high transmission. Similar results were found when testing the URC mosquito population with two ZIKV strains isolated in 2015 from other Brazilian cities [42]. These differences in vector competence may be explained by the concept that the outcome of transmission depends on the specific pairing of vector and virus genotypes [43]. Similar to other ZIKV strains isolated during the epidemic in Brazil, sequences of virus strains used in the present study clustered with Asian clade, including sequences from New World, Malaysia, Micronesia and Pacific. Thus, the New Caledonian [26] and Brazilian strains are genetically nearly identical. Phylogenetic and molecular clock analyses are consistent with a single introduction of ZIKV from the Pacific area into the Americas, probably more than 12 months before the detection of ZIKV in Brazil [21]. It is possible that some genome evolution not yet identified has rapidly shaped ZIKV to New World Ae. aegypti populations, highlighting the genetic specificity and potential for local adaptation between arboviruses and mosquito vectors previously described for dengue [44].
To evaluate the potential role of a mosquito species to transmit an arbovirus like ZIKV requires examination of multiple components governing vectorial capacity, of which vector competence is simply one. Ecological, epidemiological, environmental and climatic factors influence both vector competence and vectorial capacity. Thus, distinct geographical populations of a mosquito species can greatly diverge in their vector competence when exposed to different virus strains, since the outcome of infection depends on the specific combination of mosquito and virus genotypes [45, 46]. Thus, our demonstration that Cx. quinquefasciatus from Rio are not able to transmit ZIKV does not completely rule out the possibility that domestic Culex mosquitoes from other origins may exhibit different vector competence.
Nevertheless, to now at least, there is no evidence that the southern house mosquito Cx. quinquefasciatus is a potential ZIKV vector. Our study with four Cx. quinquefasciatus populations from Rio challenged with two recently isolated virus strains from the same location where mosquitoes were collected showed that this species is not competent to transmit ZIKV. Similar result was obtained when the closely related species Cx. pipiens from USA was challenged with a ZIKV isolated from Puerto Rico [47]. Moreover, besides being incompetent to transmit ZIKV in the laboratory, neither Cx. quinquefasciatus nor any other species of the Pipiens Assemblage has been found naturally infected in the American ZIKV transmission area [48, 49] or during the 2007 Zika outbreaks in the South Pacific island of Yap (Micronesia) [1, 16] and in Gabon [24] where thousands of Cx. quinquefasciatus have been screened.
Therefore, there is no reason to think that mosquito control efforts against Cx. quinquefasciatus to reduce Zika transmission, at least in Rio de Janeiro, Brazil. Mosquito measures to mitigate ZIKV transmission should remain focused on Ae. aegypti.
|
10.1371/journal.ppat.1001080 | Formation of Mobile Chromatin-Associated Nuclear Foci Containing HIV-1 Vpr and VPRBP Is Critical for the Induction of G2 Cell Cycle Arrest | HIV-1 Viral protein R (Vpr) induces a cell cycle arrest at the G2/M phase by activating the ATR DNA damage/stress checkpoint. Recently, we and several other groups showed that Vpr performs this activity by recruiting the DDB1-CUL4A (VPRBP) E3 ubiquitin ligase. While recruitment of this E3 ubiquitin ligase complex has been shown to be required for G2 arrest, the subcellular compartment where this complex forms and functionally acts is unknown. Herein, using immunofluorescence and confocal microscopy, we show that Vpr forms nuclear foci in several cell types including HeLa cells and primary CD4+ T-lymphocytes. These nuclear foci contain VPRBP and partially overlap with DNA repair foci components such as γ-H2AX, 53BP1 and RPA32. While treatment with the non-specific ATR inhibitor caffeine or depletion of VPRBP by siRNA did not inhibit formation of Vpr nuclear foci, mutations in the C-terminal domain of Vpr and cytoplasmic sequestration of Vpr by overexpression of Gag-Pol resulted in impaired formation of these nuclear structures and defective G2 arrest. Consistently, we observed that G2 arrest-competent sooty mangabey Vpr could form these foci but not its G2 arrest-defective paralog Vpx, suggesting that formation of Vpr nuclear foci represents a critical early event in the induction of G2 arrest. Indeed, we found that Vpr could associate to chromatin via its C-terminal domain and that it could form a complex with VPRBP on chromatin. Finally, analysis of Vpr nuclear foci by time-lapse microscopy showed that they were highly mobile and stable structures. Overall, our results suggest that Vpr recruits the DDB1-CUL4A (VPRBP) E3 ligase to these nuclear foci and uses these mobile structures to target a chromatin-bound cellular substrate for ubiquitination in order to induce DNA damage/replication stress, ultimately leading to ATR activation and G2 cell cycle arrest.
| HIV-1, the causative agent of AIDS, encodes several proteins termed accessory, which play a critical role in viral pathogenesis. One of these accessory proteins, viral protein R (Vpr), has been found to block normal cell division. This impairment of cell division by Vpr is thought to increase viral replication and to trigger immune cell death. However, how Vpr is able to block cell growth remains unknown. We and other investigators recently showed that Vpr was performing this activity by interacting with a cellular protein complex involved in ubiquitination. Ubiquitination is characterized by the conjugation of a small protein called ubiquitin to various other proteins to regulate their degradation or activities. In this report, we demonstrate that Vpr forms mobile punctuate structures called foci on the DNA of host cells. We also show that formation of these foci by Vpr is required to block cell division. We propose that Vpr recruits the ubiquitination complex to these nuclear foci and uses these mobile structures to target a DNA-bound cellular protein for degradation, resulting in the activation of a host cell response leading to a cell division block. Identification of the unknown cellular factor targeted by Vpr will contribute to the understanding of the role of Vpr during HIV infection and AIDS pathogenesis.
| HIV-1 encodes several proteins termed accessory that have been implicated in the modulation of host cell environment to promote efficient viral replication and evasion from innate and acquired immunity [1]. One of these accessory proteins, viral protein R (Vpr), is a small amphipathic protein of 96 amino acids. In addition to being expressed in infected cells, Vpr is packaged into virions through an interaction with the p6 domain of the Gag polyprotein precursor [2], [3], [4]. The molecular structure of Vpr was recently resolved and found to consist of a hydrophobic core comprising three interacting alpha helices flanked by N- and C-terminal flexible domains [5]. Of note, the third alpha helix includes a leucine-rich region essential for the stability of the core and the flexible C-terminus comprises a functionally important stretch of positively charged arginine residues [6]. Several biological functions have been attributed to Vpr including transactivation of the viral long terminal repeat (LTR), enhancement of infection in macrophages, induction of apoptosis, and promotion of a cell cycle arrest at the G2/M phase [7].
Vpr-mediated G2 arrest likely plays an important role in vivo for viral replication or pathogenesis given that this activity is highly conserved among primate lentiviruses [8], [9] and since abnormal accumulation of cells in G2/M can be observed in HIV-infected individuals [10]. Indeed, recent studies reported that Vpr upregulated the expression of ligands for the activating NKG2D receptor and promoted natural killer (NK) cell-mediated killing by a process that relied on Vpr ability to induce a G2 arrest, thus suggesting an immunomodulatory role for Vpr that may not only contribute to HIV-1-induced CD4+ T-lymphocyte depletion but may also take part in HIV-1-induced NK cell dysfunction [11], [12]. Several investigators have reported that Vpr-induced cell cycle arrest involves the activation of the ATR (ataxia telangiectasia-mutated and Rad3-related; NM_001184)-mediated G2/M checkpoint [10], [13], [14]. ATR is a kinase of the phosphatidylinositol 3 kinase-like family and is involved in the activation of the G2/M checkpoint and in the coordination of DNA repair following the occurrence of DNA damages or DNA replication stress. Activation of ATR by exogenous DNA damaging agents such as UV leads to phosphorylation of several effector molecules, including Chk1 and H2AX (histone 2A, variant X; NM_002105), inducing the formation of DNA repair foci containing γ-H2AX (phosphorylated H2AX), MDC1 (mediator of DNA damage checkpoint 1), 53BP1 (p53 binding protein 1; NM_001141979), BRCA1 (breast cancer 1), as well as the RPA (replication protein A), 9-1-1 (Rad9-Hus1-Rad1), and Rad17 complexes on the sites of DNA damage [15], [16]. Activation of ATR by Vpr similarly leads to phosphorylation of Chk1 and to the formation of DNA repair foci containing γ-H2AX, 53BP1, RPA, Hus1, Rad17, and BRCA1 [13], [14], [17], [18]. The immediate cause of the activation of ATR following Vpr expression has remained elusive but implicates in part the recruitment by Vpr of the host DDB1 (damage DNA binding protein 1; NM_001923)-CUL4A (cullin 4A; NM_003589) E3 ubiquitin ligase complex via a direct binding to the substrate specificity receptor VPRBP (Vpr-binding protein, also known as DCAF1; NM_014703) [19], [20], [21], [22], [23], [24], [25]. Specifically, RNA interference-mediated depletion of VPRBP or mutations in the hydrophobic leucine-rich core domain of Vpr impaired association to the E3 ligase complex and induction of G2 arrest. In contrast, G2 arrest-defective mutants of Vpr in the C-terminal arginine-rich domain, which maintained their association to the E3 ligase, nevertheless failed to induce G2 arrest [19], [20], [21], [22], [23], [24], [25]. These results indicate that association of Vpr to the E3 ligase complex is required but not sufficient to induce G2 arrest, thus supporting a model in which Vpr would act as a connector between a ubiquitin ligase complex and a yet-unknown cellular protein. We recently provided evidence that Vpr-induced K48-polyubiquitination and proteasomal degradation of this protein(s) would lead to DNA damage/stress, activation of ATR, and ultimately G2 cell cycle arrest [26]. HIV-2 and some species of simian immunodeficiency virus (SIV) encode a paralog of Vpr, called Vpx, which does not induce G2/M arrest but instead counteracts a putative restriction factor expressed in macrophages and dendritic cells that affects infection at a post-entry step [1]. Interestingly, Vpx also interacts with DDB1-CUL4A (VPRBP) via its hydrophobic leucine-rich core domain. This association is required for the inactivation of the restriction factor and probably leads to its proteasomal degradation [27], [28], [29].
The subcellular localization of Vpr and its importance for the induction of G2 arrest has remained a source of controversy. Several investigators reported that Vpr expressed in absence of any other viral proteins primarily localized to the nucleus in a diffuse pattern [30], [31], [32], [33], [34], [35], [36] while others observed a significant accumulation at the nuclear envelope [37], [38], [39], [40], [41], [42]. Of note, Sherman et al. showed that Vpr shuttles between the cytoplasm and nucleoplasm [43]. Moreover, Vpr has been shown to form punctuate structures in the nucleus [17] as well as to induce and co-localize with nuclear membrane herniations [44]. C-terminal mutations impairing G2 arrest did not alter localization of Vpr whereas other mutations, predominantly in the first alpha-helix, impaired both nuclear localization and G2 arrest, implying that nuclear/nuclear-envelope localization of Vpr would be required but not sufficient for this activity [33], [38]. In agreement with this model, Lai et al showed that nuclear punctuate structures formed by Vpr were associated to chromatin and partially co-localized with γ-H2AX, suggesting that Vpr might target host cell DNA and interfere with DNA replication [17]. In contrast, the F34I, V57L, R62P, L68S, and I70S mutations in Vpr caused a re-localization of the protein to the cytoplasm without significantly affecting the induction of G2 arrest [30], [36], [37], [41]. Although inconsistent results were also obtained for some of these mutants [38], these data would suggest instead that Vpr does not induce G2 arrest from the nucleus but from an extra-nuclear compartment.
Therefore, the spatial prerequisites for the induction of Vpr-mediated G2 arrest remain unclear. Additionally, while recruitment of the DDB1-CUL4A (VPRBP) E3 ubiquitin ligase complex has been shown to be critical for G2 arrest, the subcellular compartment where this association occurs and becomes functionally relevant is still unknown. We thus sought to locate the Vpr-VPRBP interaction and to determine the relevance of this localization for the induction of G2 arrest with the goal of furthering our understanding of the mechanism underlying Vpr activation of ATR and providing important information on the potential substrate targeted by Vpr. Herein, we show that Vpr forms nuclear foci that contain VPRBP and that partially co-localize with DNA repair foci components, such as γ-H2AX, RPA32 (replication protein A2, 32kD; NM_002946) and 53BP1. Moreover, we provide evidence that formation of these Vpr nuclear foci constitute a critical early event in the induction of G2 arrest. We also show that Vpr associates to chromatin via its C-terminal domain and that it binds VPRBP on chromatin. Finally, we observed that Vpr foci were highly mobile nuclear bodies. Our results suggest that Vpr recruit the DDB1-CUL4A (VPRBP) E3 ubiquitin ligase complex within mobile nuclear structures to target a chromatin-bound substrate whose ubiquitination and proteolysis would activate ATR and induce G2 arrest.
The interaction between Vpr and VPRBP was previously revealed to be required for the induction of a G2 cell cycle arrest [19], [20], [21], [22], [23], [24], [25]. However, the subcellular localization where this event might take place still remains to be determined. To this end, we performed laser-scanning confocal fluorescence immunohistochemistry to identify the respective subcellular localization and potential co-localization of Vpr and VPRBP. HeLa cells were transduced with a lentiviral vector co-expressing HA-tagged Vpr (HA-Vpr) and GFP or a control lentiviral vector expressing GFP alone. Forty-eight hours after transduction, cells were fixed, permeabilized and stained with antibodies against HA, VPRBP, and nucleoporin. The localization of HA-Vpr was mostly diffuse in the nucleus at standard amplification gain (data not shown). However, when the gain was reduced, we could observe that HA-Vpr formed small circular nuclear structures of variable relative sizes that we refer to thereafter as Vpr nuclear foci (VNF) (Figure 1A). The number of Vpr nuclear foci varied from cell to cell and from experiment to experiment but generally averaged 35 foci (SD±10) per cell. Formation of these foci was not due to the HA tag because we observed that native Vpr could also form nuclear foci (Figure S1A). Endogenous VPRBP was found to be mostly localized to the nucleus in a punctuate pattern (Figure 1A). We observed that HA-Vpr colocalized with endogenous VPRBP in the nucleus. Strikingly, a significant fraction but not all of Vpr nuclear foci co-localized with VPRBP foci, suggesting that Vpr might be able to recruit the E3 ubiquitin ligase complex to these discreet structures. Of note, in presence of Vpr, we also observed some nuclear membrane perturbations reminiscent of the previously described Vpr-induced membrane herniations [44]. Importantly, transduction of activated primary CD4+ T-lymphocytes with a lentiviral vector expressing HA-Vpr also resulted in the formation of Vpr nuclear foci that co-localized with VPRBP (Figure 1B), indicating that these foci are not solely the result of overexpression of Vpr in transformed cell lines but that their formation also occurs in a physiological cellular host of HIV. Infection of HeLa cells with a VSV-G- pseudotyped virus expressing HA-Vpr (HxBru HA-Vpr) also resulted in the formation of Vpr nuclear foci in a minor fraction of cells (Figure S1B). However, the majority of cells displayed a relocalization of HA-Vpr to cytoplasmic compartments (Figure S1B), suggesting that formation of these foci would be a dynamic process, regulated over time during the infection cycle. Indeed, preventing Vpr interaction with Gag and subsequent packaging of the protein into virions using mutations in the p6 domain of Gag (LF/PS) or in Vpr (L23F) [2], [32], resulted in the accumulation of Vpr nuclear foci (Figure S1C). These results provide evidence of the dynamic interplay between Vpr nuclear foci and Gag during infection.
To show that the observed co-localization between Vpr and VPRBP foci was not fortuitous and that Vpr foci truly contained VPRBP, we used an in situ proximity ligation assay (PLA) [45]. This assay is based on the ligation of antibody-coupled DNA molecules when these are in close proximity (when secondary antibodies are less than 400 angströms apart). Amplification of ligation products and hybridization with fluorochrome-labelled probes allow the detection of physiological interaction in situ without the need to overexpress proteins fused to fluorescent markers. Using this technique, HA-Vpr was found in close proximity to endogenous VPRBP in dense nuclear foci (Figure 1C). We did not observe similar interactions in mock-transfected cells (Figure S2), in cells expressing a Vpr mutant (Q65R) impaired for its interaction with VPRBP [19], [20], [22] (Figure S2), or when any of the primary antibodies where omitted (data not shown). These results therefore suggest that Vpr forms nuclear foci containing VPRBP.
To investigate the nature and composition of these Vpr nuclear foci, we first evaluated whether these would correspond to known well-defined nuclear bodies with similar sizes and numbers. We did not however find any significant co-localization with the canonical nuclear speckle marker SC35 (also known as SFRS2) or with PML (promyelocytic leukemia) bodies (Figure S3). Lai et al. previously reported formation and partial co-localization of Vpr nuclear foci with γ-H2AX [17]. We thus evaluated if the Vpr nuclear foci described herein where the same foci that Lai et al. reported. Interestingly, we observed a partial co-localization between HA-Vpr nuclear foci and 53BP1 (Figure 2A). Indeed, expression of HA-Vpr induced the re-localization of 53BP1 from its sites of residence in the nucleus to DNA repair foci, some of which were positive for HA-Vpr foci. We also observed a partial co-localization between some HA-Vpr nuclear foci and phosphorylated RPA32 (Figure 2B). Similar results were obtained for γ-H2AX (Figure S4A).
Co-localization of Vpr with components of DNA repair foci suggest that formation of Vpr nuclear foci might represent an early event in the induction of G2 arrest that would be responsible for the generation of DNA replication stress or DNA damage. Conversely, those might simply reflect the re-organization of the nuclear compartment following the activation of the ATR checkpoint by Vpr. To distinguish between these two possibilities, we transduced HeLa cells with a lentiviral vector expressing HA-Vpr and concomitantly treated the cells with caffeine, a non-specific inhibitor of ATR and ATM (ataxia telangiectasia mutated). In these conditions, the addition of caffeine inhibited Vpr-induced cell cycle arrest (data not shown; [12]). However, we did not detect significant changes in the number of Vpr nuclear foci (Figure 3A, 33±10 for non-treated cells vs 32±9 for caffeine-treated cells), suggesting that formation of these foci would take place independently of the activation of ATR. Moreover, consistent with the observation that not all Vpr nuclear foci co-localized with VPRBP (Figure 1A), depletion of VPRBP by siRNA (95%±3.5% knockdown relative to scrambled siRNA) in HeLa cells (Figure 3B) did not significantly alter the number of foci (36±10 for control siRNA vs 33±8 for VPRBP siRNA) (Figure 3C), suggesting that VPRBP is dispensable for the formation of Vpr nuclear foci. Similar results (data not shown) were obtained in a HEK293T monoclonal cell line stably depleted of VPRBP [26]. Moreover, in contrast to its absence of effect on Vpr foci, knockdown of VPRBP abrogated Vpr-induced formation of DNA repair foci containing γ-H2AX and 53BP1 (Figures S4A and S4B). These results indicate that Vpr forms nuclear foci prior to and independently of the activation of ATR and suggest that it is Vpr that recruits VPRBP to these foci and not the inverse.
To evaluate the potential role of these Vpr nuclear foci in the induction of G2 arrest, we monitored the capacity of several G2 arrest-defective Vpr mutants to form these foci. HeLa cells were transfected with plasmids expressing HA-tagged Vpr mutants and formation of nuclear foci was evaluated by fluorescence immunohistochemistry and confocal microscopy (Figure 4). We found that Vpr (R80A), which still interacts with the E3 ligase but is strongly attenuated for the induction of G2 arrest, was defective for the formation of nuclear foci (2.5±1.1), even though its subcellular localization was nuclear. Deletion of the C-terminus of Vpr (Vpr 1–78), which also maintains the association with the E3 ligase [22] but impairs the induction of G2 arrest [46], similarly resulted in a defect in the formation of nuclear foci (Figure 4). Similar results were also obtained with the C-terminal mutants Vpr (S79A) and Vpr (1–86) (data not shown). Vpr (Q65R), which is unable to associate with the E3 ligase and is consequently defective for G2 arrest, was found to be defective for the formation of nuclear foci and also accumulated in cytoplasmic aggregates. Similar localization phenotypes where observed for Vpr (H71R), a mutant of Vpr also defective for its interaction with VPRBP [21] (data not shown). The results obtained with the Q65R and H71R mutations are in contrast with the siRNA-mediated depletion of VPRBP which did not block the formation of Vpr nuclear foci, suggesting that these mutant proteins might have additional defects besides an impaired interaction with VPRBP (see below). These results thus suggest that the C-terminal domain of Vpr, which is required for the induction of G2 arrest, is also critical for the formation of Vpr nuclear foci.
The observation that C-terminal G2 arrest-defective mutants of Vpr are compromised in their capacity to form nuclear foci suggests that these nuclear foci might constitute an important early event in the induction of G2 arrest by Vpr. To directly address this possibility, we first evaluated the functional effect of artificially sequestering Vpr in the cytoplasm by overexpression of Gag-Pol. Co-transfection of HeLa cells with HA-Vpr- and Gag-Pol-expressing plasmids produced a sequestration of HA-Vpr in p24-positive cytoplasmic compartments (Figure 5A). This sequestration abrogated Vpr nuclear foci formation (Figure 5A). Similar results were obtained in HEK293T cells (data not shown). To evaluate the functional effect of this cytoplasmic sequestration of Vpr, HEK293T cells were co-transfected with plasmids expressing HA-Vpr and Gag-Pol or with adequate empty plasmid controls. Two days later, the cell cycle and expression profiles of transfected cells were evaluated by flow cytometry and western blot (Figures 5B and 5C). Expression of HA-Vpr alone produced an accumulation of cells in G2/M (G2+M:G1 = 1.81 vs 0.66 for mock-transfected cells). Interestingly, overexpression of Gag-Pol completely abrogated HA-Vpr-induced G2 arrest (G2+M:G1 = 0.67) in absence of any significant effect on the cell cycle when expressed alone (G2+M:G1 = 0.77). Inhibition of Vpr-induced G2 arrest by overexpression of Gag-pol was dependent on the Gag-Vpr interaction and was not the result of some non-specific effects on the cell cycle since Vpr (L23F), a mutant of Vpr unable to bind the p6 domain of Gag [2], could form nuclear foci (Figure S5A) but was impervious to the effect of Gag-Pol on Vpr nuclear localization (Figure S5B) and induction of G2 arrest (Figure S5C). Although overexpression of Gag-Pol led to a reduction of the affinity between HA-Vpr and endogenous VPRBP, the overall increase in the expression of HA-Vpr resulted in an increase in the levels of Vpr-bound VPRBP (Figure 5D), excluding the possibility that overexpression of Gag-Pol inhibited G2 arrest by preventing the Vpr-VPRBP interaction. The observed inhibition of G2 arrest by overexpression of Gag-Pol is however unlikely to have a significant role at physiological levels of expression given that infection with a wild type virus led to a G2 arrest that was as efficient as the one obtained with a virus unable to relocalize Vpr from the nucleus because of a mutation in the P6 domain of Gag (LF/PS) (Figures S1C and S5D). Overall, these results imply that nuclear localization of Vpr and possibly the formation of nuclear foci would be required for the induction of G2 arrest.
To further show that the formation of these Vpr nuclear foci is critical for the induction of G2 arrest, we evaluated the capacity of SIV Vpr and its paralog Vpx to form these foci. Both of these proteins are able to associate with the E3 ligase complex but in contrast to Vpr, Vpx does not induce G2 arrest but counteract a putative restriction factor in macrophages and dentritic cells [27], [28], [29]. HeLa cells were transfected with plasmids expressing either HA-tagged sooty mangabey Vpr (HA-Vpr sm) or Vpx (HA-Vpx sm). Two days after transfection, cells were fixed, permeabilized, and stained with antibodies against HA and nucleoporin (Figure 6). Consistent with its ability to induce G2 arrest (data not shown and [9]), we found that Vpr sm could accumulate into nuclear foci (16±4 foci per cell) in contrast to the G2-arrest incompetent Vpx that did not form any foci despite being present in the nucleus (Figure 6).
Taken together, these results indicate that formation of Vpr nuclear foci is an early event that is required to induce G2 arrest. These results also indicate that nuclear localization of Vpr is not sufficient to induce formation of foci.
Given that these foci constitute an early event in the induction of G2 arrest, we sought to determine how they would form. These foci are likely the results of a local observable accumulation of Vpr either through oligomerization of the protein or following its recruitment by a locally abundant tethering factor. To distinguish between these two possibilities, we first monitored the dimerization efficiency of the Vpr mutants Q65R and R80A, which are defective for foci formation. HEK293T cells were co-transfected with plasmids expressing enhanced yellow fluorescence protein (eYFP) fused to the N-terminus of wild type Vpr and renilla luciferase (Rluc) fused to the N-terminus of wild type Vpr and mutants. Two days after transfection, self-affinity was assessed by bioluminescence resonance energy transfer (BRET). Figure 7A reveals that all Vpr fusion proteins were efficiently expressed. In this system, we observed a specific energy transfer between eYFP-Vpr (WT) and Rluc-Vpr (WT) (Figure 7B). The maximum energy transfer at saturation (BRETmax) was 0.983 and the concentration of acceptor at 50% of BRETmax (BRET50) was 0.397. In contrast, co-expression of eYFP and Rluc-Vpr did not lead to any significant energy transfer, demonstrating the specificity of the eYFP-Vpr/Rluc-Vpr interaction. The Q65R mutant, showed a significant decrease in its affinity for wild type eYFP-Vpr (BRET50 = 0.791, 50% self-affinity) as well as a drastic decrease in BRETmax (0.314 for Q65R vs 0.983 for wild type Vpr), suggesting that in addition to a reduction in dimerization efficiency, formation of higher-order complexes (multimerization) would also be synergistically decreased. In contrast, the R80A mutant displayed an affinity for wild type Rluc-Vpr that was at least comparable to wild type Vpr (BRET50 = 0.326, 121% self-affinity relative to wild type). Similar results were obtained when eYFP-Vpr R80A and Rluc-Vpr R80A were co-expressed (data not shown). Thus, these results suggest that the ability of Vpr to oligomerize does not directly correlate with nuclear foci formation and does not explain the defect in foci formation observed in the context of C-terminal mutants. To determine if oligomerization of Vpr could still be involved in formation of Vpr nuclear foci, we performed trans-complementation experiments in HeLa cells and monitored formation of Vpr foci by immunofluorescence confocal microscopy. Trans-complementation of HA-Vpr R80A with eYFP-Vpr could rescue the defective phenotype of the R80A mutant by re-localizing the protein to eYFP-Vpr foci (Figure 7C). In contrast, eYFP-Vpr was unable to re-localize the HA-tagged Q65R mutant (Figure 7C), suggesting that oligomerization of Vpr, although not sufficient to induce formation of Vpr foci, may however contribute to the process to some degree.
Since oligomerization does not fully account for the ability of Vpr to form foci, Vpr could thus be tethered to specific sites by a cellular co-factor. Co-localization of Vpr nuclear foci with chromatin-bound factors detected at DNA repair sites suggests that this tethering co-factor could be a chromatin-bound protein or structure or DNA itself. To assess this possibility, HeLa cells were first transiently transfected with an empty plasmid or a plasmid expressing HA-Vpr and cells were lysed with 0.5% Triton X-100, resulting in the release of soluble proteins. Treatment of Triton-insoluble cellular material, including chromatin, with microccocal nuclease resulted in the solubilization of chromatin-bound cellular proteins including RPA70 (replication protein A1, 70 kDa) (data not shown) and histone 3 (Figure 8A). These proteins were not detected when cell extracts were incubated in buffer without microccocal nuclease. Importantly, chromatin extracts were not contaminated with cytoplasmic proteins as revealed by the absence of GAPDH (glyceraldehyde-3-phosphate dehydrogenase) (Figure 8A). Using this system, we found that a fraction of HA-Vpr was released in extracts treated by microccocal nuclease but not with buffer alone, indicating that Vpr associates with chromatin directly or indirectly via other proteins (Figure 8A). A specific association of a fraction of endogenous VPRBP with chromatin was also observed in presence and in absence of Vpr (Figure 8A). To determine whether the defects of foci formation observed with C-terminal mutants of Vpr would correlate with a defect in chromatin association, we analyzed the capacity of several Vpr mutants to associate to chromatin in HeLa cells. Interestingly, both Vpr (R80A) and a C-terminal deletion mutant (Vpr 1–78) showed a drastic reduction in their association to chromatin (Figure 8B). Of note, Vpr (Q65R) (Figure 8B) and Vpr (H71R) (data not shown) also failed to associate with chromatin, possibly explaining their unexpected incapacity to form foci. In contrast, knockdown of VPRBP did not significantly alter the affinity of Vpr for chromatin (Figure 8C), suggesting that VPRBP does not contribute to this association and that the absence of chromatin association with the Q65R and H71R mutants is not due to its impaired binding to VPRBP. Therefore, the ability of Vpr to form foci correlates with its ability to associate with chromatin.
Co-localization of Vpr nuclear foci with VPRBP and the association of both proteins to chromatin suggest that they might interact on chromatin. To evaluate this possibility, we transfected HeLa cells with an empty plasmid or a plasmid expressing HA-Vpr and performed anti-HA immunoprecipitations on proteins released from chromatin by microccocal nuclease (Figure 9A). Interestingly, we could detect co-immunoprecipitation of endogenous VPRBP specifically in cells extracts containing HA-Vpr, in the soluble fraction as well as in the chromatin fraction (Figure 9A). Deletion of the C-terminal domain of Vpr abrogated its interaction with VPRBP on chromatin but not in the soluble fraction (Figure 9B), demonstrating the specificity of these interactions. These data suggest that Vpr interacts with VPRBP on chromatin. Importantly, histone 3 did not co-immunoprecipitate with HA-Vpr in the chromatin fraction (Figure 9A). Moreover, treatment with high concentrations of ethidium bromide during the immunoprecipitation, a treatment that displace proteins from DNA [47], did not disrupt the Vpr-VPRBP interaction in the soluble fraction as well as on chromatin (Figure 9B), thus excluding the possibility that the observed Vpr-VPRBP interaction was mediated by incompletely digested chromatin fragments.
Nuclear bodies stably or transiently associating with chromatin are generally dynamic structures, either in mobility or in stability. For instance, PML bodies display varying levels of mobility in the nucleus. Conversely, DNA repair foci show limited mobility but can rapidly form in response to genotoxic stress and can disassemble following checkpoint recovery [48], [49], [50]. To investigate the possible dynamic nature of Vpr nuclear foci, we performed time-lapse confocal microscopy in living HeLa cells expressing eYFP-Vpr. Strikingly, observation of eYFP-Vpr foci for two minutes (at two-second intervals) revealed that these were highly mobile structures (Figure 10A; Videos S1 and S2). Software-assisted tracking of over fifty Vpr foci (Figure 10B and data not shown) revealed rates of displacement ranging from 0.05 µm/min to 8.30 µm/min for an average of 0.73 µm/min (SD = 1.00 µm/min; median = 0.30 µm/min). The mobility of Vpr foci was not dependent on the presence of VPRBP since its knockdown (Figure 10C) did not significantly alter their dynamic behavior (average rate of displacement of 0.85 µm/min for VPRBP siRNA vs 0.92 µm/min for scrambled siRNA; P = 0.78). Because some eYFP-Vpr foci seemingly appeared and disappeared during the course of these observations, we performed time-lapse spinning-disk microscopy analyses to evaluate whether these foci were translating in and out of the focal plane or instead assembling and disassembling. Tracking of eYFP-Vpr nuclear foci for 15 minutes at intervals of 5 seconds highlighted translational movements in the three axes (Figure S6A). Moreover, these analyses did not reveal any apparition or disappearance of nuclear foci (data not shown), suggesting that these are structurally stable. Similar results were obtained from observations over longer periods of time (30 minutes). Monitoring of the mean fluorescence of eYFP-Vpr in foci showed relatively stable signal over time (Figure S6A, right panel). Some transient fluctuations in fluorescence were however detected. To determine if these fluctuations could be the result of quick exchange of Vpr molecules in and out of nuclear foci, we performed fluorescence recovery after photobleaching (FRAP) analyses on eYFP-Vpr foci (Figure S6B). Photobleaching of eYFP-Vpr foci did not however lead to any fluorescence recovery even after an extensive period of time (350 seconds), suggesting that the inter-exchange of Vpr molecules is minimal.
Overall, our results suggest that Vpr would associate to chromatin-bound nuclear foci via its C-terminus. These would serve as a mobile scaffold to recruit the DDB1-CUL4A (VPRBP) E3 ubiquitin ligase to induce the ubiquitination and degradation of a chromatin-bound substrate, resulting in DNA damage or replication stress.
Our results show that Vpr mainly localizes to the nucleus in transformed epithelial cells, such as HeLa and HEK293T cells, as well as in primary CD4+ T-lymphocytes (Figure 1 and data not shown). We noticed that the localization of Vpr in HeLa cells closely resembles that observed in primary CD4+ T-lymphocytes, prompting us to select this cellular model for most of our study. Moreover, we found that ectopically expressed HA-tagged Vpr had a subcellular localization similar to that of the native protein (Figure S1A). In infected cells, the nuclear localization of Vpr appears transient because Gag interacts with Vpr to package the protein into assembling viral particles (Figures S1B and S1C). Our localization data show that Vpr can form nuclear punctuate structure that we termed Vpr nuclear foci (Figure 1), as was reported previously by Lai and colleagues [17]. It is noteworthy that these foci are not readily apparent and require careful calibration of gain to be observed (data not shown). Importantly, we observed a strong co-localization of Vpr with VPRBP in the nucleus, particularly in these foci. In situ proximity ligation assays confirmed the close proximity of the two proteins in these foci (Figure 1C), suggesting that Vpr interacts with the E3 ubiquitin ligase at the levels of these punctuate structures. In contrast to the observations of other investigators [37], [38], [39], [40], [41], [42], we did not observe a significant accumulation of Vpr at the nuclear membrane in these cell types. Several technical reasons might explain these discrepancies, including cell types, levels of expression, fixation and permeabilization conditions, or the tag used. Indeed, we did observe an enrichment of eYFP-Vpr at the nuclear membrane of Hela cells (Videos S1 and S2).
We obtained several lines of evidence demonstrating that Vpr nuclear foci are involved in Vpr-mediated G2 arrest. First, we observed a partial co-localization between these foci and RPA32, 53BP1 and γ-H2AX, which are usually detected at DNA repair sites (Figures 2 and S4). Similar results were obtained by Lai and colleagues with γ-H2AX [17]. Secondly, C-terminal mutants of Vpr defective for G2 arrest failed to induce formation of Vpr foci despite their nuclear localization (Figure 4). Thirdly, cytoplasmic sequestration of Vpr by overexpression of Gag inhibited G2 arrest as well as foci formation (Figure 5). Fourthly, only Vpr from sooty mangabey SIV but not its G2 arrest-defective paralog Vpx was able to form these foci (Figure 6). Lastly, the reduced number of foci formed by sooty mangabey Vpr in comparison to HIV-1 Vpr correlated with reduced G2 arrest activity in human cells (data not shown and [9]). All these results suggest that formation of foci is linked to G2 arrest. Moreover, these results also suggest that nuclear localization of Vpr is required but not sufficient to induce formation of these foci. Our results and conclusions are in contrast with previous reports, including one of ours, describing cytoplasmic mutants of Vpr that retain their G2 arrest activity [30], [36], [37], [41]. We had reported over a decade ago that the V57L and R62P mutations induced the relocalization of Vpr to the cytoplasm, while these mutants were still able to induce G2 arrest [36]. However, careful re-examination of the localization of these mutants showed that both mutants could localize to the nucleus to some degree. While, the V57L mutant had a reduced capacity to form foci, the R62P mutant was completely defective for foci formation (Figure S7A). The reduced capacity of V57L mutant and the defect of the R62P mutant in foci formation correlated, respectively, with attenuation and abrogation of G2 arrest (Figure S7B). These differences between our present localization data and our previously published results can probably be explained by improved imaging sensitivity, whereas the discrepancies in G2 arrest activity are unclear. Nevertheless, these results highlight an important technical limitation in these types of localization experiments: lack of detection in a subcellular compartment does not necessarily indicate an absence of protein.
Correlation between G2 arrest and formation of Vpr nuclear foci implies that the formation of these foci could either be an early event leading to G2 arrest or could be a consequence of this G2 arrest. We observed that treatment with the ATR/ATM inhibitor caffeine (Figure 3A) did not abrogate formation of Vpr foci, thus indicating that these foci likely constitute an early event in the induction of G2 arrest by Vpr. In fact, formation of Vpr foci was not affected by an almost complete knockdown of VPRBP suggesting that their formation is independent of the recruitment of the E3 ligase complex and would therefore precede ubiquitination and degradation of the putative G2 arrest substrate (Figures 3B and S4). In contrast, we found that the Q65R mutant of Vpr was unable to form foci. In addition to a reduced affinity for VPRBP [19], [20], [22], this mutation also leads to other defects including accumulation of Vpr in the cytoplasm (Figure 4), reduced dimerization efficiency (Figure 7), and absence of binding to chromatin (Figure 8B), indicating that the Q65R mutation has pleiotropic effects on the functions of Vpr. Yet, this mutation did not prevent efficient packaging of Vpr into virions [12]. Cautions should thus be used when interpreting results obtained with this mutant. Despite these pleiotropic defects, we cannot completely exclude the possibility that, in addition to the C-terminal domain, binding to VPRBP would also contribute to foci formation and chromatin association.
Given that Vpr foci containing VPRBP partially co-localize with chromatin-bound protein such as RPA32 and that Vpr associates with DNA in vitro [51] and in vivo (Figure 8A and [17]), we propose that Vpr might be able to target a chromatin-bound cellular factor. In support of this hypothesis, Lai et al. showed that in situ nuclease treatment of Vpr-expressing cells eliminates Vpr nuclear foci [17], suggesting that Vpr nuclear foci are anchored to chromatin. Deletion of the C-terminal domain of Vpr drastically reduced foci formation (Figure 4) and its chromatin association (Figure 8B). Similar results were obtained by Lai and colleagues [17]. Moreover, mutation of the arginine at position 80 did not affect direct binding to nucleic acids in vitro [51] but nevertheless impaired association to chromatin in vivo (Figure 8B), implying that a cellular factor rather than a direct binding to DNA would be implicated in association to chromatin. This cellular factor does not appear to be VPRBP since its knockdown did not significantly reduce the binding of Vpr to chromatin (Figure 8C). Moreover, we also observed protein-protein interaction between Vpr and VPRBP on chromatin (Figure 9), suggesting that Vpr would be able to recruit the E3 ligase DDB1-CUL4A (VPRBP) onto chromatin.
Strikingly, analysis of Vpr nuclear foci by time-lapse microscopy (Figures 10 and S5, Videos S1 and S2) revealed that these foci moved rapidly in the nucleus (average of 0.73 µm/min). As a comparison, passive diffusion of chromatin-bound DNA repair foci was estimated at 1–2 µm per 6 hours [52]. These results suggest that instead of stably interacting with chromatin, Vpr nuclear foci would do so in a dynamic manner, allowing movement of the foci along chromatin strands. One possible model to integrate all our results is that Vpr could interact with its putative substrate via its C-terminus in these chromatin-bound nuclear foci and could recruit the DDB1-CUL4A(VPRBP) E3 ligase to degrade the substrate, thus preventing its potential role in DNA replication or DNA repair. This model implies that Vpr would initially require binding with the substrate to localize in these nuclear bodies and that the subsequent degradation of this substrate would not exclude Vpr from these structures nor would it disrupt them. Another possibility is that Vpr would interact with a nuclear foci-associated co-factor via its C-terminus and would utilize these mobile structures to scan chromatin for its putative substrate. This second model requires that either Vpr possesses an additional functional domain mediating the interaction with the substrate or that Vpr targets VPRBP's own natural substrates. Irrespective of the above models, as was recently documented, the substrate would be covalently modified with classical K48-linked polyubiquitin chains in a DDB1-CUL4A (VRPBP)-dependent manner and degraded by the proteasome [26]. Moreover, multiple units of the putative substrate/co-factor are probably required in these nuclear bodies in order for Vpr to accumulate in these structures. Even though Vpr multimerization was shown to occur in these foci (Figure 7), it is unlikely that it would play a major role in this process given that the L23F mutation was previously shown to block dimerization [40], [53] but did not significantly affect foci formation and induction of G2 arrest (Figure S5). Similar conclusions were also previously obtained with the I70S mutation which was shown to block dimerization without affecting the induction of G2 arrest [30]. It however remains unclear whether Vpr would bind VPRBP before or after localizing to these foci, particularly when considering the important level of interaction observed in the Triton-soluble fraction (Figure 9). Moreover, the significance of the partial co-localization observed between Vpr and DNA repair foci containing RPA32, 53BP1 and γ-H2AX (Figures 2 and S4) is also unclear. On one hand it could mean that degradation of the chromatin-bound substrate would induce DNA damage or DNA replication stress in situ and that this partial co-localization would be explained by the high mobility of Vpr foci. On the other hand, we cannot exclude the possibility that degradation of the substrate could induce global genomic instability and that this partial co-localization would only be fortuitous.
Overall, our results show that Vpr forms highly mobile nuclear foci containing VPRBP and demonstrate that formation of these structures constitutes a critical early event in the induction of DNA damage/stress and G2 arrest by Vpr. The characterization of these chromatin-bound nuclear foci hijacked by Vpr will likely contribute to better delineate the mechanism by which Vpr activates ATR and induces G2 arrest. Importantly, our results further suggest that the putative cellular substrate targeted by Vpr is likely to be a chromatin-associated protein.
Peripheral blood samples were obtained from adult donors who gave written informed consent under research protocols approved by the research ethics review board of the Institut de recherches cliniques de Montreal.
HeLa and HEK293T cells were cultured as previously described [54]. Primary CD4+ T-lymphocytes were isolated and cultured as previously described [26]. The development of the HEK293T cell line stably depleted of VPRBP was described previously [26]. Caffeine and DAPI (4′,6-Diamidino-2-phenylindole) were purchased from Sigma-Aldrich (St. Louis, MO, USA). SiRNA targeting VPRBP (siGenome SMARTpool M-021119-00) and scrambled control siRNA (non-targeting siRNA #2) were obtained from Dharmacon (Chicago, IL, USA). The anti-HA (clone 12CA5) and anti-p24 (catalog no. HB9725) monoclonal antibodies were produced from hybridomas obtained from the American Type Culture Collection (Manassas, VA, USA). The monoclonal antibody against Vpr (clone 8D1) was a kind gift of Dr Y. Ishizaka (International Medical Center of Japan, Tokyo, Japan) [55]. The following commercially available antibodies were used: mouse anti-nucleoporin (Abcam, Cambridge, MA, USA), mouse anti-RPA70 (Abcam), rabbit anti-53BP1 (Abcam), rabbit anti-GAPDH (Cell Signaling Technology, Danvers, MA, USA), rabbit anti-H3 antibodies (Abcam) rabbit anti-phospho RPA32 (S4/S8) (Bethyl Laboratories, Montgomery, TX, USA), rabbit anti-VPRBP (Accurate Chemical and Scientific Corporation, Westbury, NY, USA), rabbit anti-actin (Sigma-Aldrich, St. Louis, MO, USA), mouse anti-phosphoS139-H2AX (clone JBW301)(Upstate, Millipore, Billerica, MA, USA), mouse FITC-conjugated anti-p24 (clone KC57, Beckman Coulter Canada, Mississauga, Ontario, Canada), mouse anti-SC35 (Sigma-Aldrich), and mouse anti-PML (Santa Cruz Biotechnology, Santa Cruz, CA, USA). All fluorochrome-conjugated secondary antibodies were obtained from Molecular Probes (Invitrogen, San Diego, CA, USA).
SVCMV-Vpr (WT), SVCMV-Vpr (L23F), SVCMV-HA-Vpr (WT), SVCMV-HA-Vpr (V57L), SVCMV-HA-Vpr (R62P), SVCMV-HA-Vpr (Q65R), SVCMV-HA-Vpr (H71R), SVCMV-HA-Vpr (R80A), SVCMV-HA-Vpr (S79A), SVCMV-HA-Vpr (1–86), SVCMV-HA-Vpr (1–78), and SVCMV-VSV-G were previously described or were constructed by PCR as previously described [19], [32], [46]. Plasmids pCDNA3.1_eYFP-MCS(MB) and pCDNA3.1_Rluc-MCS(MB) for the expression of eYFP and renilla luciferease (Rluc) N-terminal fusion proteins were kind gifts of M. Baril and D. Lamarre [56]. Wild type Vpr was amplified by PCR from SVCMV-HA-Vpr (WT) and subcloned into pCDNA3.1_eYFP-MCS(MB) and pCDNA3.1_Rluc-MCS(MB) to generate respectively pCDNA3.1-eYFP-Vpr(WT) and pCDNA3.1-Rluc-Vpr (WT). Vpr (R80A) and Vpr (Q65R) were subcloned into pCDNA3.1_Rluc-MCS(MB) to generate pCDNA3.1-Rluc-Vpr (R80A) and pCDNA3.1-Rluc-Vpr (Q65R) using the same strategy. The lentiviral vector pWPI as well as the packaging plasmid psPAX2 expressing Gag-Pol, Tat and Rev were obtained from Dr. D. Trono (School of Life Sciences, Swiss Institute of Technology, Lausanne, Switzerland). The lentiviral vector pWPI-HA-Vpr (WT) transducing HA-tagged Vpr and GFP was generated from the parental vector pWPI using a strategy described previously [19]. The plasmids expressing sooty mangabey HA-tagged Vpr and Vpx were obtained from S. Benichou (Institut Cochin, Paris, France) [4]. The infectious molecular clones HxBru (Vpr-), HxBru (HA-Vpr), and HxBru Vpr L23F, were described previously [26], [32], [57]. The HxBru VprWT LF/PS molecular clone with mutations (L44P, F45S) in the p6 domain of Gag disrupting interaction with Vpr was described previously [58].
The production and titration of VSV-G-pseudotyped HIV particles and lentiviral vectors were performed as described previously [19], [46].
HeLa cells were transfected using the Lipofectamine 2000 reagent (Invitrogen Canada, Burlington, Ontario, Canada) according to the manufacturer's instructions. HEK293T cells were transfected by a standard calcium phosphate precipitation protocol. SiRNA were transfected using Lipofectamine RNAi Max (Invitrogen Canada, Burlington, Ontario, Canada), according to the manufacturer's instructions. HeLa cells were transduced with the lentiviral vectors WPI and WPI-HA-Vpr in presence of 8µg/ml polybrene at a multiplicity of infection of 0.5 to 2.5, as indicated for each experiment. Primary CD4+ T-lymphocytes were transduced by spinoculation at a multiplicity of infection of 1. Briefly, cells were mixed with lentiviral vector particles in presence of 8µg/ml polybrene and centrifuged for 2 hours at 1200g. HeLa cells were infected, in presence of 8 µg/ml polybrene, with VSV-G-pseudotyped HIV-1 viruses at a concentration of 100 cpm/cell or at a MOI of 1.0, as indicated for each experiment.
Fifty thousand HeLa cells were seeded on cover slips in 24-well plates. Cells were transfected, transduced, or infected as indicated for each experiment. Two days later, cells were processed for fluorescence immunohistochemistry and laser-scanning confocal microscopy as previously described [59]. For analysis of CD4+ primary T-lymphocytes, 5×105 cells were first adhered on poly-Lysine-treated coverslips for two hours in PBS and then processed as described [59]. Quantification of Vpr nuclear foci was performed in at least 30 randomly selected cells by manual counting. Time-lapse confocal microscopy was performed on living cells in a PeCON environmental chamber maintained at 37°C and 5% CO2. Images were acquired using a Zeiss LSM 710 system with the ZEN 2009 software. Spinning-disk confocal microscopy was performed on living cells using a Quorum WaveFX-X1 spinning-disc confocal system (Quorum Technologies Inc, Guelph, Ontario, Canada). Cells were maintained at 37°C in 5% CO2 in a Live Cell Instruments Chamlide TC environmental chamber. Images were acquired with a Hamamatsu ImagEM C9100-13 camera using the Metamorph software. FRAP (fluorescence recovery after photobleaching) experiments were conducted using the Quorum WaveFX-X1 spinning-disc confocal system equipped with a Photonic Instruments Mosaic 405 nm laser. Images were processed using AxioVision v.4.7. Videos were generated with the ZEN 2009 software. Software-assisted fluorescence quantification and tracking of Vpr foci was performed with the Volocity software v.5.2.1. Statistical analysis was performed using Sigma Plot software v.10.
In situ proximity ligation assays (PLA) were performed using the Duolink kit 613 (Olink bioscience, Uppsala, Sweden). Briefly, HeLa cells were transfected with a plasmid encoding HA-Vpr or an empty plasmid as negative control. At 48h post-transfection, the cells were cytospun for 7 min at 1,100 rpm onto a glass slides and were fixed and permeabilized as described above. The fixed cells were incubated with the following antibodies: mouse monoclonal antibody against HA (clone 12CA5) or Vpr (a gift from Dr Y. Ishizaka. The antibody was shown to recognize both Vpr WT and Q65R [12]) and a rabbit polyclonal antibody against VPRBP (Accurate Chemical and Scientific Corporation). The Duolink system provides oligonucleotide-labeled secondary antibodies (PLA probes) to each of the primary antibodies that, in combination with a DNA amplification-based reporter system, generate a signal only when the two primary antibodies are in close proximity. The signal from each detected pair of primary antibodies was visualized as a spot (please see the manufacturer's instructions for more details). Nuclei were delineated using Hoechst 33342.
Cell cycle analysis was performed using propidium iodide staining and flow cytometry as previously described [12], [19].
Immunoprecipitations using anti-HA-conjugated agarose beads were performed as previously described [26]. Analysis of proteins by western blot was performed as previously described [26].
HEK293T cells were transfected in 24-well plates with 10ng of the BRET donor plasmids pCDNA3.1_Rluc-MCS(MB), pCDNA3.1-Rluc-Vpr (WT), pCDNA3.1-Rluc-Vpr (R80A) or pCDNA3.1-Rluc-Vpr (Q65R) and increasing concentration (0 to 500 ng) of the BRET acceptor plasmids pCDNA3.1_eYFP-MCS(MB) or pCDNA3.1-eYFP-Vpr (WT) using Lipofectamine 2000. Two days after transfection, cells were harvested, washed twice in PBS, and aliquoted in two wells of a 96-well plate (Costar 3917). Total eYFP fluorescence was measured with an excitation wavelength of 485 nm and an emission wavelength at 520±10 nm. BRET was initiated by adding 5µM of the renilla luciferase substrate coelenterazine H (Prolume Ltd., Lakeside, AZ, USA). Luminescence was then measured 10 minutes later at 475±15 nm and BRET fluorescence was measured at 535±15 nm. All measurements were performed on a PheraStar microplate reader (BMG Labtech, Cary, NC, USA). BRET ratios were calculated using this formula: (emission at 535 nm/emission at 475 nm)-(background emission at 535nm/background emission at 475 nm), as previously described [60].
Cells were lysed in triton lysis buffer (50 mM Tris pH 7.5, 150 mM NaCl, 0.5% Triton X-100, and complete protease inhibitors cocktail (Roche) for 15 minutes. Insoluble cell debris, including chromatin, was pelleted by centrifugation (2500g for 10 minutes). The supernatant was harvested and represented the soluble input control. Pellets were washed once with nuclease buffer (50 mM Tris pH 8.0, 5 mM CaCl2, and 100 µg/ml BSA), split in two, and resuspended in nuclease buffer alone or nuclease buffer containing 200 U/ml microccocal nuclease (New England Biolabs, Ipswich, MA, USA). Pellets were incubated for 30 minutes on ice and then centrifuged at 12000g for 10 minutes. The supernatant was harvested and represented the chromatin-bound fraction. The corresponding supernatant obtained in absence of nuclease was used to control for non-specific release. For immunoprecipitation experiments, soluble and nuclease-treated fractions were incubated with 25 µl of anti-HA-conjugated agarose beads (Sigma-Aldrich) for 2h at 4C. In some experiments, immunoprecipitations were supplemented with 25 µg/ml ethidium bromide to displace proteins from DNA [47].
|
10.1371/journal.ppat.1002452 | Inflammasome-dependent Pyroptosis and IL-18 Protect against Burkholderia pseudomallei Lung Infection while IL-1β Is Deleterious | Burkholderia pseudomallei is a Gram-negative bacterium that infects macrophages and other cell types and causes melioidosis. The interaction of B. pseudomallei with the inflammasome and the role of pyroptosis, IL-1β, and IL-18 during melioidosis have not been investigated in detail. Here we show that the Nod-like receptors (NLR) NLRP3 and NLRC4 differentially regulate pyroptosis and production of IL-1β and IL-18 and are critical for inflammasome-mediated resistance to melioidosis. In vitro production of IL-1β by macrophages or dendritic cells infected with B. pseudomallei was dependent on NLRC4 and NLRP3 while pyroptosis required only NLRC4. Mice deficient in the inflammasome components ASC, caspase-1, NLRC4, and NLRP3, were dramatically more susceptible to lung infection with B. pseudomallei than WT mice. The heightened susceptibility of Nlrp3-/- mice was due to decreased production of IL-18 and IL-1β. In contrast, Nlrc4-/- mice produced IL-1β and IL-18 in higher amount than WT mice and their high susceptibility was due to decreased pyroptosis and consequently higher bacterial burdens. Analyses of IL-18-deficient mice revealed that IL-18 is essential for survival primarily because of its ability to induce IFNγ production. In contrast, studies using IL-1RI-deficient mice or WT mice treated with either IL-1β or IL-1 receptor agonist revealed that IL-1β has deleterious effects during melioidosis. The detrimental role of IL-1β appeared to be due, in part, to excessive recruitment of neutrophils to the lung. Because neutrophils do not express NLRC4 and therefore fail to undergo pyroptosis, they may be permissive to B. pseudomallei intracellular growth. Administration of neutrophil-recruitment inhibitors IL-1ra or the CXCR2 neutrophil chemokine receptor antagonist antileukinate protected Nlrc4-/- mice from lethal doses of B. pseudomallei and decreased systemic dissemination of bacteria. Thus, the NLRP3 and NLRC4 inflammasomes have non-redundant protective roles in melioidosis: NLRC4 regulates pyroptosis while NLRP3 regulates production of protective IL-18 and deleterious IL-1β.
| The disease melioidosis is caused by the intracellular bacterium Burkholderia pseudomallei, a potential bioterrorism agent. Here we examined the interaction of B. pseudomallei with the inflammasome, an important innate immune pathway that regulates at least two host responses protective against infections: 1) secretion of the proinflammatory cytokines IL-1β and IL-18 and 2) induction of pyroptosis, a form of cell death that restricts intracellular bacteria growth. Using a mouse model of melioidosis we show that two distinct inflammasomes are activated by B. pseudomallei infection. One, containing the Nod-like receptor (NLR) NLRP3, mediates IL-1β and IL-18 induction. The other contains a different NLR called NLRC4 and mediates pyroptosis. Pyroptosis and IL-18 production were equally important for resistance to B. pseudomallei. Surprisingly, IL-1β was found to be deleterious in melioidosis. The detrimental role of IL-1β during melioidosis was due, in part, to excessive recruitment of neutrophils to the lung. We show that neutrophils do not express NLRC4, fail to undergo pyroptosis, and, therefore, may be permissive to B. pseudomallei intracellular replication leading to increased bacterial burden and morbidity/mortality. Thus, the NLRP3 and NLRC4 inflammasomes have non-redundant protective roles in melioidosis: NLRC4 regulates pyroptosis while NLRP3 regulates production of protective IL-18 and deleterious IL-1β.
| The ability to detect infection by pathogenic microbes and to restrict their growth are fundamental for the wellbeing of multicellular organisms. Pattern recognition receptors, including the Toll-like receptor (TLR) and the NLR, recognize microbial products and “danger signals” released by stressed cells and, in turn, activate signaling pathways that initiate the inflammatory response and regulate development of adaptive immunity. TLR are expressed on the cell surface or in endosomal compartments and their stimulation results in activation of the NF-κB, MAPK, and IRF signaling pathways culminating in transcriptional induction of a large number of genes. NLR, in contrast, are located in the cytoplasm, which they survey for evidence of danger or infection (reviewed in ref. [1]). Some NLR control activation of the inflammasome, a multiprotein complex that contains, in addition to a NLR, the adaptor molecule ASC and the protease caspase-1. Activation of caspase-1 in the context of the inflammasome is responsible for the proteolytic processing of the immature forms of IL-1β and IL-18, a modification required for the secretion and bio-activity of these proinflammatory cytokines. Activation of caspase-1 also triggers a form of cell death, known as pyroptosis, that effectively restricts intracellular bacterial growth [2], [3]. Production of IL-1β and IL-18 and induction of pyroptosis have been shown to be protective effector mechanisms against many infectious agents. NLRP3 and NLRC4 are the best characterized NLR molecules. NLRP3 controls caspase-1 activation in response to “danger signals”, several particles and crystals, and various bacteria, virus, and fungi. Although the logic that oversees the activation of the NLRP3 inflammasome is still elusive, it appears that disruption of cell membrane integrity may be a common event triggered by the NLRP3 activators. The NLRC4 inflammasome is responsive to a narrower spectrum of activators including cytoplasmically delivered bacterial flagellin and the basal rod constituent of various bacterial Type III secretion systems (T3SS). The T3SS apparatus is used by several bacteria, including Salmonella, Yersinia, Pseudomonas, Shigella, Legionella, and Burkholderia to inject virulence factors into the cytoplasm of target cells. Recent works demonstrated that the specificity of the mouse NLRC4 for flagellin or rod proteins is determined by its interaction with the NLR molecules NAIP5 or NAIP2, respectively [4], [5].
Burkholderia pseudomallei is a Gram-negative flagellate bacterium that causes melioidosis, a disease endemic to South-East Asia and other tropical regions [6], [7] and the most common cause of pneumonia-derived sepsis in Thailand. Because melioidosis carries a high fatality rate, B. pseudomallei is classified as category B potential bioterrorism agent by the Center for Disease Control and NIAID. B. pseudomallei infection can be contracted through ingestion, inhalation, or subcutaneous inoculation and leads to broad-spectrum disease forms including pneumonia, septicemia, and organ abscesses. Following infection of macrophages and other non-phagocytic cell types, B. pseudomallei is able to escape the phagosome and invade and replicate in the host cell cytoplasm, directly spreading from cell to cell using actin-tail propulsion. Macrophages and IFNγ have been shown to play a critical role in protection from melioidosis [8]–[10] and several B. pseudomallei virulence factors have been identified including the bacterial capsule [11], the lipopolysaccharide [12], and one of the three T3SS possessed by B. pseudomallei [13]. Analysis of mouse strains with different susceptibility to B. pseudomallei infection indicates that the early phases of the infection are crucial for survival [14], [15], emphasizing the necessity for better understanding of innate immune responses during melioidosis. With this goal in mind, using a murine model of melioidosis we have performed a detailed analysis of the role of the inflammasome components NLRP3, NLRC4, ASC, and caspase-1 and the effector mechanisms IL-1β, IL-18, and pyroptosis.
To identify the pathway responsible for IL-1β and IL-18 secretion in response to infection with B. pseudomallei, bone marrow-derived macrophages (BMDM) or dendritic cells (BMDC) derived from WT mice or mice deficient in the inflammasome components ASC, NLRP3, NLRC4, or caspase-1 were infected in vitro with B. pseudomallei and secretion of IL-1β in culture supernatants was measured. As shown in figure 1A, secretion of IL-1β by Asc-/-, Nlrp3-/-, and Casp1-/- BMDM was markedly reduced compared to WT BMDM. Production of IL-1β during the first hours of the infection was also significantly reduced in Nlrc4-/- cells. However, later in the infection process (8 hours) Nlrc4-/- cells secreted IL-1β at levels considerably higher than WT cells. Secretion of IL-18 followed a similar pattern (data not shown). Immunoblotting of the supernatants confirmed processing of IL-1β and of caspase-1 to the mature 17 kDa and p20 forms, respectively (figure 1B). Interestingly, although caspase-1 was activated in Asc-/- cells, processing and secretion of IL-1β was not observed. NLRC4 possesses an amino-terminal CARD domain that can recruit and activate caspase-1 independently of ASC. It is unclear at present why activation of caspase-1 in Asc-/- cells is not sufficient to trigger secretion of mature IL-1β, a phenomenon previously reported by other groups [16]. The differences in IL-1β and IL-18 secretion were observed regardless of the number of bacteria used to infect cells (MOI 10, 50, or 100, data not shown) and were not due to differential induction of pro-IL-1β, which was present at comparable amounts in all the cell lysates. Thus, the NLRC4 and NLRP3 inflammasomes are both mediating release of IL-1β and IL-18 by myeloid cells infected with B. pseudomallei.
Inflammasome-mediated induction of pyroptosis has been demonstrated to be a mechanism that restricts growth of certain intracellular bacteria [2], [3]. To measure induction of pyroptosis in cells infected with B. pseudomallei we used a kanamycin protection assay that allows only replication of intracellular bacteria whereas cells that undergo pyroptosis expose the bacteria to the microbicidal action of the antibiotic present in the medium. Induction of pyroptosis and intracellular bacterial replication were measured in WT or inflammasome-deficient BMDM infected with B. pseudomallei. As shown in figure 1C (upper graph), pyroptosis of infected cells (as measured by release of LDH in culture supernatants) was significantly reduced in Casp1-/- and Nlrc4-/- cells compared to WT and Nlrp3-/-. Importantly, induction of pyroptosis was not lost in Asc-/- cells. NLRC4-mediated pyroptosis induced by other bacteria is also reported to be ASC-independent [16]–[19]. Consistent with the role of pyroptosis as a mechanism to restrict intracellular bacteria growth, considerably less intracellular bacteria were recovered from WT, Nlrp3-/-, and Asc-/- cells than Casp1-/- or Nlrc4-/- cells at all time points (figure 1C, lower graph). Similar results regarding IL-1β processing and secretion and induction of pyroptosis were obtained using BMDC derived from the inflammasome–deficient mice (supplementary figure S1).
Taken together these results show that infection of macrophages and dendritic cells with B. pseudomallei leads to activation of the NLRC4 and NLRP3 inflammasomes. NLRC4 contributes to IL-1β during the early phase of the infection and induction of pyroptosis that restricts bacterial growth. NLRP3 does not control pyroptosis and primarily controls IL-1β secretion. It should be noted that the defective IL-1β production of Nlrc4-/- and Nlrp3-/- cells cannot be ascribed to the difference in induction of pyroptosis: thus Nlrp3-/- cells produce less cytokine than WT cells despite undergoing pyroptosis to the same extent as WT cells. Conversely, Nlrc4-/- cells, which are resistant to pyroptosis, still produce less cytokine than WT cells at the early time point. However, at later time points Nlrc4-/- cells produce considerably more IL-1β than WT cells. This is likely due to the fact that WT cells rapidly die after infection while Nlrc4-/- cells remain viable and continue to synthesize and secrete IL-1β.
The role of the inflammasome during in vivo B. pseudomallei infection was next analyzed using a mouse model of melioidosis (figure 2). WT mice or inflammasome-deficient mice were infected intranasally with B. pseudomallei (100 CFU) and their weight (not shown) and survival were monitored (figure 2A). All mice started to lose weight 2 days post-infection. Generally, mice that survived the infection started to recover weight 7 days post-infection. Casp1-/-, Nlrc4-/-, and Asc-/- mice were extremely susceptible to melioidosis compared to WT mice. Nlrp3-/- mice were also considerably more susceptible than WT mice but slightly more resistant than the other inflammasome deficient mice. Measurement of the bacterial burdens in lungs, spleens, and livers of infected mice 24 hours (data not shown) and 48 hours post-infection revealed that Nlrc4-/- and Casp1-/- mice carried considerably higher burdens in all three organs than WT mice (figure 2B). Surprisingly, the bacterial burden of Asc-/- and Nlrp3-/- mice was not significantly different from that of WT mice at the tested time points despite their higher mortality.
Cytokine levels were measured in bronchio-alveolar lavage fluids (BALF) obtained from infected mice (figure 2C). Confirming the in vitro results, IL-1β and IL-18 levels were severely reduced in Asc-/-, Casp1-/- and Nlrp3-/- mice. In contrast, IL-1β and IL-18 were present in the lungs of Nlrc4-/- mice in amounts considerably higher than WT mice. Immunoblotting experiments confirmed that the IL-1β measured by ELISA was in fact the p17 mature form of IL-1β (figure 2D). Thus, although the in vitro experiments demonstrated that both the NLRP3 and the NLRC4 inflammasome contribute to IL-1β and IL-18 production in response to B. pseudomallei infection, it is the NLRP3 inflammasome that primarily mediates production of these cytokines in vivo. The levels of several other proinflammatory cytokines, including IL-1α (figure S2), were significantly elevated in Nlrc4-/- BALF. It is interesting to note that the levels of IL-18 in BALF of Asc-/- and Casp1-/- mice, although very low, were higher than uninfected mice suggesting the existence of inflammasome-independent mechanisms to produce IL-1β and IL-18, as it has been previously shown in models of highly neutrophilic inflammation [20]-[23].
Histological analysis of the infected lungs revealed extensive inflammatory cell infiltration in the lung parenchyma (data not shown). The area of the inflammatory nodules, relative to the total area of the lung lobe, was calculated for each given section and found to be significantly greater in Nlrc4-/- mice compared to WT mice (figure 2E). This result was consistent with the elevated levels of inflammatory cytokines and chemokines produced by Nlrc4-/- mice. Taken together these results suggest a scenario where failure of Nlrc4-/- infected macrophages to undergo pyroptosis results in higher bacterial burden and continued production of IL-1β and other factors that attract more inflammatory cells, perpetuating lung inflammation and promoting bacteria dissemination.
Thus, our results identified two distinct infammasome-mediated mechanisms that efficiently restrict B. pseudomallei growth and pathogenesis: production of the cytokines IL-1β and IL-18 and induction of pyroptosis. The high susceptibility of Nlrp3-/- and Asc-/- mice to meliodiosis is due to defective cytokine production while that of the Nlrc4-/- mice likely results from defective pyroptosis. Casp1-/- mice are impaired in both inflammasome effector mechanisms and, therefore, we predicted that they would be more vulnerable to B. pseudomallei than Asc-/- or Il1-r1-/--Il-18-/- double knock-out mice (DKO) (that are defective in cytokines but retain pyroptosis) or Nlrc4-/- mice (that retain IL-1β/IL-18 functionality but are deficient in pyroptosis). This prediction was found to be correct. As shown in figure 2F, when mice were infected with only 25 CFU (a non-lethal dose for WT mice) the mean time to death of Nlrc4-/- and Il-1r1-/--Il-18-/- DKO mice was slightly but significantly (p<0.05, Kaplan-Meier test) increased compared to Casp1-/- mice. Surprisingly, Asc-/- mice, which should be equivalent to DKO because of the absence of IL-1β or IL-18, survived the infection. This may be explained by the observation that IL-18, although drastically reduced, it is still detectable in Asc-/- mice at higher level than uninfected mice (figure 2C).
We next analyzed the role of the inflammasome-dependent cytokines IL-1β and IL-18 during murine melioidosis. IL-18-deficient mice were extremely susceptible to B. pseudomallei infection even when infected with 25 CFU, a dose of bacteria that caused no mortality and only mild weight loss in WT mice (figure 3A). In contrast, Il-1r1-/- mice displayed increased resistance to B. pseudomallei infection compared to WT mice (figure 3A and see below). The survival of mice deficient in both IL-18 and IL-1RI (DKO) was indistinguishable from the Il-18-/- mice when the animals were infected with 100 CFU. However, in DKO mice infected with 25 CFU (figure 3A, right panel) the concomitant absence of IL-18 and IL-1RI provided a significant advantage over Il-18-/- mice (p<0.05) suggesting a detrimental role of IL-1RI-mediated signaling in melioidosis (see below).
Confirming the different susceptibilities of Il-18-/- and Il-1r1-/- mice to melioidosis, the bacterial burdens observed in the lungs, spleens, livers, and BALF of infected Il-18-/- mice were dramatically higher than that of WT mice even at early time points (24 hours post infection, figure 3B). In contrast, significantly lower amounts of bacteria were recovered 48 hours post infection from Il-1r1-/- mice compared to WT mice confirming their higher resistance.
Measurements of cytokines in the BALF obtained from mice at 24 and 48 hours post-infection (figure 3C) indicated that the levels of IFNγ were drastically reduced in Il-18-/- mice, a finding consistent with the established function of IL-18 as an IFNγ-inducing cytokine. Remarkably, IFNγ levels in Il-1r1-/- mice were greatly increased compared to WT mice. The levels of the neutrophil attractants Mip-2, KC, and IL-17 were also decreased in Il-1r1-/- mice and increased in Il-18-/- mice (figure S2). The number of inflammatory cells recovered from the BALF of infected Il-1r1-/- mice was significantly decreased compared to WT mice (see figure 4B, left panel). Histological analysis of the infected lungs revealed extensive inflammatory cell infiltration in the lung parenchyma of Il-18-/- mice (see figure 4C). The area of the inflammatory nodules, relative to the total area of the lung lobe, was calculated for each given section and found to be significantly greater in Il-18-/- mice compared to WT mice (figure 4D).
Considering that IFNγ is known to play a protective role during several bacterial infections, including B. pseudomallei [8]–[10], these results suggested that the reduced resistance of Il-18-/- mice to B. pseudomallei infection may be due to lack of IFNγ induction. To test this hypothesis, a group of Il-18-/- mice infected with B. pseudomallei were given daily intraperitoneal injections of either recombinant IFNγ or PBS. As shown in figure 3D, exogenous IFNγ completely protected the mice suggesting that IL-18 exerts its protective action primarily through induction of IFNγ.
The results of figure 3 showed that Il-1r1-/- mice were more resistant to lung infection with B. pseudomallei. This appeared even more evident when mice were infected with higher doses of B. pseudomallei that killed all WT mice but only a fraction of the Il-1r1-/- mice (figure 4A). Recruitment of neutrophils, macrophages, and dendritic cells into alveolar spaces was decreased in Il-1r1-/- mice compared to WT mice (figure 4B, left graph). Lower levels of the neutrophil enzyme myeloperoxidase (MPO) were detected in the BALF of Il-1r1-/- mice compared to WT (figure S3). The extent of lung inflammation, as measured by the number and size of inflammatory nodules, was also significantly decreased in Il-1r1-/- mice (figure 4C, and 4D).
To further test the hypothesis that IL-1R-mediated signaling has a deleterious role in this model of melioidosis, WT mice were infected with 100 CFU B. pseudomallei and were given daily intraperitoneal injections of IL-1β or PBS (figure 4E). All mice that received the cytokine succumbed to the infection compared to significantly higher survival of the control group. Injection of IL-1β in non-infected mice had no deleterious effect aside from a transient, negligible weight loss (not shown). The bacteria burdens in organs of IL-1β-treated mice 72 hours post infection were dramatically higher than the control group and bacteremia was detected in IL-1β-treated mice but not control mice (figure 4F). Higher number of neutrophils, macrophages, and dendritic cells were found in the BALF of IL-1β-treated mice (figure 4B, center graph). This correlated with increased level of MPO in BALF (figure S3). The increased inflammatory cell recruitment to the lungs of IL-1β-treated mice was likely due to the induction, by IL-1β, of neutrophil chemoattractans KC (CXCL1) and MIP-2 (CXCL-2), which in fact were detected at very high levels in the BALF of IL-1β-treated mice (figure S2). Histological analysis of lung sections of mice treated with IL-1β showed a dramatic increase in the number and size of the foci of infiltrating inflammatory cells (figure 4C, lower left panels) and evidence of perivascular edema and airway obstruction (figure 4C, lower right panels).
If IL-1β in fact has a detrimental effect during melioidosis, inhibition of its activity should lower morbidity and mortality of mice infected with B. pseudomallei. As shown in figure 4G, administration of the IL-1 receptor antagonist IL-1ra protected mice from infection with lethal doses of B. pseudomallei. Mice treated with IL-1ra had decreased recruitment of inflammatory cells to the alveolar spaces (figure 4B, right graph) lower level of MPO in BALF (figure S3), and less severe lung pathology (data not shown).
Surprisingly, in our experiments lower numbers of neutrophils in Il-1r1-/- mice correlated with lower bacterial burdens while IL-1β administration resulted in increased neutrophil recruitment but also increased bacterial burdens and systemic dissemination. These results would be consistent with the notion that neutrophils are not very effective at containing B. pseudomallei infection and, in fact, may foster its spread. In support of this idea, human neutrophils infected with B. pseudomallei underwent pyroptosis at a much slower rate than infected monocytes (figure 5A). Concomitantly, intracellular bacteria growth increased with time in infected neutrophils but decreased in monocytes. Consistent with previously published results [3], neutrophils did not express NLRC4 mRNA (figure 5B) suggesting they may be resistant to pyroptosis. Similar results were obtained using neutrophils and CD11b+ monocytic cells isolated from mouse bone marrow (figure 5C). WT monocytes infected with B. pseudomallei underwent pyroptosis and failed to support bacteria replication whereas Nlrc4-/- cells were resistant to pyropotosis and supported B. pseudomallei intracellular replication. In contrast, both WT and Nlrc4-/- neutrophils did not undergo pyroptosis and supported B. pseudomallei intracellular replication to the same extent. These results suggest that the deleterious role of IL-1β during melioidosis may be due, in part, to excessive recruitment of neutrophils, a cell type that may be permissive for B. pseudomallei replication. We decided to test this hypothesis in Nlrc4-/- mice. As shown in figure 2E, infected Nlrc4-/- mice showed a significantly higher degree of lung inflammation. Consistent with higher neutrophil influx in the lung of Nlrc4-/- mice, the levels of the neutrophil enzyme MPO were significantly increased in their BALF compared to WT mice (figure 6A). To test the hypothesis that excessive neutrophil influx is deleterious during melioidosis, Nlrc4-/- mice were injected with IL-1ra or with antileukinate, a hexapeptide that acts as a CXCR2 neutrophil chemokine receptor antagonist. Both factors have been shown to inhibit neutrophil recruitment to inflammatory sites in different animal models including lung inflammation [24]–[26]. As shown in figure 6B, administration of IL-1ra or antileukinate protected Nlrc4-/- mice infected with low doses of B. pseudomallei. The number of inflammatory cells in the BALF of Nlrc4-/- mice treated with IL-1ra or antileukinate was reduced compared to mice who received PBS injection (figure 6C) and lower levels of MPO were detected in the BALF of injected mice (figure S3). Moreover, systemic spread of bacteria to spleen or liver was reduced by administration of either drug (figure 6D).
The inflammatory response to infection consists of several protective effector mechanisms that must be activated and orchestrated in order to maximize microbicidal functions and stimulation of adaptive immunity while, at the same time, minimize damage to the host tissues. Alteration in this balance may result in excessive and non-resolving inflammation that leads to severe morbidity and mortality [27]. It is becoming clear that to be effective but non-pathogenic the inflammatory response must be tailored to each specific pathogen. Here we have analyzed the role of a very important inflammatory pathway during infection with the lung pathogen B. pseudomallei. Using a murine model of melioidosis we have determined the role of various components of the inflammasome and the downstream effector mechanisms (production of IL-1β, IL-18, and pyroptosis) and we report several novel discoveries that greatly increase our understanding of the pathogenesis of melioidosis (see model in figure 7).
First, we found that both NLRC4 and NLRP3 play non-redundant roles during detection of B. pseudomallei. Analysis of in vitro infected macrophages or dendritic cells allowed us to estimate the relative contribution of NLRC4 and NLRP3 to IL-1β production. Our findings indicated that production of IL-1β is primarily dependent on the NLRP3 inflammasome. During the early phase of the infection the NLRC4 inflammasome also significantly contributes to IL-1β production. We posit that this pattern likely reflects the fact that the NLRC4 inflammasome responds to T3SS deployment, which occurs early in the infection cycle, while activation of NLRP3 may require escape from the phagosome, which is a relatively slower event [28]. B. pseudomallei , including the strain used in our study, possesses at least three T3SS gene clusters, one of which is similar to the Salmonella SP-1 pathogenicity island and has been shown to be an important virulence factor required for escape from the phagosome, induction of IL-1β production, and pathogenicity [13], [28]. In addition to mediating host recognition of cytosol-delivered flagellin, NLRC4 also recognizes a structural motif found in the basal body rod components of the T3SS of various bacteria, including B. pseudomallei [29]. We have determined (data not shown) that transfection of B. pseudomallei flagellin protein into the cytoplasm of BMDC leads to NLRC4-dependent production of IL-1β. This result agrees with previously published evidence and indicates that B. pseudomallei (like some other bacteria) expresses multiple factors (e.g. flagellin, basal rods) that are recognized by the NLRC4 inflammasome. The mechanism responsible for NLRP3 activation by B. pseudomallei remains unclear.
In addition to controlling IL-1β and IL-18 production, NLRC4 also mediates pyroptosis, a form of cell death that is an effective mechanism to restrict growth and dissemination of intracellular bacteria [3]. Here we showed that B. pseudomallei-induced pyroptosis was caspase-1-dependent but ASC-independent, in agreement with works that showed ASC redundancy for pyroptosis induced by other bacteria [16]–[19]. However, the fact that production of IL-1β in response to B. pseudomallei infection is seriously compromised in Asc-/- cells indicates that this adaptor molecule plays a critical role in NLRC4-mediated cytokine production and suggests that NLRC4 can form two distinct inflammasomes: one that contains ASC and regulates IL-1β processing, and one devoid of ASC that activates caspase-1 and triggers pyroptosis, as recently proposed [30]. It has been recently shown [4], [5] that NAIP molecules determine the specificity of NLRC4 for its activators and, we would further speculate, for its down-stream effector mechanisms. Whether NLRC4 relies on other molecules to recognize B. pseudomallei remains to be ascertained. We tested the susceptibility to B. pseudomallei of C57BL/6J-Chr13A/J/NaJ mice, a consomic C57BL/6 strain that carries the A/J NAIP5 allele that renders them susceptible to Legionella infection [31]-[34], and found that they were indistinguishable from WT mice (data not shown).
Analysis of inflammasomes-deficient mice intranasaly infected with B. pseudomallei confirmed the importance of ASC, caspase-1, and both NLRP3 and NLRC4 inflammasomes for resistance to melioidosis. However, quite surprisingly, although production of IL-1β and IL-18 in vitro is mediated by both NLRP3 and NLRC4, in vivo it is exclusively dependent on the NLRP3-ASC-caspase-1 inflammasome (figure 2C). In contrast, Nlrc4-/- mice produce these cytokines in amounts that exceed even those detected in WT mice. Remarkably, despite the abundance of IL-1β and IL-18, Nlrc4-/- mice were dramatically more susceptible to melioidosis than WT mice, rapidly succumbed to the infection, and had very high organ's bacteria burden and worst neutrophilic lung inflammation. Thus, the critical role of NLRC4 during melioidosis is independent of IL-1β and IL-18 production. Rather, our results suggest that pyroptosis, which we show is defective in Nlrc4-/- cells, is a critical NLRC4 effector mechanism to fight B. pseudomallei and, in its absence, bacterial replication and IL-1β production proceeds unrestrained causing severe inflammation, morbidity and mortality. Moreover, our analysis indicates that pyroptosis and IL-18 are both required and contribute equally to resistance to melioidosis. Thus, deficiency of either is equally lethal while deficiency of both (Casp1-/- mice) further worsens the outcome. It is important to emphasize that our study is the first to demonstrate the importance of pyroptosis in the context of an infection with a clinically relevant human pathogen that has not been genetically manipulated, as opposed to the previous seminal work by Miao et al. [3] that elegantly employed genetically manipulated bacteria and mouse strains to identify pyroptosis as an effective innate immune defence mechanism against bacterial infections. Previous reports have demonstrated activation of both NLRP3 and NLRC4 inflammasomes in response to infection with Legionella pneumophila [16], Listeria monocytogenes [35], and Salmonella typhimurium [36]. However, in those infection models NLRP3 and NLRC4 appeared to play redundant roles while in our model we were able to assign distinct functions to each inflammasome.
A great number of publications have documented the role of IL-18 and IL-1β during infections with a variety of pathogens. Almost invariably, both cytokines were found to have a protective function. Remarkably, our results show that while Il-18-/- mice are profoundly vulnerable to melioidosis, as previously shown [37], Il-1r1-/- mice were unexpectedly more resistant than WT mice. The protective role of IL-18 during melioidosis appears to be related to its ability to induce IFNγ, as administration of exogenous IFNγ completely rescued the survival of Il-18-/- infected mice. IFNγ activates the microbicidal activity of macrophages and has been shown to be important for resistance against infection with many pathogens including B. pseudomallei [8]–[10]. It is interesting and surprising to see that Asc-/- and Nlrp3-/- mice, which are defective in IL-1β and IL-18 production, are more resistant to B. pseudomallei than mice lacking IL-18. It is worth noting that although IL-18 production is drastically reduced in Asc-/- and Nlrp3-/- mice, it is still detectable in these mice at higher level than uninfected mice. It is conceivable that this inflammasome-independent production of IL-18 may be sufficient to provide some level of protection to Asc-/- and Nlrp3-/- mice against infection with low B. pseudomallei CFU.
Our discovery that Il-1r1-/- mice were more resistant than WT to B. pseudomallei infection is quite surprising considering that this cytokine has been shown to be protective in several bacterial, viral, and fungal infection models [38]. Studies in humans have also shown that inhibition of the function of IL-1 using the IL-1R antagonist IL-1ra (Kineret) is associated with increased susceptibility to bacterial infection. Infected Il-1r1-/- mice had lower BALF levels of proinflammatory cytokines as well as a reduction of neutrophil influx into the lungs, bacterial burdens, and lung pathology. Consistent with a deleterious role of IL-1β in melioidosis, administration of recombinant IL-1β drastically increased mortality, inflammation, pathology, and bacteria burdens while administration of IL-1ra (Kineret) rescued the survival of WT mice infected with a lethal dose of B. pseudomallei.
The reason for the detrimental effect of IL-1β during melioidosis is unclear and it is likely that several factors determine this outcome. IL-1β is one of the most powerful proinflammatory cytokines, it affects virtually every organ, and several human pathologies are primarily driven by unrestrained IL-1β production. One possible mechanism to account for IL-1β's deleterious role in melioidosis may be related to its ability to inhibit IFNγ production through the induction of the cycloxigenase COX-2 and release of prostaglandin PGE2 [39], [40]. Our observation that the level of IFNγ, a protective factor against B. pseudomallei, was significantly higher in infected Il-1r1-/- than WT mice (figure 3C) supports this type of scenario in melioidosis. Interestingly IL-8, a potent neutrophil chemoattractant, was shown to enhance the intracellular growth and survival of B. cepacia in bronchial epithelial cell lines [41]. Whether IL-1β promotes B. pseudomallei intracellular replication is not known but our preliminary results indicated that induction of pyroptosis by B. pseudomallei was not affected by IL-1β.
IL-1β regulates neutrophil recruitment to inflammatory sites through multiple mechanisms including induction of KC, MIP-2, and IL-17, inflammatory mediators whose expression in our experiments correlated with the presence/absence of IL-1RI-mediated signaling. Excessive PMN recruitment is known to cause tissue damage leading to functional impairment of multiple organs, including the lungs [42], [43]. One of the most remarkable observations reported here is that the absence of IL-1 signaling was associated with reduced lung neutrophilic inflammation but also lower bacterial burdens in the lungs (figure 3B, 4B). Conversely, IL-1β administration resulted in increased neutrophil recruitment but also increased bacterial burdens and systemic dissemination. These results would be consistent with the idea that neutrophils are not very effective at containing B. pseudomallei infection and, in fact, may foster its spread despite their strong microbicidal activities. This notion is supported by our observation that human or mouse neutrophils infected with B. pseudomallei failed to undergo pyropotosis (figure 5), consistent with the finding that neutrophils do not express NLRC4 [3]. At the same time, intracellular B. pseudomallei replication proceeded unaffected in both WT and Nlrc4-/- neutrophils in agreement with a report that showed that B. pseudomallei is intrinsically resistant to killing by infected PMN [44]. In support for a deleterious role of neutrophils in melioidosis we found that inhibition of their recruitment by administration of IL-1ra or the CXCR2 neutrophil chemokine receptor antagonist antileukinate protected Nlrc4-/- mice from infection with low doses of B. pseudomallei and decreased systemic spread of bacteria (figure 6).
Taken together our results suggest the following scenario: failure of Nlrc4-/- infected macrophages to undergo pyroptosis results in higher bacteria burden and continued production of IL-1β and other factors that attract more inflammatory cells, including neutrophils, perpetuating excessive lung inflammation and promoting bacteria dissemination. It is tempting to speculate that IL-1β promotes B. pseudomallei growth possibly by increasing the local pool of infectable permissive cells, including PMN. Our conclusion that neutrophils are a permissive cell type for B. pseudomallei replication seems to contrast with a report [45] that indicated that depletion of neutrophils resulted in severe increase in mortality in a model of murine melioidosis. However, caution should be used in the interpretation of these types of experiments because systemic depletion of neutrophils devoids the host not only of their microbicidal function but also of the many immunomodulatory functions these cells exert [46]. Of note, mice deficient in osteopontin, a pleiotropic cytokine that is chemotactic for neutrophils, were shown to be more resistant to B. pseudomallei infection [47], supporting our conclusion that neutrophils have a detrimental role in melioidosis.
The notion that excessive inflammation may be detrimental in certain infection models is well accepted. For example, TLR-mediated signaling negatively affects the outcome of infections with West Nile Virus [48] or influenza virus [49]. The fact that IL-1β is deleterious in melioidosis but protective against other lung pathogens like Klebsiella, Francisella, Mycobacterium, Respiratory Syncytial Virus, and influenza virus likely reflects differences between the virulence strategy of B. pseudomallei and those other pathogens. The intensity, kinetics, and quality of the inflammatory response elicited by B. pseudomallei and its ability to suppress the induction of anti-inflammatory circuitries are phenomena that we are interested to investigate in detail. Despite an extensive literature search we could identify only a single report [26] where IL-1β was shown to be deleterious in bacterial infections. It was demonstrated that this cytokine had a negative effect on bacterial clearance in a model of pneumonia caused by Pseudomonas aeruginosa, an organism that shares features with Burkholderia, which was in fact previously classified in the Pseudomonas genus. Surprisingly, the same group also reported a deleterious role for IL-18 in this type of infection [50], a further indication that each pathogen displays unique virulence strategies. It has been shown that activation of the inflammasome exacerbates inflammation without restricting bacterial growth in a model of Mycobacterial infection [51]. That report did not examine the role of IL-1β but other work showed it is protective during tuberculosis [23].
This is the first report that has analyzed in detail the role of the inflammasome during melioidosis. Previous work has implicated caspase-1 [52] and IL-18 [37] in this infectious disease although the pathways that led to their activation were not investigated. Other species of Burkholderia have been used as model organisms to study aspects of inflammasome biology. Surprisingly, B. thailandensis, which is has been used as a model for melioidosis although it rarely causes disease in humans, was reported to cause similar disease in WT and IL-18- IL-1β-double deficient mice [3] suggesting that species of Burkholderia other than B. pseudomallei may not be reliable models for melioidosis.
In summary, our work shows that NLRP3 and NLRC4 play non-redundant roles during B. pseudomallei infection by differentially regulating pyroptosis and production of IL-1β and IL-18; it demonstrates that pyroptosis is an efficient effector mechanism to restrict in vivo bacterial growth and dissemination; it identifies a deleterious role of IL-1β in melioidosis possibly due to excessive recruitment of neutrophils, a cell type that may be permissive to replication of B. pseudomallei; and, finally, it indicates that inhibition of IL-1RI-mediated signaling may be a beneficial therapeutical approach for the treatment of melioidosis.
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 University of Tennessee Health Science Center (UTHSC) Institutional Animal Care and Use Committee (IACUC, protocol #1854). All efforts were made to minimize suffering and ensure the highest ethical and humane standards. Research involving human blood is exempt from the human subjects regulations. Human neutrophils and monocytes were isolated from healthy donors Leukopacks obtained from Lifeblood Mid-South Regional Blood Center, Memphis TN. All leukopaks are obtained anonymously. The gender, race, and age of each donor are unknown to the investigators.
C57BL/6, Il-1r1-/-, Il-18-/-, C57BL/6J-Chr13A/J/NaJ mice were purchased from Jackson lab. Il-18-/--Il-1r1-/- double deficient mice (DKO) were obtained by crossing the parental single knock-out mice. Asc-/-, Nlrp3-/-, Nlrc4-/- (from Vishva Dixit, Genentec) and Casp1-/- (from Fayyaz Sutterwala) were bred in our facility. All mouse strains were on C57BL/6 genetic background and were bred under specific pathogen-free conditions. Age-(8–12 weeks old) and sex-matched animals were used in all experiments. Generally, experimental groups were composed of at least 5 mice. Animal and in vitro experiments involving B. pseudomallei were performed under biosafety level 3 conditions in accordance with standard operating procedures approved by the Regional Biocontainment Laboratory at UTHSC.
For all experiment the B. pseudomallei 1026b strain (a clinical virulent isolate) was used. Bacteria were grown in Luria broth (Difco) to mid-logarithmic phase, their titer was determined by plating serial dilutions on LB agar, and stocks were maintained frozen at −80°C in 20% glycerol. No loss in viability was observed over prolonged storage. For infections, frozen stocks were diluted in sterile PBS to the desired titer. Aliquots were plated on LB agar to confirm actual CFU. Mice were anesthetized with isoflurane using a Surgivet apparatus and 50 µl of bacteria suspension were applied to the nares. In some experiments, mice were injected i.p. daily with recombinant mouse IL-1β (1 µg) or IFNγ (1 µg). IL-1ra (Biovitrum) was administered by alternating s.c. and i.p. injections every 12 hours (60 mg/kg body weight). Antileukinate (American Peptide Company) was administered by s.c. injection (8 mg/kg body weight).
Mouse macrophages or dendritic cells were generated by incubating bone marrow cells in RPMI 1640-10%FCS supplemented with either rmM-CSF or rmGM-CSF (20 ng/ml) for 8 days, respectively.
Neutrophils and monocytic cells were isolated from the bone marrow cells of WT or Nlrc4-/- mice using Miltenyi Ly6G microbeads. Flow-through cells, consisting mostly of monocytic cells, were further purified using Miltenyi CD11b microbeads.
Human neutrophils and monocytes were isolated from healthy donors Leukopacks obtained from Lifeblood Mid-South Regional Blood Center, Memphis TN. Blood was mixed with Isolymph (CTL Scientific Supply Corp.) (5∶1 ratio) and RBC were allowed to sediment for 60 min at RT. The leukocytes-enriched supernatant was washed, resuspended in PBS, and stratified over Isolymph cushion and centrifuged at 1,350 rpm for 40 min. The cell pellet containing RBC and neutrophils was treated with 0.2% NaCl for 30 seconds to lyse RBC and immediatedly treated with an equal volume of 1.6% NaCl. The PBMC containing ring from the Isolymph centrifugation step was collected, washed, and monocytes were purified using CD14 microbeads (Miltenyi). The procedure routinely yield populations of purity greater than 95%.
Release of LDH in tissue culture media, a reflection of pyroptosis, was measured using the Roche Cytotox detection kit. BMDM, PMN, or monocytes (5×105 cells) were plated in 24 well plates. Bacteria at different MOI were added to the cell culture and the plates were centrifuged at 1500 rpm for 10 minutes to maximize and synchronize infection and incubated for 30 minutes at 37°C. Cells were washed with PBS to remove extracellular bacteria and medium containing kanamycin (200 µg/ml) was added to inhibit extracellular bacteria growth. Media were collected at 1, 2, 4, 8, 12 hours post infection for LDH measurement. Cells were lysed in PBS-2% saponin-15% BSA and serial dilutions of the lysates were plated on LB agar plates containing streptomycin (100 µg/ml) using the Eddy Jet Spiral Plater (Neutec). Bacterial colonies were counted 48 hours later using the Flash & Grow Automated Bacterial Colony Counter (Neutec).
Organs aseptically collected were weighted and homogenized in 1 ml PBS using 1 mm zirconium beads and the Mini16 bead beater. Serial dilutions were plated as described above.
Conditioned supernatants were separated by 12% PAGE electrophoresis, transferred to PVDF membranes, and probed with rabbit anti-caspase-1 (Upstate Biotechnologies) or goat anti-mIL-1β (R&D Systems).
BALF were collected from euthanized mice by intratracheal injection and aspiration of 1 ml PBS. Cytokines levels in tissue culture conditioned supernatants and BALF were measured using the Milliplex mouse cytokine/chemokine panel (Millipore) and confirmed by ELISA using the following paired antibodies kits: mIL-1β and mIFNγ (eBioscience), mIL-18 (MBL Nagoya, Japan). MPO level in BALF were measured using the HyCult Biotech ELISA kit.
Cells obtained from BALF were counted and stained with CD45, CD11b, CD11c, F4/80, GR1 (Ly6G), and analyzed with a LSRII BD flow cytometer.
Formalin-fixed paraffin-embedded lung sections were stained with H&E and scanned using the Aperio Scanscope XT. The Aperio ImageScope software was used to quantitate the area of the inflammatory foci compared to the total lung lobe area. Results from lungs from 5 animals per group were combined.
Total RNA was extracted using Trizol (Invitrogen) and 100 ng were amplified (27 cycles) using Superscript III One-step RT-PCR (Invitrogen) and primers specific for human Nlrc4, Nlrp3, Asc, and GAPDH (primers' sequence available upon request).
All data were expressed as mean ± S.E.M. Survival curves were compared using the log rank Kaplan-Meier test. 1way ANOVA and Tukey Post-test was used for analysis of the rest of data unless specified in the figure legends. Significance was set at p<0.05. Statistical analyses were performed using the GraphPad Prism 5.0.
UniProtKB/Swiss-Prot ID: IL-1β, P10749; IL-1R1, P13504; IL-18, P70380; NLRP3, Q8R4B8; NLRC4, Q3UP24; ASC, Q9EPB4; Casp-1, P29452; NAIP5, Q8CGT2.
|
10.1371/journal.pgen.1006260 | Novel Genetic Variants for Cartilage Thickness and Hip Osteoarthritis | Osteoarthritis is one of the most frequent and disabling diseases of the elderly. Only few genetic variants have been identified for osteoarthritis, which is partly due to large phenotype heterogeneity. To reduce heterogeneity, we here examined cartilage thickness, one of the structural components of joint health. We conducted a genome-wide association study of minimal joint space width (mJSW), a proxy for cartilage thickness, in a discovery set of 13,013 participants from five different cohorts and replication in 8,227 individuals from seven independent cohorts. We identified five genome-wide significant (GWS, P≤5·0×10−8) SNPs annotated to four distinct loci. In addition, we found two additional loci that were significantly replicated, but results of combined meta-analysis fell just below the genome wide significance threshold. The four novel associated genetic loci were located in/near TGFA (rs2862851), PIK3R1 (rs10471753), SLBP/FGFR3 (rs2236995), and TREH/DDX6 (rs496547), while the other two (DOT1L and SUPT3H/RUNX2) were previously identified. A systematic prioritization for underlying causal genes was performed using diverse lines of evidence. Exome sequencing data (n = 2,050 individuals) indicated that there were no rare exonic variants that could explain the identified associations. In addition, TGFA, FGFR3 and PIK3R1 were differentially expressed in OA cartilage lesions versus non-lesioned cartilage in the same individuals. In conclusion, we identified four novel loci (TGFA, PIK3R1, FGFR3 and TREH) and confirmed two loci known to be associated with cartilage thickness.The identified associations were not caused by rare exonic variants. This is the first report linking TGFA to human OA, which may serve as a new target for future therapies.
| Osteoarthritis (OA) is the most common form of arthritis and a leading cause of chronic disability in the western society affecting millions of people. OA is a degenerative joint disease characterized by changes in all joint tissues, including cartilage, bone and synovium, causing chronic pain and loss of function. There are no effective therapeutic treatments available for OA and therefore finding novel biological pathways through genetic association studies can open up new treatment options. The number of known DNA variants associated with OA-risk is limited.
To identify new loci, we have performed a Genome Wide Association Study meta-analysis on cartilage thickness, one of the joint tissues affected in OA in a total sample of more than 20,000 individuals from twelve cohorts. This analysis revealed six variants associated with cartilage thickness, four of these being novel associations, including TGFA as the most prominent one. A systematic prioritization for underlying causal genes, using diverse lines of evidence, highlighted genes underlying the disease associations, including TGFA, RUNX2 and PIK3R1. Large scale exome sequencing data (n = 2,050 individuals) indicated that there were no rare exonic variants that could explain the identified associations. This is the first report linking TGFA to human OA, which may serve as a new target for future therapies
| In spite of advances in the understanding of OA, the absence of effective therapeutic targets demonstrates that a better comprehension of its causes and pathophysiological mechanisms is needed. Since genome-wide genetic studies are hypothesis-free and do not suffer from the bias of previous knowledge, they have the potential to identify novel biological pathways involved in OA. The discovery of novel genes has the potential to identify novel treatment options. In addition, more personalized medicine approaches for OA can be explored through prediction of risk for disease as well as classification of disease subtypes.
Heritability of hip OA has been estimated to be around 40–60%. However, to date only few genetic variants have been successfully identified [1,2]. The reasons for finding only a modest number of genetic loci associated with hip OA can be attributed partially to relatively modest samples sizes in comparison to other complex diseases, such as myocardial infarction [3]. In addition, phenotype heterogeneity is an important issue in OA genetics as this is well known to reduce power to robustly detect signals. The problem of heterogeneity in genetic association studies of OA has been highlighted before [4,5] and is exemplified by the fact that the definition of the phenotype is a combination of bone and/or cartilage features as well as clinical complaints. Moreover, there is growing consensus that OA can be divided into multiple sub-phenotypes each with their own etiology and risk factors. For example, it has been demonstrated that individuals with hip OA, where only cartilage degradation is involved (atrophic OA form), are linked to a different systemic bone phenotype compared to individuals with OA where bone formation is also present [6].
As a way to overcome this, we examined a quantitative trait, which is one of the structural components of joint health, cartilage thickness, as a distinct phenotype.
Joint Space Width (JSW) is considered to be a proxy for cartilage thickness measured on hip radiographs. Minimal JSW (mJSW) has been shown to be a more reliable measure for hip joint health compared to the classical Kellgren & Lawrence score [7]. Previously, we have demonstrated that using only a modest discovery sample size (n = 6,000), we were able to successfully identify a genome-wide significant association of the DOT1L locus with mJSW as well as hip OA [1,8]. We now aimed to perform a more powerful analysis by combining data from five studies in the discovery phase, and subsequent replication in seven additional studies, amounting to a total sample size of 21,240 to identify new genes implicated in joint health using mJSW as a proxy for cartilage thickness. Using whole exome sequence data from 2,050 individuals we screened the discovered genes for potential functional variants. Subsequently we used multiple approaches that leverage different levels of information to enforce evidence of candidate genes annotated close to the associated signals.
Genome-wide Association analysis of mJSW of the hip with genetic variants was performed in a discovery set that included 13,013 individuals (see S1 Table and S1 Text for cohort specifics) with data on ±2,5 million genotyped or HapMap Phase II imputed SNPs. We applied extensive quality control measures (see S2 Table and S3 Table for details on quality control and exclusion criteria) leaving a total of 2,385,183 SNPs available for association analyses. Genomic control inflation factors for the P values of the RS, TwinsUK, MrOS, and SOF GWAS were low (λ = 1.02, 1.01, 1.02 and 0.99 respectively), and the interquantile-quantile plot (S1 Fig) also indicated no residual population stratification due to cryptic relatedness, population substructure or other biases.
The discovery analysis yielded eighteen independent SNPs with suggestive evidence for association (P <1*10−5) with mJSW, of which five (four genetic loci) met the genome-wide significance threshold of P-value ≤ 5*10−8 (see Fig 1). The top SNPs from these eighteen loci were selected for replication in additional 8,227 individuals from seven different cohorts. We observed that six of the eighteen SNPs significantly replicated (P<0.05) with the same direction of effect (see Table 1). When we combined discovery and replication results in a meta-analysis, the five SNPs that met genome-wide significance in the discovery analysis became more significant and another two SNPs that replicated in independent cohorts reached suggestive evidence (P≤ 1*10−6) for association in the combined meta-analysis.
The top signal in the combined meta-analysis, rs1180992 (Table 1, Pcombined = 3.2x10-16), is located in the intronic region of the previously OA associated DOT1L gene. This variant is very close to and in linkage disequilibrium with rs12982744 (D’ = 1, r2 = 1), which was previously found in association with mJSW and hip OA [1,8].
The DOT1L signal was followed in strength of association by rs2862851 (Pcombined = 5.2x10−11), which is annotated to the intronic region of TGFA (Fig 2A). Two variants near RUNX2, rs10948155 and rs12206662, also reached genome-wide significance for association with mJSW (Fig 2B). The two variants in the RUNX2 locus were weakly correlated (r2<0.2). Conditional analysis, using GCTA, showed that both SNPs represented different signals (S4 Table). Finally, the last signal that reached genome-wide significance was rs10471753, an intergenic variant closer to PIK3R1 (~450 Kb) than to SLC30A5 (~750Kb) (Table 1, Pcombined = 3.8*10−9).
Other suggestive signals for association with mJSW at a Pcombined≤ 1x10−6 including signals with significant replication were rs496547 (p = 1.5x10−7), a downstream gene variant located 3' of TREH and, an intron variant annotated near SLBP (rs2236995; p = 9x10−7). All other additional signals selected in the discovery stage did not replicate.
We examined whether the five GWS and two suggestive mJSW loci were also associated with hip OA in a total of 8,649 cases and >57,000 controls. Detailed description of the cohorts and OA definitions is given in S1 Table. Table 2 shows the associations found with hip OA. We observed that five of the seven identified mJSW loci were also associated with hip OA (p-value<0.05). Apart from the known DOT1L locus, the variant near TGFA was significantly associated with hip OA (Table 2, P = 4.3x10−5). In addition, the SNP near SLBP and the two SNPs near RUNX2 were associated with hip OA. One of the latter SNPs, rs10948155, is in high LD with a variant (rs10948172, D’ = 0.95 and r2 = 0.90) previously found in association with hip OA in males at borderline GWS level (2). However, in our study, rs10948155 was just marginally associated with hip OA in the overall analysis (Table 2, P = 0.021). We observed the second variant in this genomic region, an intronic variant in RUNX2, rs12206662, to have a larger effect size (β = 0.14, P = 1.1×10−4 r2 = 0.09 with rs10948172).
We further examined whether the identified loci were found associated with other phenotypes in earlier reports (Table 3). Five of the seven identified mJSW SNPs mapped to loci that have previously been associated with other bone-related phenotypes, primarily height. However, many of the identified height loci were not highly correlated with the mJSW signal (Table 3). Additional adjustment for height did not have an effect on the described association with mJSW; they showed an independent, possibly pleiotropic effect, on both traits. A particularly dense number of associations with different bone related phenotypes were present in the RUNX2 5’ region, where variants have been associated to BMD [10], height [11], osteoarthritis [2] and ossification of the spine [12]. Given the low LD between the variants underlying the different GWAS signals, it is likely that these represent independent associations.
We used multiple approaches that leverage different levels of information (e.g., gene expression, regulatory regions, published literature, mouse phenotypes) to prioritize candidate genes at each mJSW locus. Table 4 shows the summarized results from these analyses. In addition to the seven loci identified in the current study, we also analyzed five previously published loci for hip OA [2].
First, we focused on two gene prioritization methods: (i) DEPICT, a novel tool designed to identify the most likely causal gene in a given locus, and identify gene sets that are enriched in the genetic associations [21], and (ii) GRAIL which uses existing literature to identify connections between genes in the associated loci [22]. The DEPICT analysis yielded seventeen significantly prioritized genes (FDR >0.05), from which three genes were also significantly prioritized in the GRAIL analysis (S5 Table and S6 Table). Next, using the Online Mendelian Inheritance in Man (OMIM) database (http://omim.org), we identified genes with mutations implicated in abnormal skeletal growth in humans; for 50% of the loci, a skeletal syndrome gene was present (S7 Table). Similarly, we investigated if any of the genes had a known bone and cartilage development phenotype in mice. Very similar to the human phenotypes, the mice knockouts of the same genes resulted in bone and cartilage phenotypes (http://mousemutant.jax.org/) (S7 Table). Other supporting biological evidence that we gathered consisted of known expression quantitative loci (eQTL) and nonsynonymous variants in LD (r2>0.6) with the lead SNP of a locus (S8 Table and S9 Table), as well as expression in bone and cartilage tissue in mice using data from the Jackson lab database (S7 Table).
To further explore which genes are possibly underlying the genetic associations identified in this study, we analyzed gene expression in a paired set of non-lesioned and OA-lesioned cartilage samples of the RAAK study acquired from 33 donors at the time of joint replacement surgery for primary OA [19]. We first examined which genes are expressed in a set of seven human healthy cartilage samples (S7 Table). Additionally, we tested which of the genes located in 1MB region surrounding the lead SNP were differentially expressed in OA-lesioned cartilage versus non-lesioned cartilage of the same hip. Of the 152 genes that were selected, 129 genes were represented on the array. Of those, 64 genes were significantly expressed in the cartilage samples. For eight of the twelve loci, we found genes that were differentially expressed in OA lesioned cartilage versus non-lesioned cartilage (Table 4, S10 Table). Differential expression in cartilage healthy vs OA affected cartilage was performed likewise (S10 Table), while additionally adjusting for sex and age. Given the relatively small number of healthy samples (n = 7) with large age range these data are less robust and we did not use these data in gene prioritization.
For each gene a prioritization score was computed, based on equally weighting of the ten lines of evidence (Table 4). Following this approach, RUNX2 is highly likely to be the causal gene associated with rs12206662 and rs10948155. Similar strong evidence is found for rs788748 (IGFBP3) and rs10492367 (PTHLH). In addition, suggestive evidence for a causal gene is found for the following: rs10471753 (PIK3R1), rs835487 (CHST11), rs2862851 (TGFA), rs6094710 (SULF2), rs9350591 (COL12A1) and rs11177 (GNL3). However for some loci the current evidence is ambiguous, suggesting more than one gene as the potentially causal one; rs2236995 (FGFR3 or SLBP), rs11880992 (GADD45B or DOT1L) and rs496547(KMT2A or UPK2) (Table 4).
In 2,628 individuals from the Rotterdam Study, exome sequencing was performed at a mean depth of 55x. Of those, 2,050 individuals also had mJSW and hip OA phenotype data. Baseline characteristics of those individuals were similar to the source population, mean age was 67.3 years, 57% of the individuals were female and mean of mJSW was 3.81 mm (SD 0.82). Details of the experimental procedure and variant calling are given in the supplementary material (S2 Text). Only the variants with a minimal allele count of three in the total population were selected for analysis. Within the sixteen prioritized genes, a total number of 158 variants were identified in the protein-coding region, of which 85 were non-synonymous and one was a stop-gain mutation (Table 5, S11 Table).
We first performed a single variant test, where we tested each of the 86 variants changing the amino-acid sequence for association with the mJSW trait (S11 Table). We observed four nominal significant associations, with rare variants in SULF2, TGFA, RUNX2 and FGFR3. None of these rare exonic variants explained the original association between the GWAS hit and mJSW or hipOA when tested in a multivariate model (S12 Table). Next, we performed a burden test (SKAT) [23], to investigate whether the cumulative effects of the variants present in the sixteen selected genes were associated to mJSW, while adjusting for age and gender (Table 5). We observed a nominal significance burden test (p<0.05) for TGFA, SULF2, CHST11 and RUNX2 for mJSW. However, none of these findings reached significance after correction for multiple testing.
For most of the loci, no obvious protein-coding variants were found that could explain the associations. In previous studies it was shown that disease-associated variants are enriched in regulatory DNA regions [24,25]. We therefore examined whether the identified DNA variants (or SNPs in high LD) resided in chondrocyte and/or osteoblast specific enhancer regions, using data from ENCODE and ROADMAP [26–28]. To this end, we compared CHIP-seq signals from five different chromatin state markers (H3K4me3, H3K4me1, H3K36me3, H3K27me3, H3K9me3) in chondroblasts and osteoblasts to four cell lines from another origin. Together, these chromatin state markers identify promoter and enhancer activity in each of the cell lines. With the exception of rs2862851, we observed that for all mJSW genetic loci, SNPs in high LD were located in cell regulatory regions in chondroblast and/or osteoblast cells (see Fig 3 for an overview and S13–S18 Tables for each locus).
Only a modest number of genetic variants has been successfully identified through genome-wide association studies for OA This can in part be explained by the phenotypic heterogeneity of OA. Therefore, we used mJSW, a proxy for cartilage thickness in the hip joint, as one of the structural components of joint health. An additional advantage of this phenotype is its continuous nature, which increases power compared to a dichotomous trait, such as OA-status. We identified six independent loci associated with cartilage thickness in the hip joint, of which four surpassed genome-wide significance (TGFA, PIK3R1, SUPT3H-RUNX2, DOT1L) and two were suggestive for association with mJSW (SLBP/FGFR3, TREH-DDX6). Four of these loci (TGFA, SUPT3H-RUNX2, DOT1L and FGFR3) were also associated with hip OA.
The fact that we were able to identify six loci with the current sample size (13K individuals in the discovery) indicates that cartilage thickness is a phenotype providing a better yield in number of discoveries than the efforts ran with traditional composite radiographic scores. As a comparison, the largest GWAS study up to now, arcOGEN with 7,4K cases and 11K controls as discovery, yielded one locus in the overall analysis, and seven additionally in a number of stratified analyses. Interestingly, in the current manuscript we report on rs10948155, which is in high LD (r2 >0.8) with a locus from arcOGEN which was only marginally associated (p below genome-wide significance threshold) with OA in males only [2]. By using a cartilage specific endophenotype, evidence for this locus is elevated here to genome-wide significance in the total population, underscoring the increased power when more specific endophenotypes are used. Endophenotypes are quantifiable biological traits intermediate in the causal chain between genes and disease manifestation (in this case osteoarthritis). JSW can be precisely measured throughout the life of individuals [7] and also displays variation in normal subjects. Therefore, mJSW may be more tractable for the genetic dissection of OA.
Across the cohorts in this manuscript, mJSW has been measured in different ways, using both hand measured JSW on radiographs as well as (semi) automatic software which could have added some noise to the overall meta-analysis. Future cross-calibration of JSW measurements might aid in a more precise measurement and additional power to pick up genetic loci.
To the best of our knowledge, we are the first to scrutinize exome variants in relation to OA identified by large scale re-sequencing. We did not find low frequency exonic variants in any of the prioritized genes that could explain the observed associations with mJSW. We do have to keep in mind that the power of the exome sequencing effort is smaller than the original discovery analysis. We were unable to examine variants with allele frequencies below 0,07%. In addition, for rare or low allele frequencies, we only had power to detect relatively large effect sizes. For example, we had 80% power to detect a beta of 0,7 mm (almost 1SD) difference for a variant with 1% allele frequency. However, we tested all of the discovered exome variants in a multivariate analysis, and found that the novel identified rare exome variants did not affect the association between the GWAS-identified variants and mJSW in the same sample. This suggests that the associations between mJSW and the identified SNPs are not explained by rare exonic variants and likely exert their effects through regulation of expression. Indeed, supporting this hypothesis, we found that many these variants (or SNPs in LD) were annotated in regions that were annotated as regulatory active in chondroblastic and/or osteoblastic cells. However, more work is needed to examine the exact biological mechanism underlying the identified genetic loci.
TGFA (Transforming Growth Factor Alpha, rs2862851) was the strongest novel locus associated with cartilage thickness and hip OA. TGFA has been suggested to be involved in endochondral bone formation in mice, specifically the transition from hypertrophic cartilage to bone [29]. Recent, TGFA has also been implicated in the degeneration of articular cartilage during OA in rats [30]. Our results now imply a relationship between TGFA and human OA. In addition to the genetic association, we also show that TGFA expression is higher in human OA affected versus non-lesioned cartilage, possibly indicating that TGFA has a role in cartilage remodeling.
Functional characterization of the TGFA- associated locus by an examination of the histone methylation marks representing promoter or enhancer activity, did not reveal an obvious explanation for the functional impact of the SNP. However, the examined histone mark data represent unstimulated cells, and it is anticipated that the promoter and enhancer activity change upon stimulation of the cells. It is becoming more clear that effects of SNPs can be stimulus and context dependent, as has recently been shown for human monocytes, where many regulatory variants display functionality only after pathophysiological relevant immune stimuli [31].
The identified SUPT3H-RUNX2 locus contains two variants, rs12206662 and rs10948155, which are partially independent of each other. Where rs12206662 is located in the first intron of the RUNX2 gene near the second transcription start site (the so-called P2 promoter), rs10948155 is located more than 500kb away from RUNX2 between CDC5L and SUPT3H. However, rs10948155 is in high linkage disequilibrium with SNPs near in the P2 promoter and SNPs located in chondroblast specific enhancer regions (S16 and S17 Tables).Possibly, these enhancer regions regulate RUNX2 gene expression during endochondral differentiation. RUNX2 (Runt-related transcription factor 2) is a master transcription factor for controlling chondrocyte hypertrophy and osteoblast differentiation [32]. Previous genome-wide association studies have identified variants in the SUPT3H-RUNX2 locus associated with other bone and cartilage related phenotypes including height [14], bone mineral density [10] and ossification of the posterior longitudinal ligament of the spine [12]. All these previously published loci are independent of the two mJSW SNPs identified in the current study. We hypothesize that the SNPs are located in long-range enhancers, which regulate RUNX2 gene expression during endochondral differentiation via a chromatin-loop mediating protein.
We have also identified rs10471753, with PIK3R1 (Phosphoinositide-3-Kinase, Regulatory subunit 1 alpha) as the closest and strongest prioritized gene, related to rs10471753 associated with mJSW. Mutations in this gene are known to cause the SHORT syndrome, which is a rare multisystem disease with several manifestations including short stature, hernias, hyper extensibility and delayed dentition [33].Taken together with the fact that PIK3R1 is differentially expressed in OA affected cartilage, these results identify PIK3R1 as the most likely causal gene. Another possibility is that not PIK3R1 but rather a long-non-coding RNA (lncRNA), lnc-PIK3R1-4:1, is causal, since a variant in LD with the lead SNP is located in the predicted transcription start site of this lncRNA potentially affecting its expression. Although conserved in mice and zebrafish, thus far no function has been ascribed to this lncRNA [34].
We confirmed a locus previously associated with cartilage thickness, the DOT1L locus. Our identified SNP, rs11880992 is in high LD with the previously reported SNP rs129827744, and both are associated with cartilage thickness and hip OA [1]. Despite the previously presented suggestive evidence for involvement of DOT1L in chondrogenic differentiation, DOT1L did not receive a high score in our systematic prioritization study; the gene GADD45B, located in the region 500Kb downstream of the lead SNP, received a similar score. GADD45B is a transcriptional co-factor for C/EBP-β, a master regulator of chondrocyte differentiation [35]. Thus, it remains unclear which gene or genes in this locus contribute to the cartilage phenotype. Further research is needed to determine whether DOT1L is the true causal gene in this locus.
Our analyses suggest that the majority of prioritized genes in hip OA associated loci are involved in cartilage and bone developmental pathways; including TGFA, RUNX2, FGFR3, PTHLH, COL12A1 and others that seem to affect bone and/or cartilage development such as PIK3R1 and KMT2A We hypothesize that the mJSW and OA associated variants influence gene expression regulation. The dysregulation of these genes and mechanisms during development may, later in life, result in an increased risk for OA.
The identified mJSW SNPs are associated with hip OA, but not with knee OA. We have analysed the identified SNPs also for association with knee OA in the TREAT-OA meta-analysis dataset [36], but found no association. This observation fits in the overall finding that many of the identified genetic loci for OA seem to be site-specific [37], and support the hypothesis that the aetiology of OA is different in each joint. Nevertheless, this observation can still be a result of low power in the GWAS studies that have been done for OA till now [38], and final conclusions on this aspect cannot be drawn at this point.
This is the first report linking TGFA to human OA most likely by affecting mJSW. It may serve as a new target for future therapies. We have identified multiple mJSW associated loci which have previously been associated with other bone and cartilage related phenotypes such as bone mineral density and height, displaying a possible pleiotropic effect for the analysed traits. It will be important to understand how mJSW and OA associated variants can affect the developmental processes that regulate morphometry of the hip joint, including the formation of articular cartilage. Therefore further expression and functional studies are warranted of genes identified to be associated with hip OA phenotypes.
The participating studies were approved by the medical ethics committees of all participating centres, and all participants gave their written informed consent before entering the study
We conducted genome-wide association studies of mJSW for each cohort of the discovery stage: Rotterdam Study I (RS-I), Rotterdam Study II (RS-II), TwinsUK, SOF and MrOS using standardized age-, gender and population stratification (four principal components) adjusted residuals from linear regression. Cohort description and details of the single GWAS studies are given in S1 Text and S1 Table. The 6 cohorts used in the discovery stage were combined in a joined meta-analysis using inverse variance weighting with METAL [39]. Genomic control correction was applied to the standard errors and P-values before meta-analysis. SNPs with a P value < = 5×10−6 were selected for replication. The top SNPs for each independent locus were taken for replication in seven studies: the Genetics of Osteoarthritis and Lifestyle (GOAL) study, the Chingford study, CHECK (Cohort Hip & Cohort Knee), Genetics osteoARthritis and progression (GARP) study, the Genetics of Generalized Osteoarthritis (GOGO), the Johnston County Osteoarthritis Project (JoCo) and additionally the Nottingham OA case-control study for association with Hip OA (see Supplemental material for detailed information of the cohorts). Association of the SNPs with mJSW was additionally adjusted for height to test its independence. Secondary analyses included: association of the top SNPs with hip OA using logistic regression analysis (age and gender adjusted and by study centres an/or relatedness when it was pertinent). We used conditional analyses to investigate whether there are any independent signals in the identified associated loci, which were implemented using GCTA-COJO analysis [40].
The mJSW was assessed at pelvic radiographs in anteriorposterior position. The mJSW was measured in mm, along a radius from the center of the femoral head, and defined as the shortest distance found from the femoral head to the acetabulum. Within the Rotterdam Study, we used a 0.5 mm graduated magnifying glass laid directly over the radiograph to measure the minimal joint space width of the hip joints [41]. Within SOF and MrOS, a handheld caliper and reticule was used to measured mJSW to the nearest 0.1mm between the acetabular rim and proximal head of the femur [42]. For CHECK, mJSW was measured semi-automatic with the Software tool HOLY [43].
Radiographic hip OA was defined in the RS-I, RS-II, RSIII, Twins-UK, Chingford, and JoCo studies using Kellgren and Lawrence (K/L) scores. Hip OA cases were defined as a K/L score ≥ 2 on either side of the hip or THR due to OA. Hip OA controls were defined as no THR for OA and K/L score ≤ 1 and JSN ≤ 1. In MrOS and SOF cohorts, radiographic hip OA case-control was defined by a modified Croft grade, as previously described [44], where cases were defined as a Croft score ≥ 2 on either side of the hip or THR due to OA and controls were defined as a Croft score ≤ 1 on both sides of the hip and no THR. Hip OA cases in the GOAL and Nottingham OA studies were defined by having THR, and controls were radiographically free of hip OA, as previously described [45]. In GARP, hip osteoarthritis was defined as pain or stiffness in the groin and hip region on most days of the preceding month in addition to femoral or acetabular osteophytes or axial joint space narrowing on radiography or prosthesis due to osteoarthritis. In GOGO, hip OA was defined as KL grade > = 2, or minimal joint space width < = 2.5 mm, or the combination of joint space narrowing grade > = 2 and any osteophyte of grade > = 1, or history of joint replacement for OA. In JoCo, hip OA cases were defined as KL grade > = 2 or THR in at least one hip. Hip OA controls were defined as KL grade < = 1 in both hips.
We have used several available tools and publicly available databases to prioritize genes in known and newly discovered osteoarthritis associated regions. Locus gene sets were constructed by taking a region of 500 Kb upstream and 500Kb downstream of the lead SNP of that locus. We analysed 152 genes in 13 independent loci associated with minimal joint space width in the hip joint (mJSW) for 7 loci, hip OA for 4 loci, total joint replacement (TJR) for 1 locus and total hip replacement (THR) for 1 locus [2]. We analysed the following biological evidence for each gene at all loci; Nearest located genes: Taken from the UCSC genome browser, GRCh37/hg19 [46]. DEPICT gene prioritization: Data-driven Expression-Prioritized Integration for Complex Traits, a novel tool designed to identify the most likely causal gene in a given locus and to gene sets that are enriched in the genetic associations [21]. DEPICT was used to prioritize genes in a 1 MB region around the found SNPs that were significant associated with the osteoarthritis phenotype, taking a region of 500 Kb upstream and 500Kb downstream of the lead SNP of that locus. Gene prioritization analysis was performed to directly investigate functional similarities among genes from different associated regions, significance was defined by false discovery rate (FDR ≤ 5%). GRAIL gene prioritization: Gene Relationships Across Implicated Loci (GRAIL), was used to determine connectively between genes across OA implicated loci based on literature associations [22]. A GRAIL analysis was performed on 10 independent OA associated loci, based on existing literature in PubMed till August 2014. Mouse gene expression and phenotype: For each investigated gene, expression in mouse bone and/or cartilage tissue during several developmental stages as well as for adult tissue was determined using data from the Jackson lab database (http://www.informatics.jax.org/). In addition mouse phenotype data was also obtained for each gene. OMIM phenotype: Using the Online Mendelian Inheritance in Man (OMIM) database we examined which genes were involved in abnormal skeletal growth syndromes when mutated (http://omom.org). Expression quantitative trait loci: eQTL information was taken from the Blood eQTL browser (http://genenetwork.nl/bloodeqtlbrowser/) and the eQTL browser (http://www.ncbi.nlm.nih.gov/projects/gap/eqtl/index.cgi) using the lead SNP in each locus [20]. Non-synonymous variants: Last we determined if there were any nonsynonymous variants in LD (r2>0.06) with the lead SNP of a locus, using HaploReg V2 and the SNP Annotation and Proxy Search (SNAP) tools [47,48]. For each gene we assigned a score based on equally weighted lines of evidence.
We have used cartilage samples from the RAAK study to study gene expression in preserved and affected cartilage from individuals undergoing joint replacement [19]. The ongoing Research Arthritis and Articular Cartilage (RAAK) study is aimed at the biobanking of blood, joint materials (cartilage, bone and where available ligaments of knees and hips) and bone marrow stem cells (hip joints only) of patients and controls in the Leiden University Medical Center and collaborating outpatient clinics in the Leiden area. At the moment of collection (within 2 hours following surgery) tissue was washed extensively with phosphate buffered saline (PBS) to decrease the risk of contamination by blood, and cartilage was collected of the weight-bearing area of the joint. Cartilage was classified macroscopically and collected separately for macroscopically OA affected and preserved regions. Classification was done according to predefined features for OA related damage based on color/whiteness of the cartilage, based on surface integrity as determined by visible fibrillation/crack formation, and based on depth and hardness of the cartilage upon sampling with a scalpel. During collection with a scalpel, care was taken to avoid contamination with bone or synovium. Collected cartilage was snap frozen in liquid nitrogen and stored at -80°C prior to RNA extraction. Tissues have been stored tailored to apply staining and immunohistochemistry (IHC). Furthermore, DNA and RNA have been isolated from the preserved and affected areas of the respective tissues in order to apply genetic, transcriptomic and epigenomic profiling with respect to the OA pathophysiological process.
After in vitro transcription, amplification, and labeling with biotin-labeled nucleotides (Illumina TotalPrep RNA Amplification Kit) Illumina HumanHT-12 v3 microarrays were hybridized. Sample pairs were randomly dispersed over the microarrays, however each pair was measured on a single chip. Microarrays were read using an Illumina Beadarray 500GX scanner and after basic quality checks using Beadstudio software data were analyzed in R statistical programming language. Intensity values were normalized using the “rsn” option in the Lumi-package and absence of large scale between-chip effects was confirmed using the Globaltest-package in which the individual chip numbers were tested for association to the raw data. After removal of probes that were not optimally measured (detection P >0.05 in more than 50% of the samples) a paired t-test was performed on all sample pairs while adjusting for chip (to adjust for possible batch effects) and using multiple testing correction as implemented in the “BH” (Benjamini and Hochberg) option in the Limma-package. Analyses for differential expression between OA and healthy and between preserved and healthy cartilage was performed likewise, adjusting in addition for sex and for age.
Exome sequencing was performed in 2628 individuals from the Rotterdam Study the average mean coverage was 55x, corresponding to approximately 80% of the targeted regions covered by at least 20 reads. The exome sequencing was performed in house (HuGe-F, www.glimDNA.org). Details of the technical procedure and variant calling are given in S2 Text. We tested the exome variants for association with mJSW and/or hip OA in two ways. Each individual variant was tested for association with mJSW using the single variant option within RV-test, while adjusting for age and sex. In addition, we did a burden test for each of the selected genes by using SNP-set kernel association test (SKAT-O). SKAT aggregates individual score test statistics of SNPs in a SNP set and computes SNP-set level p-values for a gene [23].
For each of the top mJSW GWAS associated SNPs the LD region was determined using the 1000G Phase 1 population using the Haploreg tool [47]. The LD threshold was set at r2≥0.8. For each of these SNPs it was determined if the variant was located in a potential enhancer region using the Roadmap consortium reference epigenomes data set [27]. Heatmaps were constructed by calculating the percentage of variants in LD with the top mJSW GWAS found SNP located in enhancer regions as defined by the Roadmap epigenome chromatin states. The reference epigenomes were downloaded from the official data portal accompanying [27]. Reference epigenome data was used from mesenchymal stem cell derived chondrocyte cultured cells, Osteoblast, Bone marrow derived cultured mesenchymal stem cells, K562, HUVEC, HeLA and NHEK cells. Reference epigenomes were chromatin state models based on ChIPseq data of 5 core histone marks (H3K4me3, H3K4me1, H3K36me3, H3K27me3, H3K9me3) and an additional H3K27ac histone mark, the Roadmap expanded 18-state model.
ChIPseq data of mesenchymal stem cell derived chondrocyte cultured cells, and bone marrow derived cultured mesenchymal stem cells were generated by the NHI roadmap epigenomics project [28]. ChIPseq data of, Osteoblast, K562, HUVEC, HeLA and NHEK cells were generated by the ENCODE consortium [26]. All data and annotation tracks were downloaded through the UCSC genome browser table tool. Visualization of all ChIPseq annotation and roadmap full epigenomes tracks was done through the UCSC genome browser on GRCh37/hg19. Heatmaps were plotted in R using the CRAN software packages gplots and RcolorBrewer. Enrichment was calculated according to methods described in Trynka et al [25].
|
10.1371/journal.pcbi.1000217 | Inferring Pathway Activity toward Precise Disease Classification | The advent of microarray technology has made it possible to classify disease states based on gene expression profiles of patients. Typically, marker genes are selected by measuring the power of their expression profiles to discriminate among patients of different disease states. However, expression-based classification can be challenging in complex diseases due to factors such as cellular heterogeneity within a tissue sample and genetic heterogeneity across patients. A promising technique for coping with these challenges is to incorporate pathway information into the disease classification procedure in order to classify disease based on the activity of entire signaling pathways or protein complexes rather than on the expression levels of individual genes or proteins. We propose a new classification method based on pathway activities inferred for each patient. For each pathway, an activity level is summarized from the gene expression levels of its condition-responsive genes (CORGs), defined as the subset of genes in the pathway whose combined expression delivers optimal discriminative power for the disease phenotype. We show that classifiers using pathway activity achieve better performance than classifiers based on individual gene expression, for both simple and complex case-control studies including differentiation of perturbed from non-perturbed cells and subtyping of several different kinds of cancer. Moreover, the new method outperforms several previous approaches that use a static (i.e., non-conditional) definition of pathways. Within a pathway, the identified CORGs may facilitate the development of better diagnostic markers and the discovery of core alterations in human disease.
| The advent of microarray technology has drawn immense interest to identify gene expression levels that can serve as biomarkers for disease. Marker genes are selected by examining each individual gene to see how well its expression level discriminates different disease types. In complex diseases such as cancer, good marker genes can be hard to find due to cellular heterogeneity within the tissue and genetic heterogeneity across patients. A promising technique for addressing these challenges is to incorporate biological pathway information into the marker identification procedure, permitting disease classification based on the activity of entire pathways rather than simply on the expression levels of individual genes. However, previous pathway-based methods have not significantly outperformed gene-based methods. Here, we propose a new pathway-based classification procedure in which markers are encoded not as individual genes, nor as the set of genes making up a known pathway, but as subsets of “condition-responsive genes (CORGs)” within those pathways. Using expression profiles from seven different microarray studies, we show that the accuracy of this method is significantly better than both the conventional gene- and pathway- based diagnostics. Furthermore, the identified CORGs may facilitate the development of effective diagnostic markers and the discovery of molecular mechanisms underlying disease.
| Analysis of genome-wide expression profiles has become a widespread technique for identifying diagnostic markers of various disease states, outcomes, or responses to treatment [1]–[5]. Markers are selected by scoring each individual gene for how well its expression pattern can discriminate between different classes of disease or between cases and controls. The disease status of new patients is predicted using classifiers tuned to the expression levels of the marker genes.
One challenge of expression-based classification is that cellular heterogeneity within tissues and genetic heterogeneity across patients in complex diseases may weaken the discriminative power of individual genes [6]–[9]. In addition, marker genes are typically selected independently although proteins are known to function coordinately within protein complexes, signaling cascades, and higher-order cellular processes. Thus, the resulting expression-based classifiers may contain unnecessarily many marker genes with redundant information which may lead to decreased classification performance [10].
Due to these types of difficulties, several groups have hypothesized that a more effective means of marker identification may be to combine gene expression measurements over groups of genes that fall within common pathways [11]–[17]. The pre-defined functional groupings of genes are drawn from canonical pathways curated from literature resources such as the Gene Ontology [18] and KEGG databases [19] or experimentally defined gene lists from microarray studies [15],[16],[20]. Recently, pathway-based analysis has been extended to perform disease classification of expression profiles. Some approaches use gene expression parametrically by representing pathway activity with a function summarizing the expression values of member genes [21],[22], while others estimate probabilities of pathway activation based on the consistency of changes in gene expression [23],[24]. Alternative approaches engineer normal cells to activate pre-selected oncogenic pathways to determine gene signatures which can distinguish tumor characteristics [20],[25]. These methods have demonstrated classification accuracies that are comparable to conventional gene-based classifiers, while providing a strong biological interpretation for why the expression profile is associated with a particular type of disease (i.e., based on the pathways found to be perturbed). On the other hand, a potential shortcoming of current pathway-based classifiers is that the pre-defined set of genes making up a pathway may be derived from conditions irrelevant to the disease of interest. Moreover, not all the member genes in a perturbed pathway are typically altered at the mRNA level.
Here, we propose a novel gene-expression-based diagnostic that incorporates pathway information in a condition-specific manner (Pathway Activity inference using Condition-responsive genes, PAC). The markers are encoded not as individual genes, nor as static literature-curated pathways, but as subsets of condition-responsive co-functional genes (Condition-Responsive Genes, CORGs). To optimally discriminate samples of different phenotypes, we identify CORGs from each static pathway in the context of the specific disease in question. The combined expression levels of the CORGs are treated as the pathway “activity” and used to build classifiers for predicting the disease status of new patients. We show that our pathway-based approach outperforms previous analyses of differential expression in classifying samples across seven different datasets. Moreover, we show that pathway activities inferred using only CORGs lead to better classification performance as compared to pathway activities inferred using various types of summary statistics of all genes which participate in a common pathway. The resulting pathway markers and their CORGs also provide models of the molecular mechanisms which define the disease of interest.
We obtained previously published mRNA expression datasets covering seven different disease classification scenarios: 24 expression profiles of HeLa cells after stimulation by Tumor Necrosis Factor (TNF) [26], expression profiles of 62 primary prostate tumors and 41 normal prostate specimen [27], expression profiles of 143 acute lymphoblastic leukemia (ALL) patients [28], breast cancer expression profiles for 295 patients from the Netherlands [29] and 286 patients from the USA [5], and lung cancer expression profiles for 86 patients from Michigan [30] and 62 patients from Boston [31].
Each dataset was divided into two populations of distinct phenotypes as per the original publications (Table S1). For the TNF study [26], 12 samples had normal IkB proteins (labeled “Wildtype”) and 12 samples expressed mutant IkB blocking NF-kB signaling (labeled “Mutant”). For the prostate cancer study [27], 62 samples were retrieved from primary tumors (labeled “Cancer”) and 41 samples were from normal prostate specimen (labeled “Normal”). For the ALL study [28], 79 patients suffered from one subtype resulting from a t(12;21)(p12,q22) reciprocal translocation (labeled “TEL-AML1”) and the other 64 patients showed hyperdiploid hyperdip >50 (labeled “HH”). For the two breast cancer datasets, metastasis had been detected in 78 [29] and 106 [5] patients during follow-up visits within five and seven years after surgery (labeled “Metastatic”); the remaining 217 and 180 patients were still metastasis free (labeled “Non-metastatic”). For the two lung cancer datasets, we defined the two phenotype populations according to Subramanian et al. [15], who labeled 24 patients in the Michigan dataset and 31 patients in the Boston dataset as having a “Poor” prognosis, while the remaining 62 and 31 patients were labeled as having a “Good” prognosis.
For pathway information, we used the C2 functional set downloaded from MsigDB v1.0 [15]. This set includes 472 canonical metabolic and signaling pathways pooled from eight manually curated databases along with 50 co-expressed gene clusters obtained from various microarray studies. Each pathway or gene cluster defines a set of genes (gene clusters are henceforth also called “pathways”). In total, the available pathways covered 5602 genes, most but not all of which were measured in the seven gene expression datasets, due to the various array platforms used.
To integrate the expression and pathway datasets, we overlaid the expression values of each gene on its corresponding protein in each pathway. Within each pathway, we searched for a subset of member genes whose combined expression levels across the samples were highly discriminative of the phenotypes of interest (Figure 1). For a particular gene set G, let a represent its vector of activity scores over the samples in a study, and let c represent the corresponding vector of class labels (e.g. good vs. poor prognosis). To derive a, expression values gij are normalized to z-transformed scores zij which for each gene i have mean μi = 0 and standard deviation σi = 1 over all samples j. The individual zij of each member gene in the gene set are averaged into a combined z-score which is designated the activity aj (the square root of the number of member genes is used in the denominator to stabilize the variance of the mean). Many types of statistic, such as the Wilcoxon score or Pearson correlation, could be used to score the relationship between a and c. In this study, we defined the discriminative score S(G) as the t-test statistic [32] derived on a between groups of samples defined by c.
For a given pathway, a greedy search was performed to identify a subset of member genes in the pathway for which S(G) was locally maximal. We refer to this subset as the set of “condition-responsive genes” (CORGs) representing the majority of the pathway activation under the relevant conditions. To identify the CORG set, member genes were first ranked by their t-test scores, in ascending order if the average t-score among all member genes was negative, and in descending order otherwise. The CORG set G was initialized to contain only the top member gene and iteratively expanded. At each iteration, addition of the gene with the next best t-test score was considered, and the search was terminated when no addition increased the discriminative score S(G). The activity vector a of the final CORG set was regarded as the pathway activity across the samples.
We also used a method proposed by Tian et al. [16] to assess the probability of a pathway being altered in disease based on the correlation between the expression of all its member genes and the disease phenotype. For each pathway P in MsigDB, Tian et al. calculated a score T by averaging the t-test statistic scores of all member genes. Higher T was indicative of stronger pathway correlation with the disease status. The top 10% of pathways (52 pathways) in each dataset were selected for further analysis and for classification. The decision of whether a pathway had been disrupted by disease was assessed on the basis of the discriminating power of the member genes between the classes of interest (using a t-test statistic). However, there may be some signatures of pathway disruption that are independent of the classification task at hand. To detect such signatures, a number of statistical functions [8],[33] can be adopted in the framework of Tian et al. Unlike the t-test, these functions are designed to detect perturbed patterns rather than mean expression changes.
To compare our PAC with other activity inference schemes, we implemented three other expression summarization methods, including a principal component analysis (PCA) similar to that used in Bild et al. [20] and the mean and median approaches used in Guo et al. [22]. Bild et al. used the first principal component of the expression of the member genes to represent the activation of a given pathway, while Guo et al. summarized the expression levels of member genes by using simple statistics like mean and median.
For each dataset, 100 alternative two-fold splits were generated of each mRNA expression profile in the dataset. Pathways were ranked on each fold using the method of Tian et al. [16], and CORGs for each pathway were identified using the samples in a single fold. Individual genes were also ranked by their discriminative power on each fold. The robustness was estimated as the average degree of overlap among top pathways/genes derived from the two folds of samples across the 100 splits.
Logistic regression models [34] were trained on both the pathway activity matrix (pathways versus samples) and the original gene expression matrix (genes versus samples—i.e., conventional gene-based classification). For within-dataset experiments, the expression samples in a dataset were divided so that four-fifths of the samples were used as the training set to build the classifier, and one fifth were used as the test set (five-fold cross validation). Each of the five subsets in the dataset was evaluated in turn as the test set and withheld during marker selection (including CORG identification) and classifier training. In order to train a generalized classifier and to minimize over-fitting, we further split the training set into three smaller subsets of equal size: two subsets were used as the marker selection set to rank markers (pathways or genes) as well as identify CORGs (pathways only), and one subset was used as the validation set for assessing which marker set was significant for classification. Thus the CORGs might be different for a specific pathway, depending on the samples used in the marker selection set. Pathways or genes were ranked by the p-value of discriminative power to classify samples in the marker selection set, after which the logistic regression model was built by adding markers sequentially in increasing order of p-value (sequential selection). The number of markers used in the classifier was optimized by evaluating its Area Under ROC Curve (AUC, see [35] for details) on the validation set. The AUC metric captured performance over the entire range of sensitivity/specificity values. The final classification performance was reported as the AUC on the test set using the classifier optimized from the validation set. For unbiased evaluation, we generated 100 alternative five-fold splits of samples in each dataset and ran cross validation on each split. The final reported AUC values were averaged across 500 randomly selected ways of partitioning the data into four-fifths training and one-fifth test samples.
For cross-dataset experiments, markers (pathways or genes) were selected using the whole first dataset and then tested on the second dataset (or vice versa). CORG identification was also performed on the first dataset. As for the within-dataset experiments, the patient samples in the second dataset were divided into five subsets of equal size: four subsets were designated as the “training” set to build the classifier using markers from the first dataset, and one subset was held for testing. One hundred alternative five-fold splits were generated to partition samples in the second dataset into four-fifths for training and one-fifth for testing. Therefore, we learned 500 classifiers for each of these two datasets, in which each classifier was associated with its own pathway marker set. The averaged AUC values among the 500 classifiers built on the second dataset were reported as the final classification performance for each marker set identified from the first dataset. Among the 500 classifiers, the pathway marker set used in classification could be different depending on which training samples were used in the second dataset. However, the CORGs of each pathway were the same across these 500 classifiers because the identification was done using the whole first dataset.
In this study, for pathway-based classifiers, the input marker set was defined as the top 10% of pathways in MSigDB ranked by Tian et al. [16] using a designated training set. In order to compare pathway and gene based methods in a fair manner that controls for the number of genes used, we provided the gene-based classifiers with the same number of top ranked genes as the number of CORGs pooled from the significant pathways selected by Tian et al. [16].
We first tested the robustness of the pathway markers selected by the method of Tian et al. [16]. The agreement between the significant pathways was higher than that between the individually scored gene markers (Figure S1). The CORGs within the top pathways were also more consistent than individually scored gene markers in different subsets of samples. The observed robustness of CORGs might imply that some non-differentially expressed genes, which are often dropped in conventional analysis, do have associations with the disease of interest.
We hypothesized that pathway information could be used to restrict the search space for truly perturbed genes whose aggregated expression is more predictive for disease status than individually considered. We began by analyzing the breast and lung cancer datasets (four datasets in total), since each dataset has available two separate cohorts of patients studied by different researchers. The top 10% of pathways were selected for each of the four datasets (see Methods). We identified the CORGs for each top pathway and aggregated their expression levels into a single activity value for each sample (Methods). By design, the inferred pathway activities had more discriminative power in distinguishing samples with different disease phenotypes than did the individual expression levels of the member CORGs (PAC versus CORGs in Figure 2A, 2C, 2E, and 2G). However, the discriminative power fell when the pathway activity was inferred using not only the CORGs but all member genes associated with each pathway (PAC_all in Figure 2A, 2C, 2E, and 2G). This result suggests that, as might be expected, not all genes in a significant pathway are transcriptionally altered or associated with the phenotype of interest.
We then compared our pathway markers to the individual gene markers selected without pathway information. We found that the PAC activity scores outperformed individual gene markers in terms of discriminating samples with different disease phenotypes in both the source datasets used for marker identification (PAC versus Genes in Figure 2A, 2C, 2E, and 2G) and the independent verification datasets (Figure 2B, 2D, 2F, and 2H). In the verification datasets, the CORGs demonstrated almost the same discriminative power as did the top genes, although the top genes were more powerful in the original datasets. These comparisons suggest that aggregating the perturbed genes in a pathway leads to a better marker for discriminating disease phenotypes. Although the expression of a single gene might not be a strong predictor, pathway integration provides a means to amplify individual weak signals at the transcriptional level.
We next tested that the inferred pathway activity levels could be used in the classification of disease status for a new expression profile. To use pathway information for classification, pathway activities were used as feature values in a classifier based on logistic regression. The technique of five-fold cross validation was applied to test the predictive power of the pathway markers (see Methods). In each run of cross validation, we only considered the top 10% of pathway markers selected by Tian et al. [16] using the designated training data.
As shown in Figure 3A, our pathway-based classifiers (PAC) significantly outperformed the conventional gene-based classifiers (Gene). The improved performance was not simply due to grouping multiple gene expression measurements, as shown by comparing our performance with that of random groups of genes (PAC_random; averaged AUCs of 1000 sets of same-size random gene sets as the significant pathways). Classifiers using pathway activity inferred by the mean or median of the member gene expression [22] or the 1st principle component (PCA) [20] had higher predictive power than those using random gene sets (PAC_random), but only comparable power to the conventional gene-based classifiers. These results indicate that there are at least two critical factors in developing an advanced molecular diagnostic: (1) a biologically meaningful definition of pathways and (2) inference of condition-specific pathway activity.
Next, we tested the reproducibility of the pathway markers selected across different microarray platforms or different cohorts of patients. For this purpose, we used expression profiles of the two lung cancer datasets and the two breast cancer datasets generated from different groups. For each cancer, significant pathways and their CORGs were identified using the whole first dataset and then tested on the second dataset, or vice versa (Figure 3B). Our pathway-based classifiers again significantly outperformed the gene-based classifiers.
To show that the better performance of PAC was not dependent on the chosen classification algorithm, we evaluated all markers and pathway activity inference methods using three additional classification approaches: k-nearest neighbors, naïve Bayes, and linear discriminative analysis. Moreover, forward selection method was also employed to show our superior performance was not beneficial from the feature selection method used. All further analyses demonstrated the same trends, i.e., our CORG-based pathway classifiers outperformed other gene-based and pathway-based classifiers (Figures S2 and S3).
Beyond achieving better classification performance, the discriminative pathway markers and their CORGs can lend insight into the biological basis for why samples are classified as a specific disease status. As an example, we examined the pathway markers selected in the above two cross-dataset experiments for classification of lung cancer prognosis (for a similar analysis of breast cancer metastasis, see Table S2 and Figure S4). We counted the frequency with which each pathway in MSigDB was selected over the 500 classifiers, and we identified the top most frequent pathways having over 100 occurrences (Table 1).
Pathways involved in glucose metabolism (“Glycolysis” in Table 1) and estrogen signaling (“Breast cancer estrogen signaling” and “Estrogen receptor modulators down-regulated genes”) were frequently used in classifying lung cancer patients, and over-expression of these pathways had poor prognosis in both datasets (Figure 4). Constitutively up-regulated glycolysis has been observed in most primary and metastatic cancers and further explored to develop potential therapeutic targets [36]–[38]. Up-regulated glycolysis enables unconstrained proliferation and invasion and may lead to a more aggressive type of lung cancer [37]. Estrogen signaling has been known to promote cell proliferation and suppresses apoptosis, and its role in the late steps of lung metastasis has recently been suggested [39]. As shown in Table 1, many pathways could be represented by CORGs of the size from two to four, although some required more than eight genes (Figure S5). Especially for larger CORG sets, it would be computationally infeasible to identify these combinations to have maximal discriminative power in the absence of prior pathway knowledge.
We have demonstrated that effectively incorporating pathway information into expression-based disease diagnosis can provide better discriminative and more biologically defensible models. Grouping gene expression responses via functional linkages can amplify individually weak signals due to the heterogeneity of samples, either genetic or technical. In addition, such gene groupings also emerge as a critical step of removing potential redundancy on expression among genes associated with the same function. In view of classification tasks, genes of the same expression pattern do not provide extra information for a classifier but may cause over-fitting. The identification of condition-responsive genes within each pathway helps to reduce noisy or variable measurements, leading to a more precise and robust classifier. Better coverage and quality of human pathway information is likely to enable more precise prediction of disease status and, accordingly, better management of patient care. In addition, human interaction databases are growing exponentially at present, enabling further opportunities for unveiling novel functional pathways or complexes [40]–[43]. Integrating known pathways and novel hypotheses from protein networks with expression profiles and phenotypic information will lead to more effective molecular characterization of human disease [17].
|
10.1371/journal.pgen.0030163 | Genome-Wide Patterns of Nucleotide Polymorphism in Domesticated Rice | Domesticated Asian rice (Oryza sativa) is one of the oldest domesticated crop species in the world, having fed more people than any other plant in human history. We report the patterns of DNA sequence variation in rice and its wild ancestor, O. rufipogon, across 111 randomly chosen gene fragments, and use these to infer the evolutionary dynamics that led to the origins of rice. There is a genome-wide excess of high-frequency derived single nucleotide polymorphisms (SNPs) in O. sativa varieties, a pattern that has not been reported for other crop species. We developed several alternative models to explain contemporary patterns of polymorphisms in rice, including a (i) selectively neutral population bottleneck model, (ii) bottleneck plus migration model, (iii) multiple selective sweeps model, and (iv) bottleneck plus selective sweeps model. We find that a simple bottleneck model, which has been the dominant demographic model for domesticated species, cannot explain the derived nucleotide polymorphism site frequency spectrum in rice. Instead, a bottleneck model that incorporates selective sweeps, or a more complex demographic model that includes subdivision and gene flow, are more plausible explanations for patterns of variation in domesticated rice varieties. If selective sweeps are indeed the explanation for the observed nucleotide data of domesticated rice, it suggests that strong selection can leave its imprint on genome-wide polymorphism patterns, contrary to expectations that selection results only in a local signature of variation.
| Domesticated Asian rice is one of the oldest and most important crops in the world. Two main rice evolutionary lineages have been identified, and are thought to have been independently domesticated in Asia. We have examined patterns of DNA sequence variation in the genomes of rice and its wild ancestor to make inferences about the origin of domesticated rice. Population bottlenecks (a reduction in the size of the founding population) in the evolutionary transition from wild to cultivated species has long been thought to be the dominant force shaping patterns of molecular evolution during domestication. We find that the nucleotide variation patterns in rice are inconsistent with a simple bottleneck model. Rice genetic variation, however, can be explained by either a model that incorporates both a bottleneck and migration among rice variety groups, or a model that incorporates a bottleneck and multiple rounds of artificial selection on rice. Selection by humans is believed to have played an important role during crop domestication, and these results may suggest that strong, recurrent selection can leave a signal that can be observed throughout the genomes of domesticated species.
| Domestication is a complex, cumulative evolutionary process in which human use of organisms leads to morphological and/or behavioral changes distinguishing domesticated species from their wild ancestors [1,2]. Beginning with Charles Darwin [3,4], there has been strong interest in the study of domestication of crop species as a means of understanding the nature of selection. Moreover, domestication and the development of agriculture are arguably the most important technological innovations in human history [5]. Crop plant domestication was the linchpin of the Neolithic Revolution 10,000–12,000 years ago, in which hunter-gatherer groups transitioned into sedentary agricultural societies that gave rise to current human cultures [6]. With domestication came the availability of food surpluses, and this agricultural development led to craft specializations, art, religious and social hierarchies, writing, urbanization, and the origin of the state [5].
One of the earliest domesticated crop species is cultivated Asian rice, Oryza sativa L., which has become the world's most widely grown crop and has also assumed the stature of a key model system in plant biology. Rice consumption constitutes about 20% of the world's caloric intake, and in Asian countries, where over half of the world's population lives, rice often represents over 50% of the calories consumed [7]. Because of its small genome size, rice has been the first crop plant to have its whole genome sequenced [8–10].
A wealth of morphological, physiological, and ecological variation exists within cultivated Asian rice, reflected in the large number of recognized cultivars or strains [11,12]. Two main rice varietal groups, O. sativa indica and O. sativa japonica, have been recognized since ancient China [13]. Although phenotypic distinctions between these groups is not always straightforward, indica varieties tend to be found throughout the tropical regions of Asia and are primarily grown in lowland conditions, while japonica types are differentiated into tropical japonica, distributed in upland tropical regions, and temperate japonica, a recently derived group cultivated in temperate regions [11,13,14]. Additional variety groups include aus, drought-tolerant rice from Bangladesh and West Bengal, and aromatic, fragrant rice from the Himalayan range [14,15]. All rice varieties have a predominantly self-fertilizing mating system [13]. Both morphological and isozyme data have established that O. rufipogon Griff., a partially outcrossing species native to southern Asia, is the wild ancestor of domesticated rice [13].
In this paper, we describe the levels and patterns of DNA sequence polymorphism across the rice genome and that of its wild ancestor, O. rufipogon. To our knowledge this is the first genome-wide characterization of sequence variation in domesticated Asian rice, and we show that rice contains a unique pattern of excess high-frequency derived single nucleotide polymorphisms (SNPs) that has not been reported in other species. We develop four models to explain patterns of genetic variation in O. sativa and O. rufipogon, including a simple selectively neutral bottleneck model that has been previously thought to be the dominant demographic force shaping levels of nucleotide variation in crop species. We demonstrate that this simple bottleneck model is inadequate to explain the origin of domesticated rice. We conclude that either positive selection has made a significant impact on genomic polymorphism patterns, or that domestication involved an extremely severe bottleneck (∼99.5% reduction) coupled with gene flow among modern varieties and between domesticated rice and its wild ancestor.
To assess levels and patterns of polymorphism in the rice genome, we sequenced one hundred eleven randomly chosen gene fragments (sequence-tagged sites or STS) in a diverse panel of Oryza accessions, including 72 from O. sativa and 21 from O. rufipogon (Tables S1 and S2). Average silent (synonymous and noncoding) site nucleotide diversity (θπ) across all sampled loci in O. sativa is approximately 3.20 × 10−3 (Table 1). Levels of polymorphism in the wild ancestral species, O. rufipogon, are predictably higher than rice, with a mean silent θπ of 5.19 × 10−3 (Table 1). These levels of polymorphism are lower than those observed for maize, a domesticated outcrossing species [16], and Arabidopsis thaliana, a selfing, wild species [17,18].
To determine if any genetic differentiation due to population structure among rice groups is evident in these STS sequences, we used the Bayesian clustering program STRUCTURE [19]. The highest likelihood obtained was with a model specifying K = 7 groups (Figure 1; Table S1). Five groups occur within O. sativa and correspond to the traditional variety designations, as described previously [14]. Evidence of some limited geographical population structure is also observed in O. rufipogon (Figure 1; Table S1). Neighbor-joining analysis of the concatenated STS sequences (Figure S1) revealed two distinct clusters within cultivated rice; one comprises a tropical japonica, temperate japonica, and aromatic rice lineage, and another consists of aus and indica rice. The apparent monophyly of these major groups is consistent with at least two domestication events in rice [14,20–24]. The nesting of the aromatic and the temperate japonica variety groups within tropical japonica suggests the first two groups originated from secondary divergence events from the latter, although the lack of support for tropical japonica branches does not exclude other possible divergence scenarios (Figure S1). Indica and aus relationships, on the other hand, are consistent with rapid divergence after domestication or separate domestication events from the same ancestral gene pool. Within-group SNP levels of cultivated rice are lower than those of the whole species (Table 1), with subpopulations harboring between 19% (temperate japonica) and 43% (indica) of the polymorphism of O. rufipogon. Assuming separate domestication events, the japonica clade contains 42% and the indica clade contains 48% of the diversity levels found in O. rufipogon.
Because of the strong population structure evident in our rice sample, it is necessary to assess patterns of variation separately for each group when making inferences about the evolutionary dynamics of domestication. Indica and tropical japonica represent the most widely grown cultivars for each of the separate domestication events, and we limited our characterization of polymorphism patterns to these two groups. We examined the frequency spectrum of segregating sites within loci using Tajima's D [25], and found that O. rufipogon and the two main rice subspecies show an excess of rare alleles, as evidenced by the biased distribution of Tajima's D toward negative values (Figure S2; Table 1). Crops are expected to have gone through a population bottleneck during domestication, as only a limited number of founding individuals were brought into cultivation. The distribution of Tajima's D in the domesticated rice varieties is inconsistent with a recent bottleneck, however, as these should reduce levels of low-frequency variants and bias measures of Tajima's D toward positive values. It is possible that subsequent population expansion, due to the spread of rice agriculture, could be responsible for the over-representation of rare alleles segregating in domesticated rice varieties, or selection may have played a role.
We further examined the derived site-frequency spectrum across SNPs (i.e., the fraction of derived polymorphisms present at various frequencies within a group) in indica and tropical japonica. To infer ancestral alleles for each SNP, we used as an outgroup O. meridionalis, a species believed to have diverged from O. sativa ∼2 million years ago [21]. In each O. sativa variety we observed a large number of high-frequency derived mutations (i.e., derived SNPs above 70% frequency in the population) leading to a U-shaped frequency distribution (Figure 2); this type of pattern has not been reported at the genomic level in any other species.
Possible explanations for the excess of high-frequency derived SNPs in O. sativa include the misidentification of ancestral states due to shared polymorphism with O. meridionalis, or the occurrence of multiple mutations at given sites since divergence from O. meridionalis. However, both misidentification of derived alleles and multiple hits would be expected to also affect the site-frequency spectrum of O. rufipogon, which is not observed (Figure 2). This suggests that the O. sativa derived site-frequency distribution is a result of the domestication process. Furthermore, derived alleles at high frequency in the O. sativa varieties occur primarily at low to intermediate frequency in O. rufipogon, suggesting that such alleles have only recently increased in frequency (Figure S3).
We also checked the ancestral state calls in O. sativa using the African wild rice O. barthii. Although O. barthii is more closely related to O. sativa than is O. meridionalis, if we assume that both wild species share ancestral polymorphisms with domesticated rice, the possibility that we always identified the same alternative allele as derived in our sample should be low. Using this approach, we find that 88% of our ancestral SNP calls in indica and 86% in tropical japonica matched in O. barthii and O. meridionalis. Even when using only the matched calls (which is a very conservative criterion, since it does not take into account drift and/or fixation processes in O. barthii), the site frequency spectrum in O. sativa varieties remains U-shaped.
An excess of high-frequency derived SNPs is often interpreted as a result of genetic hitchhiking during recent selective sweeps [26]. Because the site-frequency spectrum in rice varieties is observed from randomly selected loci, and the loci contributing high-frequency derived SNPs are distributed across the genome (Figure S4), this pattern suggests that strong linkage to positively selected mutations occurred within most of the genome. However, demographic forces may have also played a role in shaping the rice genomes. We developed several demographic models and a multiple selective sweeps model to test which evolutionary processes may best explain the observed patterns of polymorphism in rice.
The most widely accepted demographic model for crop domestication is a neutral bottleneck model [27–29]. In this model, rice domestication is assumed to be a result of recent population divergence, with one of the two daughter populations experiencing a reduction in population size at divergence associated with the founder effect at the time of domestication, followed by population growth as cultivation of the crop increases. To fit this model to our data, we used a diffusion-based approach [30–32] to predict the pattern of allele frequencies in domestic and ancestral populations under selective neutrality.
Details of the inference procedure can be found in the Materials and Methods section. The composite-likelihood function we employed uses the reduction in diversity observed in either of the domesticated rice subspecies and the shift in allele frequency distribution to estimate four parameters: the time back until the start of domestication (τ1), duration of the bottleneck (τ2), ratio of current population to ancestral population size (ν2), and relative size of the bottleneck population to the ancestral population (νb). The duration of the bottleneck was assumed to be 25% of the time back until domestication (τ2 = 0.25 × τ1), which is consistent with archeological data suggesting it took ∼3,000 y from the time of initial cultivation (∼12,000 y ago) until the appearance of domesticated rice grains [33,34].
Bottleneck parameter estimates for indica and tropical japonica are broadly comparable, with a slightly more severe bottleneck in tropical japonica (Table 2). Assuming the time back to the beginning of domestication for both variety groups was ∼12,000 y [35], we can independently derive estimates of the current O. rufipogon effective population size, Nrufi, using the relationship τ1 × 2Nrufi = 12,000 (because τ1 is scaled by 2Nrufi). From the indica analyses, Nrufi is equal to 12,000/(2 × 0.1044) = 57,471, and from the tropical japonica analyses is equal to 12,000/(2 × 0.0508) = 118,110 (this exact value of Nrufi is important in scaling all of the estimated parameters into years and number of individuals). The indica-derived Nrufi estimate implies bottleneck and current estimated population size (Ne) for indica of (νb × Nrufi) = 1,413 and (ν2 × Nrufi) = 40,229 respectively. The second estimate suggests a bottleneck and current Ne sizes for tropical japonica of (νb × Nrufi) = 1,334 and (ν2 × Nrufi) = 46,889, respectively.
The differences in estimates of Nrufi from each analysis could be attributable to differences in the founding population of each variety group or differences in the timing of each domestication event. We note, however, that a bottleneck model conditioned on coincident domestication for indica and tropical japonica (equal τ1 values) differs only by 1.8 log likelihood units (unpublished data), suggesting that equal timing of domestication is likely to have occurred. An independent estimate of Nrufi can be found by using the estimated scaled population silent mutation rates (θW = 4Nrufi μ = 5.42 × 10−3 per bp; Table 1) and the observation that the O. rufipogon site-frequency spectrum is consistent with that of a population of long-term constant size (Figure 2). Assuming a neutral mutation rate of 10−8 per bp, yields a point estimate of Nrufi = 135,500, which is slightly higher, but close to the estimates found by conditioning on the start of domestication.
It is important to note that population bottlenecks alone would not generate the strong excess of high-frequency derived alleles and strong U-shaped site-frequency spectrum observed in O. sativa (Figure 2) [36]. In order to explain this aspect of the data, we considered several demographic models that included ancient subdivision in the ancestor of rice, a bottleneck at the time of domestication for each domesticated varietal group, and limited gene flow between the independently domesticated rice groups indica and tropical japonica. Ancient, strong subdivision is not evident in our O. rufipogon sample (Figure 1); Fst between Chinese and non-Chinese O. rufipogon is low, about 0.16, and no interior modes are evident in the site-frequency spectrum of O. rufipogon, as expected under subdivision. However, it is possible for limited gene flow in O. rufipogon to lead to some differentiation of allele frequency between groups, but not so much that it would have a strong effect on a combined O. rufipogon sample. Furthermore, the population bottlenecks induced by independent domestication events could amplify any allele frequency differentiation between indica and tropical japonica, and limited gene flow between these two groups could introduce ancestral alleles into each population, causing mutations previously fixed in one group to be observed as high-frequency derived alleles in the other.
To test the effect of ancestral population substructure within O. rufipogon prior to the domestication of the two O. sativa groups, we fit the parameters of a complex demographic model to our data using a composite likelihood technique (see Materials and Methods). We began by exploring a model with seven demographic parameters, which consists of O. rufipogon being subdivided into two demes of equal size, sharing on average MR migrants per generation. Current-day indica varieties are descended from one of these demes, while tropical japonica varieties descend from the other. During the domestication process, each population underwent a bottleneck that began τ1 generations ago (in units 2Nrufi) and had severity νb (the ratio of the reduced population size to the ancestral size). After τ2 = 0.25 × τ1 generations (∼3,000 y), both indica and tropical japonica partially recovered, instantaneously reaching a fraction υI and υJ of the ancestral size, respectively. Contemporary gene flow (since domestication) between tropical japonica, O. rufipogon, and indica is captured by the last parameter, the average number of migrants per generation between these demes (MS). This model was conceived because it incorporates key demographic features of rice or crop domestication (e.g., bottlenecks, two domestication events) and could conceptually generate the observed derived SNP site frequency spectrum.
In preliminary analyses, we found that the migration rate (MR) between the two ancestral O. rufipogon demes was very large, with the marginal likelihood surface for this parameter near its maximum value whenever MR > 7. This is consistent with our observations of limited population structure in O. rufipogon (above), and we therefore discarded ancestral population structure as a main contributor to the patterns observed in our dataset, and simplified the demographic model to consider only a single ancestral population from with both indica and tropical japonica derive (with migration rates among the three remaining demes, MS = 4Nrufim). This assumption reduced the computational complexity, so that the remaining parameters could be estimated via a grid search using an initial size of over 2,000 points with 1,000,000 coalescent simulations per point. The resulting model (which we refer to as the bottleneck plus migration model) has five free parameters with composite maximum likelihood estimates of MS = 7.0 (migration between demes), νb = 0.0055 (domestication bottleneck size), νI = 0.27 (ratio of indica to O. rufipogon Ne), νJ = 0.12 (ratio of tropical japonica to O. rufipogon Ne), and τ1 = 0.04 (start of domestication in units of 2Nrufi) (Table 2). It is important to note that coalescent simulations scale the migration based on population size, so the number of migrants entering into the tropical japonica population is smaller (0.5 × M × νJ = 0.42), than into indica (0.5 × M × νI = 0.945), and O. rufipogon (0.5 × M = 3.5).
In Figure 3, we report the profile composite-likelihood contours for the three key demographic parameters in the bottleneck plus migration model: migration rate, start of the bottleneck, and severity. The figure is constructed by holding two parameters fixed at a given point in the (x,y) plane, optimizing over the third parameter, and reporting the maximum likelihood attained for the (x,y) point (due to computational limitations the figure was constructed holding the ratio of current-day indica and tropical japonica populations at their maximum composite-likelihood estimates). We note that the three parameters are moderately to strongly correlated, but only a restricted set of values in high dimensional space is consistent with the data. These solutions all include: a very strong bottleneck (>99% reduction), high rates of migration within and between domesticated and wild populations of Asian rice (M > 5), and current-day effective population sizes for cultivated rice that are substantially smaller than those seen in the ancestral population. We also note that the model solutions show a positive correlation between size of bottleneck population and timing of the bottleneck, a negative correlation between size of the bottleneck and migration, and a negative correlation between migration and timing (consistent with the ∼2-fold difference in the estimated time of the bottleneck between the model with migration and the model without).
As can be seen in Figure 4, the expected site-frequency spectrum under the best fitting bottleneck plus migration model matches the observed frequency distributions fairly well for both O. rufipogon as well as indica, but not as well for tropical japonica. As expected, the total number of SNPs in each of the three populations is predicted quite well by the model. We quantified the fit of the model to the observed data using a modified Pearson Chi-square goodness-of-fit (GOF) statistic, and found that the best-fitting complex demographic model is an excellent fit to the marginal indica (GOFI = 20.26, p = 0.72) and O. rufipogon site-frequency spectra (GOFR = 7.57; p = 0.99), and an adequate fit to the tropical japonica site-frequency spectrum (GOFT = 37.83, p = 0.22). One interesting observation is that the demographic model underpredicts the excess of high-frequency derived alleles observed in tropical japonica—a potential indication of recent positive selection. Given that artificial selection was probably quite strong and frequent during and after domestication, we further explored models that incorporate selection during the domestication process of O. sativa.
Since strong selection is known to accompany crop domestication, we developed two alternative models incorporating multiple selective sweeps to explain the unusual polymorphism patterns in indica and tropical japonica. In a neutral locus linked to a single, recent selective sweep, let fi be the probability of observing a neutral mutation segregating at frequency i in a sample of size n, conditional on the locus being variable. An expression for fi has been derived [26] and further extended [37,38], and includes the genomic distance d (measured in bp) between neutral and selected loci, a compound parameter α, which represents the combined contributions of recombination, selection, and population size, and the “background” allele frequency distribution (i.e., the expected site-frequency spectrum for loci unlinked to a selected site).
These results for a single sweep can be used to predict the site-frequency spectrum at randomly chosen loci if multiple sweeps have recently occurred. Assuming that selective sweeps occur at random positions in the genome at a density of κ sweeps per bp, the distance between a random neutral locus and the nearest sweep will be approximately exponentially distributed with mean 1/(2κ). Define the function φi(d, α, κ) to be the probability of observing i copies of a neutral mutation in a sample of n chromosomes, given that a sweep occurred at a distance d bp away with compound parameter α [38], and background site-frequency spectrum q. By integrating over the distance between the sampled locus and the unknown target of the sweep, the marginal probability, Pi, of observing a randomly chosen SNP at frequency i in a sample of n chromosomes is a function of κ, α, and q [38]:
This probability can be used to calculate the composite likelihood of the data and estimate the parameters κ and α (see Materials and Methods). It should be noted that this equation assumes that the neutral locus is affected only by the nearest selective sweep.
We considered two distinct models. The first is a model in which strong selection is the only force that has acted in domesticated rice populations, and uses the normalized O. rufipogon site-frequency spectrum as the background frequency distribution. The second, a bottleneck plus sweeps model, allows multiple selective sweeps to affect patterns of variation immediately following a population size change. The background site-frequency spectrum in the latter case can be approximated using the predictions of a simplified neutral bottleneck model. The bottleneck plus sweeps model incorporates the sweep density κ, the compound parameter α (the combined contributions of recombination, selection, and population size), and a bottleneck severity parameter ν.
The likelihood surfaces for both the pure selection and the bottleneck plus sweeps model in rice each contains a long ridge where different parameter combinations have almost equally high likelihoods, implying that a model with high sweep density and relatively weak selection is just as likely as a model with low sweep density and strong selection (Figure 5). For both models, the ridge of maximum likelihood is shifted to the right in tropical japonica, indicating that for a given value of the selection severity parameter α, the sweep density in tropical japonica is estimated to be twice that in indica.
Sweep density is confounded with selection strength due to the effect of a mating system change on recombination rate. In domesticated rice, the transition to selfing likely occurred simultaneously with the sweeps, making it difficult to disentangle the recombination rate and selfing parameters. Under a recent selective sweep in a randomly mating population, the compound parameter α ≈ rs−1 ln(2N), where r is the per-basepair recombination rate, s is the selection coefficient and N is the population size [39]. In a partially selfing population such as domesticated rice, however, both effective recombination rate and population size are affected by selfing rate. While the rate of coalescence (and hence the effective population size) is at most doubled by the rate of selfing, the rate of recombination can be radically altered. An expression for effective recombination rate is r(1 − σ/[2 − σ]), where σ is the selfing rate [40]. For domesticated rice, estimates of selfing rates are typically ∼0.99 [13], resulting in a reduced recombination rate by approximately 10−3. If we assume 400 selective sweeps occurred in the rice genome since domestication (κ = 10−6), we estimate that α = 2 × 10−12 for indica. With r = 10−9 recombination events per generation per base pair and ln(2N) ≈ 10, this estimate of α corresponds to an unreasonably high estimate of a 5,000-fold fitness advantage. Substituting an effective recombination rate of 10−12 (corresponding to a reduced effective rate due to selfing), we find more reasonable values for the strength of selection for the selective sweeps, with s ≈ 5. This example illustrates how high selfing rates can amplify the signal of selection and contribute to the pattern of polymorphism in the rice genome.
Visually, it appears both the bottleneck plus sweeps model and the bottleneck plus migration model predict the site-frequency spectrum of domesticated rice better than the bottleneck model alone (Figure 4) or the pure selection model (unpublished data). To compare likelihoods and determine which model best fits the data, we used the Akaike information criterion (AIC) [41]. Since SNPs in our dataset are linked, we used a composite likelihood function and simulations to assign p-values to the observed AIC statistic (see Materials and Methods).
For indica, the bottleneck plus sweeps model is significantly better than the neutral bottleneck model (Λ = −17.18, p < 0.05) as is the bottleneck plus migration model (Λ = −14.19, p < 0.05). For tropical japonica, we also reject the neutral bottleneck model in favor of both the bottleneck plus sweeps model (Λ = −56.88, p < 0.01) and the bottleneck plus migration model (Λ = −53.60, p < 0.01). For both rice variety groups, the AIC for the bottleneck plus sweeps model was slightly lower than for the bottleneck plus migration models (Λ = −2.26, indica; Λ = −3.28, japonica), but this difference is likely not statistically meaningful given the various assumptions made. A separate (but not independent) assessment is comparing the fit of the predictions of each model to the data. The bottleneck plus sweeps model fits the marginal site-frequency spectrum of indica quite well (GOF = 13.86; p = 0.92), and does a slightly better job explaining the site-frequency spectrum of tropical japonica than does the complex demographic model incorporating bottlenecks plus migration (GOFsweeps + bottleneck = 31.21, p = 0.33; GOFbottlenecks + migration = 37.83; p = 0.22). These results underscore the importance of jointly modeling demographic and selective effects when considering the evolution of domesticated crop species.
Population bottlenecks are believed to be the primary demographic event associated with crop species origins, and are the accepted mechanism to explain observed genome-wide polymorphism levels among these taxa. There have been concerted efforts to model the impact of population bottlenecks on domesticated species genomes [27–29,42–44]. It appears from our results, however, that a population bottleneck alone is inadequate to explain the observed nucleotide polymorphism patterns in rice, one of the oldest and the most predominant food crop species in the world.
A more complex demographic scenario involving very strong bottlenecks that led to the fixation of alternate alleles during the two rice domestication events, with concurrent gene flow between variety groups, can explain the site-frequency spectrum of indica and O. rufipogon. However, this pure demography model requires a bottleneck 4-fold stronger in indica and twice as strong in tropical japonica relative to the model that incorporates selection (Figure 5; Table 2), and a relatively high migration rate between domesticated rice and wild O. rufipogon populations. It is also important to note that the model is a poor fit to the observed frequency distribution of alleles in tropical japonica.
Domestication, however, is characterized by strong directional selection on a suite of traits that lead to the establishment of cultivated species as distinct entities from their wild progenitors within agricultural settings. We show that, in contrast to the complex demographic model, a simple bottleneck with sweeps model fits data from both tropical japonica and indica well without requiring an extremely strong domestication bottleneck. Since domesticated Asian rice has been subject to artificial selection, the selection plus demography model is a very plausible explanation for the observed strong excess of high-frequency derived alleles in domesticated rice varieties, and is consistent with recent reports about domestication genes in rice [45,46].
Positive selection on specific genes results in reductions in variation within a genome through selective sweeps [47,48]. Unlike bottlenecks, however, selection is thought to have largely localized effects on genome variation. Our results suggest that a model that incorporates selection can explain patterns of nucleotide variation in a set of genome-wide markers. We suggest two reasons why selective sweeps during domestication could cause a genome-wide effect in O. sativa and not in other cereal crop species such as maize. First, the origin of domesticated Asian rice is associated with a transition to self-fertilization, which results in a low effective recombination rate and greatly increases the genomic distance affected by selection. Second, O. sativa possesses such a small genome (<400 Mb) that it is likely that a few dozen to hundreds of selective sweeps could leave a genome-wide imprint.
Interestingly, under the bottleneck plus selective sweeps model, the dynamics of domestication appear to differ in significant ways between indica and tropical japonica. Despite the fact that these two variety groups were domesticated from the same species and both have contributed significantly to Asian agriculture, it appears that the number of selective events and/or the bottleneck severity differs between them. It is possible that the two subspecies would diverge from each other in the demographic patterns associated with domestication, given that they were established by different cultures. If this is correct, then tropical japonica appears to have undergone a more severe bottleneck associated with domestication. Alternatively, it may be that the establishment of tropical japonica, which includes landraces that expanded to upland growing areas, may be associated with stronger selection pressures on a larger number of traits.
The process of domestication is one of recent, rapid species evolution, and studies on the dynamics of this process inform our understanding of the origins and diversification of new species. Simple demographic scenarios that have been employed in the past may not fully capture the domestication process of some crop species such as Asian rice. Our models indicate that selection and population bottlenecks together, or more complex scenarios that invoke very strong bottlenecks and current gene flow, could be responsible for determining genome-wide variation in the rice genome, a finding that has not been described in other domesticated species. Domesticated crop species are particularly suitable subjects in which to study the interaction between demographic events and selection in shaping species characteristics, and exploring the relative contributions of these forces require developing predictions for patterns of DNA polymorphism using models that allow selection to vary in timing (i.e., both during and after population bottlenecks) and strength. Nevertheless, our findings do underscore the possible role that selection may play in shaping genomic variation in domesticated species, reinforcing our appreciation of the foresight showed by Charles Darwin nearly a century-and-a-half ago [3] when he sought to illustrate the power of selection by drawing on the lessons learned from the evolution of domesticated species.
A panel of 72 O. sativa accessions was chosen to represent the diversity found within the species. These include representatives of five major subpopulations identified in a previous study [14], including 21 indica, 18 tropical japonica, 21 temperate japonica, six aus, and six aromatic accessions (Table S1). Most accessions are landraces, but five accessions studied correspond to modern cultivars. Also included in the panel were 21 accessions of the wild progenitor of rice, O. rufipogon, along with one sample each of O. nivara (a close relative of O. rufipogon not believed to have contributed to the ancestry of cultivated rice) and the outgroup species O. barthii and O. meridionalis (Table S1).
DNA was extracted from single plants as described in [49] with minor modifications. All O. sativa and one O. rufipogon accession (International Rice Germplasm Collection [http://www.irri.org/grc/] #105491) were self-fertilized for two generations prior to initiating the study. Seeds from O. rufipogon from Nepal were collected in the field by H. J. Koh and colleagues (Seoul National University); all other seeds were obtained from germplasm repositories as summarized in Table S1.
A total of 121 approximately 400–600 bp gene regions across the rice genome were chosen at random for sequencing from a set of 6,591 ESTs [50]. Four fragments were also selected from genes coding for well-known allozymes, including: catalase, acid phosphatase, pgi-a, and Adh. Primers were designed from the Nipponbare genomic sequence available from Gramene using Primer3 [51]. Primers were designed in exons, and attempts were made to include both exon and intron sequence within each fragment. DNA sequencing was carried out in Genaissance's sequencing facilities (New Haven, Connecticut, United States) as described in [52]. Amplification and sequencing were successful for 111 fragments referred to as STS (Table S2). Approximately 54 kbp per accession were sequenced, composed of, on average, 55% coding and 45% noncoding sequence.
Base-pair calls, quality score assignment, and construction of contigs were carried out using the Phred and Phrap programs (Codon Code). Sequence alignment and editing were carried out with BioLign Version 2.09.1 (Tom Hall, North Carolina State University, Raleigh, North Carolina, United States). Heterozygous sites were identified with Polyphred (Deborah Nickerson, University of Washington, Seattle, Washington, United States) and by visually inspecting chromatograms for double peaks. Heterozygous sites were rare for O. sativa. For heterozygous O. sativa and O. rufipogon sequences, heterozygous sites were labeled with ambiguity codes. For all analyses, the published sequence of Nipponbare was included.
To assess the sequencing error rate, 18 randomly chosen STS fragments were resequenced in a single direction for four Oryza accessions. Only three discordant base pairs within a single individual in a single fragment sequence were observed. This corresponds to three errors in 33,193 resequenced bp, or a sequencing error rate of less than 0.01%.
Population structure among O. sativa and O. rufipogon accessions was evaluated with STRUCTURE 2.1 [19] using an admixture model with no linkage. To limit the effect of correlation between SNPs due to linkage, one SNP per fragment (the SNP with the highest minor allele frequency across the entire accession set) was used in the analysis. O. sativa is primarily selfing, and most accessions exist as homozygotes; thus, SNP data were considered haploid for this species. O. rufipogon is partially outcrossing, a condition that cannot be adequately represented by considering each locus as diploid; thus, SNP data for O. rufipogon were also considered haploid. Because alternate alleles could occur at a given site in heterozygous O. rufipogon accessions, ten datasets were created with randomly chosen alternative base pairs in heterozygous individuals. Analyses were carried out for all ten datasets. All analyses had a burn-in length of 50,000 iterations and a run length of 100,000 iterations. Three replicates at each value of K (population number) were carried out. Simulations were run with uncorrelated allele frequencies. Results were entirely consistent among replicate runs within datasets and among datasets; the results from one run are presented in Figure 1 and Table S1.
To assess relationships among Oryza accessions, all STS fragment alignments were concatenated to form a single dataset. Relationships were estimated with a neighbor-joining analysis as implemented in PAUP* version 4.0 b3 [53]. Distances were calculated using the Kimura two-parameter model. Branch bootstrap estimates were obtained from 1,000 replicates.
Perl scripts were written to assess levels of nucleotide variation (θW) and nucleotide diversity (θπ) and Tajima's D across rice groups for all STS fragments, and to calculate the frequency distributions of derived SNPs across the genome. For O. sativa accessions, where heterozygotes were rare, all measures were calculated considering each accession as contributing a single haplotype; for O. rufipogon population measures, each accession was considered to contribute two haplotypes, except for one accession (International Rice Germplasm Collection [http://www.irri.org/grc/] #105491) from Malaysia, which had been selfed for several generations prior to this study.
Under a neutral bottleneck model, the history of rice domestication is represented by recent population divergence, with one of the two daughter populations experiencing a size bottleneck at divergence associated with the founder effect at the time of domestication. We use the sample frequencies of variable noncoding and synonymous nucleotides in the STS alignments (i.e., the site-frequency spectrum of putatively neutral SNPs) to infer the parameters of the bottleneck model. Our analytical approach makes use of standard Wright-Fisher population genetic theory within a Poisson random field setting [54–57]. The assumptions of this model include independence among SNPs, no selection, an underlying Poisson process governing mutations, and a piecewise constant population of large size amenable to modeling using diffusion approximations.
The model we employ is an extension of Williamson et al. [58], where we present the relevant population and statistical inference theory for modeling a population experiencing a recent size change. The key addition to our previous model is a second size change event, corresponding to the post-bottleneck growth phase. This amounts to modeling the components of the site-frequency spectrum (X1, X2, . . ., Xn) as independent Poisson random variables with mean:
where θ is the genome-wide mutation rate, x represents the (unknown) population frequencies of mutations, and f(x;Θ) is the distribution of mutation frequencies given demographic history parameters Θ = {ν,τ1,τ2}. These parameters are: the time back until the start of domestication (τ1), duration of the bottleneck (τ2), ratio of current population to ancestral population size (ν2), and relative size of the bottleneck population to the ancestral population (νb). The duration of the bottleneck was assumed to be 25% of the time back until domestication (τ2 = 0.25 × τ1), which is consistent with archaeological data suggesting domestication took 3,000 y and began 12,000 y ago. The mutation rate, θ, was estimated from the number of synonymous and noncoding segregating SNPs assuming O. rufipogon represented a population of constant size. This assumption is quite reasonable given the excellent concordance between the O. rufipogon and the predictions of the standard neutral model (Figure 4), and is equivalent to using Watterson's (1975) estimator of θ. In order to account for missing data, we fitted the population bottleneck model using the projected site-frequency spectrum for a sample of n = 16 chromosomes.
We considered alternative demographic scenarios, in which ancestral population subdivision, followed by gene flow between rufipogon, indica, and tropical japonica, led to an excess of high-frequency derived alleles in domesticated rice groups, as well as a simpler model that has no ancestral substructure. For these models, the composite likelihood function was based on the marginal site-frequency spectrum of each of the three groups analyzed. For ease of notation, let Sind, Sjap, and Sruf be the number of SNPs for which we could distinguish ancestral from derived alleles using the outgroup (223, 172, and 636, respectively). Let y denote the set of derived allele counts for each SNP, with y•ind, y•jap, and y•ruf referring to set of SNPs for indica, tropical japonica, and O. rufipogon (with lengths Sind, Sjap, and Sruf , respectively). To account for missing data, let n refer to the number of chromosomes sequenced at each SNP, with n•ind, n•jap, and n•ruf the vector for each group (again with lengths Sind, Sjap, and Sruf, respectively). For a given demographic model discussed above (the parameters of which we collectively denote Θ), the composite likelihood function is written as
where Pr(S•|Θ) is assumed to follow a Poisson probability of observing S• SNPs in a given population under the demographic model Θ assuming the population scaled mutation rate θ = 148.6 (estimated using the observed number of SNPs in O. rufipogon), and
is the probability of observing a SNP configuration in a given population under the demographic model. It is important to note that the inference scheme assumes the allele frequency distributions, conditional on the observed number of segregating sites and demographic parameters, are independent among populations. This composite-likelihood function (like all composite-likelihood functions) must, therefore, be taken as an approximation of the true likelihood function since it ignores dependencies among SNPs due to linkage and among populations due to shared variation. To account for missing data at an arbitrary SNP k in population x, we set
where Pz(Θ,Nx) is the expected proportion of SNPs at a frequency z in a sample of Nx chromosomes under the demographic model Θ, and the fraction within the summation represents the hypergeometric probability of sampling
derived alleles in a subsample of
chromosomes if the unknown frequency of the SNP were j out of Nx (summed over all possible underlying SNP frequencies, j). Details on calculating the expected number of SNPs in each population as well as Pz(Θ,Nx) are described below.
For a given set of parameters, Θ, we determine the expected site-frequency spectra for all three populations (O. rufipogon, indica, and tropical japonica) using 100,000 iterations of the coalescent simulation program ms [1] conditional on the observed genome-wide estimate of θ for O. rufipogon. To generate data under this model, we used the following code:
ms 80 200000 –t 148.6487 –r 148.6487 111 –I 3 21 18 41 M –en 0.5*0.75*τ1 1 νB –en 0.5*0.75*τ12νB –ej 0.5*τ1 1 3 –ej τ1 2 3 −em 0.5*τ1 3 1 0 –em 0.5*τ1 3 2 0 −n 1 νI –n 2 νJ.
Note that the factor 0.75 enters from the assumption that the bottleneck lasted 3,000 y of the 12,000 y time since domestication began, and 0.5 enters since ms scales time in units of 4N generations.
To optimize the three- and five-dimensional likelihood surface, we used an iterative technique, whereby a very coarse grid is initially chosen for each parameter, followed by successively tighter intervals containing the previous iteration's maximum likelihood estimates. Because we were pooling data across 111 STS loci, we generated our expected site-frequency spectrum accordingly. Although recombination within or between STS loci will not affect the expected number of segregating sites or the expected site frequency spectrum under a neutral demographic model, it does impact the rate at which simulations will approach them. We therefore assumed 111 mostly independent loci of equal size when generating our expectations.
In order to compare the fit of the demographic model to the observed data accounting for missing genotypes and partial selfing, we considered a projection of the observed and predicted site-frequency spectra into a sample of size n = 16 chromosomes from each of the three populations using the hypergeometric distribution. The “observed” data can be thought of as the predicted SFS in a subsample of n =16 based on the actual SNP data assuming each of the O. sativa accessions contributes one chromosome to the observed allele frequency spectrum, and each of the O. rufipogon accessions contributes two, with the exception of one accession that was known to have been purified. The “expected” data are the predicted marginal site-frequency spectrum at the maximum composite-likelihood estimates of the parameters from the complex demographic model that includes bottlenecks in the two domesticated populations, migration within domesticated populations, and migration between domesticated and ancestral populations. There were 45 observed data points (15 segregating site-frequency spectrum components multiplied by three populations), and the GOF statistic for a given population was tabulated as
. In order to assign a p-value, we simulated 10,000 datasets each containing 111 independent loci with no recombination within loci under the best-fitting demographic model. For each dataset, we then calculated the GOF test statistic using the expected site-frequency spectrum from Figure 4 scaled to the observed number of segregating sites within each of the subpopulations. Ideally, one would re-estimate the demographic parameters in order to fully mimic the inference procedure we used. Unfortunately, estimation of the demographic parameters was extremely computationally intensive for each dataset; the single observed STS data point analyzed here, for example, took over a week of computer time on a dedicated 100-node computing cluster.
Conditioning on the observed number of segregating sites in the dataset, the site-frequency spectrum is multinomially distributed with frequency probabilities according to Equation 1. For the pure selection model, the composite likelihood is:
where qr is the normalized site-frequency spectrum of O. rufipogon. For the bottleneck plus multiple sweeps model, the composite likelihood is:
where qν is the predicted spectrum from a neutral bottleneck model with severity ν. Equations 5 and 6 can be maximized to quantify the number and strength of selective sweeps in domestic rice, and the optimization of Equation 5 provides an estimate of the severity of the population bottleneck that preceded the selective sweep.
The bottleneck plus sweeps model assumes that a short bottleneck (representing to the founding of domestic populations) precedes the selective sweeps. To calculate the background site-frequency spectrum at the end of the bottleneck and the beginning of the selective sweeps, we again used numerical methods to solve the one-population diffusion equation with population size changes:
In this case, the recovery time, τ1, was set to 0, corresponding to the assumption that new mutations since the bottleneck do not make a strong contribution to the observed SFS. Because the bottleneck duration, τ2, and the severity, ν, are confounded parameters, we set τ2 = 0.01 and allow ν to vary. With f(q,τ2) as the numerical solution to Equation 7 evaluated at time τ2, we calculate the background site-frequency spectrum qν as:
To properly interpret differences in AIC between models, we simulated 10,000 datasets of 111 nonrecombining loci under the null hypothesis of the best-fitting neutral bottleneck model using the ms coalescent simulation program [59]. Because we did not allow recombination within loci, these simulations conservatively account for the effects of linkage. For each simulated dataset, we found the maximum composite likelihoods under each model (bottleneck, bottleneck plus migration, multiple sweeps, and bottleneck plus sweeps) and calculated the AIC value. The AIC statistic of model i is defined as: AICi = −2(lmaxi − ki) where lmaxi is the maximum likelihood under model i and ki is the number of free parameters in model i. We used Λ = AIC1 − AIC2 as a test statistic for comparing the bottleneck and alternative models using a one-tailed test: the p-value was estimated as the proportion of simulations under the null distribution with Λ > Λobs.
The National Center for Biotechnology Information GenBank (http://www.ncbi.nlm.nih.gov/Genbank) ID numbers for the sequences and alignments discussed in this article are EF000002–EF010509. |
10.1371/journal.pbio.1001811 | Motivational Salience Signal in the Basal Forebrain Is Coupled with Faster and More Precise Decision Speed | The survival of animals depends critically on prioritizing responses to motivationally salient stimuli. While it is generally believed that motivational salience increases decision speed, the quantitative relationship between motivational salience and decision speed, measured by reaction time (RT), remains unclear. Here we show that the neural correlate of motivational salience in the basal forebrain (BF), defined independently of RT, is coupled with faster and also more precise decision speed. In rats performing a reward-biased simple RT task, motivational salience was encoded by BF bursting response that occurred before RT. We found that faster RTs were tightly coupled with stronger BF motivational salience signals. Furthermore, the fraction of RT variability reflecting the contribution of intrinsic noise in the decision-making process was actively suppressed in faster RT distributions with stronger BF motivational salience signals. Artificially augmenting the BF motivational salience signal via electrical stimulation led to faster and more precise RTs and supports a causal relationship. Together, these results not only describe for the first time, to our knowledge, the quantitative relationship between motivational salience and faster decision speed, they also reveal the quantitative coupling relationship between motivational salience and more precise RT. Our results further establish the existence of an early and previously unrecognized step in the decision-making process that determines both the RT speed and variability of the entire decision-making process and suggest that this novel decision step is dictated largely by the BF motivational salience signal. Finally, our study raises the hypothesis that the dysregulation of decision speed in conditions such as depression, schizophrenia, and cognitive aging may result from the functional impairment of the motivational salience signal encoded by the poorly understood noncholinergic BF neurons.
| Humans and animals face the constant challenge of identifying the subset of incoming sensory stimuli that are most behaviorally relevant and prioritizing behavioral responses accordingly. Critical to this decision is the ability to determine whether a stimulus is motivationally salient—that is, whether the stimulus predicts important behavioral outcomes such as reward or punishment. While it is generally assumed that motivational salience is related to faster decision speed and shorter reaction time, it remains unclear how motivational salience actually modulates the decision-making process. This study investigates how the motivational salience signal in the basal forebrain controls the fundamental properties of the decision-making process—decision speed and its variability. In rats performing a reward-biased simple reaction time task, we show that the basal forebrain motivational salience signal is associated with a faster and also precise decision speed. In support of a causal role for this association, artificially augmenting this basal forebrain motivational salience signal by electrical stimulation also leads to faster and more precise reaction times. These results suggest that decision speed and its variability are jointly determined by an early and previously unrecognized step in the decision-making process, dictated largely by the motivational salience signal encoded by poorly understood noncholinergic neurons in the basal forebrain.
| The overall speed of information processing and decision-making has been studied for more than a century by measuring reaction time (RT) [1]–[3]. Significant increases in RT, reflecting a slower decision speed, represents a key feature in several conditions such as depression [4],[5], dementia [6],[7], schizophrenia [8]–[10], and cognitive aging [11],[12]. Therefore, it is important to understand how decision speed is modulated by cognitive variables and by underlying neural circuit mechanisms.
An important modulator of decision speed is motivational salience [13]–[16]. Determining whether a stimulus is motivationally salient—that is, whether the stimulus predicts important behavioral outcomes such as reward or punishment—allows humans and animals to select among incoming sensory information the subset of stimuli that are behaviorally relevant. Thus, motivational salience plays a key role in goal-directed decision-making to prioritize behavioral responses. As a result, neural correlates of motivational salience are commonly defined or inferred through the modulation of RT [13]–[15]. However, this logical interdependency poses a fundamental challenge in understanding the relationship between motivational salience and decision speed.
The alternative approach we took to avoid this circular logic was to investigate the relationship between RT and a neural correlate of motivational salience defined independently of RT. Recent studies identified a neural correlate of motivational salience in the basal forebrain (BF) [16], where a distinct group of BF neurons respond to motivationally salient stimuli that predict either reward or punishment with similar and robust phasic bursting responses [16]–[20]. The strength of the BF motivational salience signal, reflected by the amplitude of bursting response, is coupled with the overall response latency [16]. The same BF neurons also respond to primary reward and punishment with similar bursting response [16]. We hypothesize that the bursting response of BF neurons can translate the motivational salience signal into widespread modulation of cortical activity [21] and therefore represents an ideal candidate mechanism to increase decision speed and reduce RT. In support of this hypothesis, slowing of RT was observed following lesion [22]–[25] or inactivation [26],[27] of the BF.
Our approach to understand how BF motivational salience signal modulates decision speed was to first determine the part of RT variability that was correlated with, and likely modulated by, BF motivational salience signal while rats responded to two motivationally salient stimuli that predicted different amounts of reward. Second, we sought to determine the part of RT variability that was present in the face of constant BF bursting response, which does not reflect the moment-to-moment fluctuation of motivational salience and likely represents how RT is modulated by the intrinsic noise in the decision-making process. By partitioning RT variability into two distinct sources that were either correlated with or independent of BF motivational salience signal, we further investigated whether the strength of BF motivational salience signal modulated the magnitude of RT variability contributed by intrinsic noise. Finally, we tested whether the observed functional coupling between BF bursting response and RT represented a causal relationship using electrical microstimulation of the BF.
To investigate the relationship between motivational salience and decision speed, we developed a reward-biased simple RT task in rats that used differential reward expectations to modulate motivational salience (Figure 1A, Figure S1). Specifically, each trial started with a light signal that instructed rats to enter a nosepoke fixation port where they maintained fixation until an auditory stimulus instructed them to collect a water reward in the adjacent reward port. In the fixation port, rats heard, with equal probability and randomly across trials, either a sound predicting a large reward (S-Large), a different sound predicting a small reward (S-Small), or no sound and no reward (Catch). S-Large and S-Small were chosen to be clearly discriminable (white noise versus clicker) and presented at a suprathreshold level (80 dB for 2 s) to minimize sensory detection and discrimination effort. After sound onset, rats exited the fixation port quickly and moved to the adjacent reward port in almost all S-Large (99.8%) and S-Small (99.4%) trials, and only did so occasionally in Catch trials (3.8%). This response pattern confirmed that rats treated both S-Large and S-Small as motivationally salient stimuli. The latency between sound onset to fixation port exit, defined as RT, reflected the speed of the initial decision-making process in response to motivationally salient sounds and is the focus of our study.
While both S-Large and S-Small predicted reward in the same port and therefore commanded the same behavioral response without the need of a choice, well-trained rats automatically responded faster in S-Large than in S-Small trials (Figure 1B, Figure S1), indicating that the stimulus paired with larger reward was motivationally more salient. The modulation of RT between S-Large and S-Small trials continued to grow and did not reach asymptotic level after 10 training sessions (Figure 1C, Figure S1). Following the reversal of sound-reward contingency, it took rats on average three sessions to reverse their RT bias and began to show faster RT toward the new S-Large (Figure 1C, Figure S1). Thus, the reward-biased simple RT task provided a large dynamic range of RT modulation between S-Large and S-Small trials.
The reward-biased simple RT task was designed to minimize several variables that also affect RT: First, the influence of trial-by-trial variation in motivational state such as fatigue and satiety was minimized by requiring rats to initiate each trial with the same nosepoke fixation response. Second, the influence of choice—that is, choosing between different response options—on RT was minimized because both S-Large and S-Small signaled reward in the same port. Third, the influence of stimulus uncertainty and sensory decision-making process on RT was minimized by using sounds well above the detection threshold. Finally, this task design ensured that any other variable that affects RT at the time of sound onset, such as temporal expectation of stimulus onset or the spontaneous activity state of the neural network, should similarly affect both S-Large and S-Small trials. These behavioral and neuronal variabilities not directly controlled by the experimenters are collectively labeled as intrinsic noise of the decision-making process in the current study and should be equivalent between S-Large and S-Small trials. As a result, the RT difference between S-Large and S-Small trials must arise from the difference in the properties of the stimulus, such as the associated motivational salience. Therefore, the reward-biased simple RT task serves to provide the necessary simple behavioral context to reveal the quantitative relationship between BF motivational salience signal and RT, while minimizing the influence of other variables.
We first investigated whether BF motivational salience signal occurred early enough to modulate RT and decision speed. In six rats recorded over 40 sessions, we recorded 309 well-isolated single units in the region of the BF where cortically projecting BF neurons are located (Figure S2) [28]–[30]. Of these BF neurons, 47% (144/309) showed prominent bursting responses to sound onset and were classified as BF bursting neurons (Figure 2B, Figure S3). The same neurons also showed bursting responses to the trial start light signal (Figure 2A) and reward delivery (Figure 2D), consistent with the encoding of motivational salience as we previously reported [16]. This bursting response began at 50 ms after sound onset and peaked at 120 ms (Figure 2B), considerably earlier than almost all RTs in all behavior sessions (Figure 1B), and largely dissipated when rats exited the fixation port (Figure 2C). Therefore, BF motivational salience signal occurred early enough in the decision process to modulate the fixation port exit RT.
Next, we investigated whether the strength of motivational salience, defined as the amplitude of the BF bursting response, was correlated with decision speed. A typical example of salience-encoding BF neuron is shown in Figure 3A, which illustrates how these neurons respond to both S-Large and S-Small onset at a fixed latency relative to stimulus onset and robustly in each trial in well-trained animals. This example neuron also illustrates the common finding that the strength of BF bursting response was stronger toward S-Large than toward S-Small onset (Figure 3B,C). This is consistent with the intuition that pairing with the larger reward should endow a stronger motivational salience to S-Large than to S-Small.
The critical comparison was whether the RT modulation between S-Large and S-Small trials was correlated with the modulation of BF bursting amplitude between these two trial types. We found that the modulation of BF bursting amplitude at both single neuron and population level were highly correlated with the modulation of mean RT between S-Large and S-Small trials in a session (Figure 3D, Figure S4). The modulation of BF bursting tracked the modulation of mean RT even during the first three sessions of reversal learning when the RT bias had not been updated to reflect expected reward (red dots in Figure 3D). These findings provide key support of our hypothesis that the difference in decision speed between the two trials types was mostly driven by the difference in their associated motivational salience, encoded in the BF. Since S-Large, S-Small, and Catch trials were randomly intermingled in a session, BF motivational salience signal is coupled with RT on a single trial basis.
Having demonstrated the strong coupling between BF bursting amplitude and RT modulation, we further investigated whether this coupling was similarly present within, and not just between, S-Large and S-Small trials. We reasoned that if BF bursting amplitudes similarly predicted RT within the same trial type, the largest difference in BF bursting amplitude should be observed between the fastest and slowest trials. However, we found that there was little modulation of BF bursting strength even between these trials that had the largest RT difference within each trial type (Figure 4A,B, Figures S5 and S6), and the modulation of BF bursting amplitude did not correlate with RT modulation (Figure 4C,D). These results suggest that the trial-by-trial RT variability within each trial type was not correlated with the moment-to-moment fluctuation of BF motivational salience signal across trials. Instead, the within-trial-type RT variability likely reflected the contribution from the intrinsic noise of decision-making process in the presence of highly similar BF bursting amplitude and behavioral states across trials.
To better understand the nature of the within-trial-type RT variability, we noted its similarity with the considerable RT variability in humans when sensory ambiguity is reduced to a minimum, which has long been proposed to reflect the contribution of intrinsic noise in the decision-making process [31]–[33]. In humans, RT variability to suprathreshold sensory stimuli like the ones used in the current study, but not RT variability toward ambiguous sensory inputs, is highly structured and best described by the recinormal distribution [31],[32],[34]. Recinormal RT distribution means that the reciprocal of RT (1/RT) is normally distributed and that the RT distribution can be transformed into a straight line by plotting 1/RT versus its z-score (Figure S7). Therefore, we tested whether recinormality can be extended to the rat and found that the within-trial-type RT variability was well-described by the recinormal distribution (Figure 5A, Figure S7). This finding supports the universality of recinormal RT distribution across species and response modalities, and supports the hypothesis that within-trial-type RT variability likely reflected the contribution of intrinsic noise in the decision-making process.
The empirical observation that the within-trial-type RT variability is well described by the recinormal distribution suggests that the recinormal distribution provides an ideal quantitative framework to understand the two sources of RT variability in our study. Specifically, we hypothesize that the variability (σ) of the underlying normal distribution provides an estimate of the influence from intrinsic noise on RT, whereas the mean speed (μ) of the normal distribution captures the between-trial-type RT variability modulated by the BF motivational salience signal. Therefore, the quantitative relationship between μ and σ parameters of S-Large and S-Small RT distributions should provide insights on the relationship between the two sources of RT variability, and by extension the relationship between BF motivational salience signal and intrinsic noise of the decision-making process.
Interestingly, we observed that S-Large and S-Small RT distributions in a session often intersected near their respective fastest RT at a fixed time point around 160 ms (Figure 5B), suggesting that all RT distributions swivel against a fixed time point. To further investigate the consequences of swiveling against a fixed time point, we solved the linear equations for the two RT distributions with parameters (μ1, σ1) and (μ2, σ2) under the constraint of a fixed intersection point (Figure 5C). The fixed swivel point predicted two invariant relationships between parameters (μ1, σ1) and (μ2, σ2), which revealed previously unknown and exceedingly strong correlations between μ and σ parameters of RT distributions (Figure 5D,E). These novel correlations indicate that RT distributions with a larger mean speed μ (faster RTs) are associated with a shrinking variability σ. When μ approaches the theoretical limit (vertical black dotted line in Figure 5F), the mean RT approaches its fastest limit equivalent to the swivel point ∼160 ms while the RT variability (σ) shrinks to zero. This extreme scenario underscores the general finding that the RT variability (σ) does not scale proportionally with RT (1/μ), as would be expected from Weber's law, but in fact shrinks much faster. In the broader context of between- versus within-trial-type RT variability, the novel correlations between μ and σ revealed here suggest that these two sources of RT variability are tightly co-regulated and not independent. The magnitude of within-trial-type RT variability (σ), reflecting the magnitude of contribution from intrinsic noise, is actively suppressed in faster RT distributions with higher speed (μ) and stronger BF motivational salience signals.
The co-regulation of μ and σ parameters of RT distributions further suggests that each RT distribution can be determined with only one free parameter instead of two. From this perspective, the organization of RTs in our study can be viewed as a family of RT distributions swiveling against a fixed time point around 160 ms, with only one degree of freedom (Figure 5F). As such, our analysis predicts that this family of RT distribution can be generated by one single neural mechanism with three predicted properties essentially those of the BF motivational salience signal. First, this neural mechanism should occur before the fixed swivel point (∼160 ms) like the BF bursting response (Figure 2). Second, the activity of this neural mechanism should determine the speed (μ) and variability (σ) parameters of RT distributions, similar to how BF bursting amplitude is correlated with the mean RT (Figure 3). Third, it predicts that the same speed (μ) and variability (σ) parameters should be shared by all trials within a recinormal RT distribution regardless of whether RT is fast or slow, similar to the invariant BF bursting amplitude across all trials within a trial type (Figure 4). Therefore, these results support that BF motivational salience signal serves as a neural correlate of RT distribution parameters.
Finally, to investigate whether the observed coupling between the strength of BF motivational salience signal and the speed and variability of a RT distribution represents a causal relationship, we tested the prediction that augmenting the strength of BF motivational salience signal via electrical stimulation should produce faster and more precise RT distributions. The experimental setting was the same as the reward-biased simple RT task, except that S-Large and S-Small sounds were replaced by a common 6 kHz tone in both trial types and either paired with or without BF microstimulation (Figure 6A, Figure S8). The BF electrical stimulation was precisely timed to coincide with the tone-induced bursting response as a way to artificially augment the BF bursting amplitude. Under this protocol, RTs in the stimulated trials indeed became faster relative to nonstimulated control trials (Figure 6B–D). This RT difference grew larger as the stimulation current increased, consistent with the observation that greater bursting amplitudes were associated with faster RT distributions. Furthermore, BF electrical stimulation also produced more precise RTs such that the coupling between μ and σ parameters of RT distributions remained largely unchanged (Figure 6E and Figure S9). This result therefore supports a causal role of the BF motivational salience signal in determining both the speed and variability of RT distributions.
This study examined how motivational salience modulates decision speed. Our results provide strong support that the motivational salience signal in the BF, encoded by the phasic bursting response [16], is a major determinant of decision speed. We found that the BF bursting response took place early in the decision-making process and occurred early enough to modulate RT in the reward-biased simple RT task (Figure 2). RT variability in this task can be partitioned into two distinct sources, with the between-trial-type RT modulation tightly correlated with the strength of BF motivational salience signal (Figure 3), whereas the within-trial-type RT variability was unrelated to the BF motivational salience signal and likely reflected the intrinsic noise of the decision-making process (Figure 4). Analysis of the organization of RTs using recinormal distribution revealed that these two sources of RT variability were highly coupled, where RT distributions with fast mean RTs were associated with shrinking RT variability (Figure 5). Artificially augmenting the BF bursting amplitude via electrical stimulation increased decision speed as a function of stimulation current amplitude and also reduced variability, consistent with a causal relationship (Figure 6). Together, these results support the hypothesis that the BF motivational salience signal increases decision speed and also suppresses the contribution of intrinsic noise on RT variability.
Although the correlation between BF motivational salience signal and RT was fully predicted based on the literature [13]–[16], this is the first study, to our knowledge, to demonstrate the quantitative relationship between RT and a neural correlate of motivational salience defined independently of RT. BF bursting amplitude and RT were correlated on a single trial basis because the three trial types (S-Large, S-Small, and Catch) were intermingled and randomly presented in the session. Therefore, BF bursting amplitude fluctuated significantly across trials (of different trial types) and provided a good predictor of the RT on that trial. Within the same trial type, however, BF bursting amplitude remained highly consistent across trials, reflecting the highly similar behavioral and motivational states.
The reward-biased simple RT task was designed to minimize the influence of other variables on RT such that RT variability was mainly driven by motivational salience and by the intrinsic noise of the decision-making process. To determine the respective contributions of BF motivational salience signal and intrinsic noise on RT, our approach was to systemically vary the contribution from these two sources of RT variability. When the BF bursting amplitude was held constant within a trial type (Figure 4), the trial-by-trial RT variability was unrelated to the moment-by-moment fluctuation of BF motivational salience signal and therefore reflected the contribution from intrinsic noise of decision-making. When BF bursting amplitude was modulated between S-Large and S-Small trials (Figure 3), BF motivational salience signal accounted for most of the between-trial-type RT variability. The contribution of BF bursting on RT was further supported by the responses to the trial start light signal (Figure S10), in which the large fluctuation of BF bursting amplitude across trials was the main contributor to RT variability and was associated with clearly visible trial-by-trial coupling between BF bursting amplitude and overall response latency. Therefore, the lack of correlation between BF bursting amplitude and RT within a trial type does not mean that BF bursting is not correlated with RT. Rather, it reflects a principled approach to estimate the contribution of intrinsic noise on RT variability.
Our findings support that the recinormal distribution provides a quantitative framework, across species and response modalities, to describe the contribution of intrinsic noise on RT variability. Although recinormal RT distributions have been described to swivel against fix points under other behavioral contexts [35]–[38], our finding is the first, to our knowledge, to describe that RT distributions can swivel against a fixed time point near their respective fastest RT (Figure 5). This novel swiveling pattern revealed previously unknown correlations between the speed (μ) and variability (σ) parameters of recinormal RT distributions, and suggests that the contribution of intrinsic noise can be actively suppressed to near zero RT variability in the presence of a fast mean RT and strong BF motivational salience signal. This analysis also suggests that RT in a single trial is jointly determined by the parameters (μ and σ) of the recinormal distribution, and by the stochastic intrinsic noise that randomly draws from the recinormal RT distribution. Determining the parameters of the recinormal distribution therefore represents a previously unrecognized yet essential step in the decision-making process, which determines both the decision speed and its precision and is dictated largely by the BF motivational salience signal.
Our study is also the first, to our knowledge, to demonstrate that artificially increasing BF bursting amplitude through BF electrical stimulation was sufficient to produce faster and more precise RTs (Figure 6). Because stimulated and nonstimulated trials (as well as catch trials) were randomly intermingled in the session, our results suggest that the effect of BF electrical stimulation was transient and did not affect RT in subsequent trials. This is consistent with the trial-by-trial coupling between BF bursting amplitude and RT (Figure 3). We believe that this transient influence on RT by BF electrical stimulation is more consistent with a transient increase of the motivational salience associated with the stimulus, and less consistent with an increase in general arousal, which should fluctuate at a much slower time scale but not in single trials.
The current study replicated and extended our previous findings linking BF bursting response to motivational salience. We found that BF bursting neurons not only responded to multiple motivationally salient stimuli in the reward-biased simple RT task (Figure 2), their bursting response also reflected the influence of other behavioral variables on motivational salience. For example, when rats were not required to maintain fixation, BF bursting amplitude in response to the trial-start light signal showed large fluctuation across trials (Figure S10), presumably reflecting the influence of fluctuations in arousal, fatigue, or satiety on motivational salience. This fluctuation of BF bursting amplitude was coupled with response latency on a trial-by-trial basis (Figure S10). Furthermore, in the early sessions of reversal learning when the RT bias had not been updated to reflect expected reward, BF bursting remained tightly coupled with RT modulation (Figure 3D). These data provide further support that BF bursting amplitude reflects motivational salience and is tightly coupled with decision speed. The role of BF motivational salience signal in the learning process is an important question that needs further investigation in future studies.
The widespread spatial distribution of BF bursting neurons is consistent with the location of cortically projecting BF neurons as revealed by placing retrograde tracers in the prefrontal cortex (Figure S2) [28]. The cortically projecting BF neurons are not restricted to a specific subregion in this area, but spread across multiple subregions, including the ventral part of globus pallidus (GP), ventral pallidum (VP), substantia innominata (SI), nucleus basalis of Meynert (NBM, or B), magnocellular preoptic nucleus (MCPO), and horizontal limb of the diagonal band (HDB), but not in the adjacent hypothalamic region lateral preoptic area (LPO) [28]. Furthermore, a recent study [39] shows that individual BF neurons tend to project to multiple subregions in the frontal cortex, unlike single neurons in other neuromodulatory systems, which tend to project to one single subregion in the frontal cortex. This finding suggests that the exact location of BF neurons, as well as the location of BF stimulation electrode, is less critical because the activity of any subset of cortically projecting BF neurons likely provide similar modulation of the entire frontal cortex. Previous studies have shown that the salience-encoding BF neurons represent a physiologically homogeneous group of noncholinergic BF neurons, which, unlike cholinergic BF neurons, do not change their mean firing rates between awake and sleep states [16],[21]. Given that the neurochemical identity of BF bursting neurons remains to be determined, electrical stimulation is the best available technique to ensure the activation of BF bursting neurons for testing causality. One appealing possibility is that salience-encoding BF neurons may correspond to the cortically projecting GABAergic BF neurons, which represent an anatomically prominent projection to the cerebral cortex [28]–[30] and primarily innervate on interneurons in the cortex [40],[41]. The activation of these long-range projecting GABAergic BF neurons should transiently inhibit cortical interneurons, releasing cortical pyramidal neurons from inhibition, leading to an increase in response gain modulation and ultimately resulting in faster decision speed.
The coupling with faster and more precise decision speed demonstrated in this study adds to the functional significance of this anatomically prominent [28]–[30] yet previously neglected neuronal population in the BF [42],[43]. Dissecting the neural circuit-level mechanisms of the BF motivational salience signal will have important translational implications. Dysregulation of motivational salience coupled with decreased decision speed are well documented in schizophrenia [8],[9],[44] and depression [4],[5],[45]. Significant decreases in decision speed also represent a key feature in dementia [6],[7] and cognitive aging [11],[12]. Although the dysregulation of motivational salience has traditionally implicated dopamine neurons and the literature on dementia and cognitive aging has largely focused on cholinergic BF neurons, our study points to a novel and previously neglected candidate mechanism in these conditions. We propose that the decline of decision speed in some of these conditions may result from either the functional impairment of the BF motivational salience system or a disrupted cortical-BF interaction.
All experimental procedures were conducted in accordance with the National Institutes of Health (NIH) Guide for Care and Use of Laboratory Animals and approved by the National Institute on Aging Animal Care and Use Committee.
Twenty-two male Long Evans rats (Charles River, NC), aged 3–6 mo and weighing 300–400 grams at the start of the experiment, were used for this experiment. Sixteen of the 22 rats were trained in the reward-biased simple RT task, with 6/16 of this group undergoing surgery for chronic neuronal activity recording. Seven rats were used in the electrical microstimulaiton experiment, including one of the six rats used for neuronal activity recording, and the other 6/22 rats were used exclusively for the electrical stimulation experiment.
Rats were housed under 12/12 day/night cycle with ad libitum access to rodent chow and water in environmentally controlled conditions. During training and recording procedures, rats were mildly water restricted to their 90% weight and were trained in a daily session of 60–90 min in length, 5 d a week. Rats received 15 min water access at the end of each training day with free access on weekends.
Twelve plexiglass operant chambers (11″L×8 ¼ ″W×13″H), custom-built by Med Associates Inc. (St. Albans, VT), were contained in sound-attenuating cubicles (ENV-018MD) each with an exhaust fan that helped mask external noise. Each chamber was equipped with one photo-beam lickometer reward port (CT-ENV-251L-P) located in the center of the front panel, with its sipper tube 7.5 cm above the grid floor. Two infrared (IR) sensors were positioned to detect reward port entry and sipper tube licking, respectively. Water reward was delivered through a custom-built multibarrel sipper tube. The delivery system was controlled by pressurized air and each solenoid opening (10 ms) was calibrated to deliver a 10 µl drop of water. The reward port was flanked by two nosepoke ports (ENV-114M), located 6.6 cm to each side and 5.9 cm above the grid floor. Each nosepoke port was equipped with one IR sensor to detect port entry. Only the right nosepoke port was used in behavioral training as the fixation port, while the left nosepoke port was inactive.
Each chamber was equipped with two ceiling-mounted speakers (ENV-224BM) to deliver auditory stimuli, and two stimulus lights (ENV-221) positioned above the reward port in the front panel. Behavior training protocols were controlled by Med-PC software (Version IV), which stored all event timestamps at 2 ms resolution and also sent out TTL signals to neurophysiology recording systems to register event timestamps.
After reaching asymptotic behavioral performance, rats were taken off water restriction for at least 3 d before undergoing stereotaxic surgery for chronic electrode implants. Rats were anesthetized with isoflurane (2%–4% isoflurane induction followed with 1%–2% maintenance) and received atropine (0.02–0.05 mg/kg, i.m.) to reduce respiratory secretion. The incision area was shaved and cleaned with betadine, and injected first with local anesthetic (1% mepivacaine HCl solution). Ophthalmic ointment was applied to prevent corneal dehydration. A heating pad was used to maintain body temperature at 37°C. Rats were placed in a stereotaxic frame (David Kopf Instrument, CA) fitted with atraumatic earbars.
Multiple skull screws were inserted first, with one screw over the cerebellum serving as the common electrical reference and a separate screw over the opposite cerebellum hemisphere serving as the ground. A custom-built 32-wire multi-electrode moveable bundle was implanted into bilateral BF. The electrode consisted of two moveable bundles, each containing 16 polyimide-insulated tungsten wires (California Fine Wire, CA) ensheathed in a 28-gauge stainless steel cannula and controlled by a precision microdrive. Eight of the wires in a bundle were 38 µm in diameter and the other eight were 16 µm diameter, with 0.1–0.3 MΩ impedance measured at 1 kHz (niPOD, NeuroNexusTech, MI). The two cannulae of the electrode were precisely positioned to target the BF on both hemispheres at AP –0.6 mm, ML ±2.25 mm relative to Bregma [46]. During surgery, the cannulae were lowered to DV 6–6.3 mm below cortical surface using a micropositioner (Model 2662, David Kopf Instrument), and the electrodes were advanced to 7 mm below the cortical surface. After reaching target depth, the electrode and screws were covered with dental cement (Hygenic Denture Resin). Rats received acetaminophen (300 mg/kg, oral) and topical antibiotics after surgery for pain relief and prevention of infection. Water restriction and behavioral training resumed 7–10 d after surgery. Cannulae and electrode tip locations were verified with cresyl violet staining of histological sections at the end of the experiment and compared with reference anatomical planes [46]. All electrodes were found at expected positions (Figures S2 and S8).
Each recording session lasted 60–90 min. At the conclusion of each recording session, BF electrodes were advanced at least 100 µm and 3–7 d elapsed before the next recording session. One recording session at each electrode depth was included in the final analysis and therefore sampled distinct BF single neuron ensembles. A total of 309 BF single units were recorded from 40 sessions in six rats, at DV 7.1–8.3 mm below the cortical surface.
Electrical signals were referenced to a common skull screw placed over the cerebellum. Electrical signals were filtered (0.03 Hz–7.5 kHz) and amplified using Brighton Omnetics or Cereplex M digital headstages and recorded using a Neural Signal Processor (Blackrock Microsystems, UT). Single unit activity was further filtered (250 Hz–5 kHz) and recorded at 30 kHz. Spike waveforms were sorted offline using OfflineSorter (Plexon Inc, TX). Only single units with clear separation from the noise cluster and with minimal (<0.1%) spike collisions (spikes with less than 1.5 ms interspike interval) were used for further analyses. Additional cross-correlation analysis was used to remove duplicate units recorded simultaneously over multiple electrodes.
Peri-stimulus time histograms (PSTHs) were generated for each BF single neuron against each event using a 10 ms bin. To determine whether each BF neuron significantly responded to sound onset, we subtracted the neuronal response in catch trials from the response in sound-present trials for each neuron. This was necessary because many BF neurons changed their activity during the foreperiod while waiting for sound onset inside the fixation port. This subtraction procedure removed the nonstationary baseline before sound onset (see Figure S5) and allowed us to ask whether BF neurons truly responded to sounds. To determine whether a significant response was present in the subtracted PSTH, we used the method developed by Wiest et al. [47]. Briefly, the statistical significance of PSTHs was determined by comparing cumulative frequency histograms (CFHs) of PSTHs after sound onsets against the cumulative sum distribution of baseline PSTH before sound onsets ({−1, 0}s), estimated based on 1,000 bootstrapped samples (with replacement). The response onset latency was defined as the first bin in which postcue CFH exceeded the cumulative sum distribution from the baseline PSTH (p = 0.01, two-sided). A minimum response amplitude of 0.2 spike (per response) was required to be considered a significant response.
Fifty-nine percent (181/309) of recorded BF neurons showed a significant response within 200 ms of sound onset (Figure S3). Based on our previous studies [16],[21], we focused on BF neurons with a short latency (40–200 ms) excitatory response, which accounted for 162 of the 181 neurons. Furthermore, since the mean firing rate of the 162 neurons were bimodally distributed (Figure S3) and BF neurons encoding motivational salience have firing rates ≤8 spikes/s [16],[21], we included only the 144/162 neurons with firing rate ≤8 spikes/s as BF bursting neurons encoding motivational salience in our analysis. These neurons represented the most prominent neuronal response type in the BF.
Bursting amplitude was calculated as the mean firing rate within the {50,160} ms window after sound onset. There was considerable variability in the bursting amplitude across BF neurons (Figure 3C), even within the same recording session. In order to address the sampling variability and compare across sessions, we used the bursting amplitude ratio between S-Large and S-Small trials, which measures the modulation of BF bursting strength between the two conditions. By comparing this bursting modulation against RT modulation, we quantified how neuronal response modulations correlated with RT modulations between S-Large and S-Small trials. The activity of the entire bursting population (Figure 3D, Figure 4C and D) was calculated by pooling the activity of all BF bursting neurons recorded in a session as a multiunit.
Consistent with a decreasing baseline activity before sound onset, we noted that longer foreperiods produced a stronger activity decrease during the prestimulus period, accompanied by faster RTs (Figure S5). Thus, in our analysis comparing faster and slower trials within the same trial type (Figure 4), it was important to properly control for the influence of foreperiod on neuronal activity and on RT. We therefore first sorted trials associated with each foreperiod and then median split the trials into faster and slower halves (Figure S5). This procedure led to a proper matching of the foreperiod activity between the faster and slower half of trials.
The behavior protocol (Figure 6A) for electrical stimulation was the same as the reward-biased simple RT task (Figure 1A) except that both S-Large and S-Small sounds were replaced by the same 6 kHz 80 dB tone in both trial types, but followed with or without a train of electrical stimulation delivered directly through the BF electrodes. Water reward, on both trial types, was the same (three drops of water). This design allowed us to assess whether the addition of BF electrical stimulation—as a way of augmenting the naturally occurring bursting response to the tone—could lead to faster RT distributions compared to nonstimulated control trials.
The stimulation was delivered through all 32 electrodes in the BF, the same electrode configuration as used in the recording experiment. This was intended to mimic the widespread presence of BF bursting neurons throughout the recording region, representing an ensemble bursting event of the entire population [16],[21].
Individual stimulation pulse was a biphasic charge-balanced pulse (0.1 ms each phase) delivered through a constant current stimulator (stimulus isolator A365R, World Precision Instruments, FL), driven by a Master-8-VP stimulator (A.M.P.I., Israel). Each stimulation train consisted of 11 pulses delivered at 100 Hz (10 ms interstimulus interval) and lasted a total of 100 ms. Stimulation current level was set at 16 µA, 24 µA, 32 µA, or 48 µA per electrode, resulting in a total of 0.5 mA to 1.5 mA over all electrodes. The timing of the stimulation was given at either {60,160} or {80,180} ms posttone onset to coincide with the BF bursting peak. We also implemented two stimulation current paths; one was a unipolar stimulation protocol with currents flowing between all BF electrodes (bilateral BF) against the reference skull screw over the cerebellum. In the second stimulation current path configuration, currents were flowing through all BF electrodes in one hemisphere against all BF electrodes in the other hemisphere.
In total, 44 sessions were tested in seven rats and 15 configurations, with 2–3 current levels tested in each configuration. One rat was first used for recording experiments in the reward-biased simple RT task, while the other six rats were never trained in the reward-biased simple RT task prior to the stimulation protocol. In Figure 6D, each gray line represents data collected from one rat, under one specific combination of current path and timing window. Linear mixed models were used to handle an unequal number of within-configuration observations in order to determine the influence of microstimulation on RT and RT modulation between stimulated and nonstimulated trials. The electrical stimulation current level was modeled as a continuous variable and its influence on RT modeled as a constant slope fixed effect. The choice of either timing windows and the choice of either current path configurations did not modulate RT differently (linear mixed models estimating the fixed effect of timing window on RT modulation, F(1,39) = 0.899, p = 0.349; and the fixed effect of current paths on RT modulation, F(1,39) = 0.303, p = 0.585; interaction term F(1,39) = 0.498, p = 0.485). Therefore, these variables were not included in the final model, which only tested whether the fixed effect of stimulation current level on RT was significantly different from zero. Electrical stimulation significantly decreased RT in stimulated trials as a function of stimulation current level (F(1,42) = 17.856, p<0.001; Figure 6C), but had no influence on RT in nonstimulated trials (F(1,42) = 0.235, p = 0.630; Figure 6C). Stimulation also significantly increased RT modulation (ratio of mean RT between nonstimulated and stimulated trials) as a function of stimulation current level (F(1,42) = 18.922, p<0.001; Figure 6D).
|
10.1371/journal.pgen.1003929 | Recurrent Tissue-Specific mtDNA Mutations Are Common in Humans | Mitochondrial DNA (mtDNA) variation can affect phenotypic variation; therefore, knowing its distribution within and among individuals is of importance to understanding many human diseases. Intra-individual mtDNA variation (heteroplasmy) has been generally assumed to be random. We used massively parallel sequencing to assess heteroplasmy across ten tissues and demonstrate that in unrelated individuals there are tissue-specific, recurrent mutations. Certain tissues, notably kidney, liver and skeletal muscle, displayed the identical recurrent mutations that were undetectable in other tissues in the same individuals. Using RFLP analyses we validated one of the tissue-specific mutations in the two sequenced individuals and replicated the patterns in two additional individuals. These recurrent mutations all occur within or in very close proximity to sites that regulate mtDNA replication, strongly implying that these variations alter the replication dynamics of the mutated mtDNA genome. These recurrent variants are all independent of each other and do not occur in the mtDNA coding regions. The most parsimonious explanation of the data is that these frequently repeated mutations experience tissue-specific positive selection, probably through replication advantage.
| DNA mutations are expected to be formed randomly, thus any reproducible pattern of DNA somatic mutations across multiple individuals or even across organs within each individual is highly unexpected. Using next generation sequencing of multiple tissues from the same individuals we found several somatic mutations in mitochondrial DNA that appear in a heteroplasmic state in all individuals examined, but only in particular tissues. These mutations were only found in known regions of replication control for the mitochondrial DNA. These data imply the presence of tissue-specific positive selection for these variants.
| Mitochondrial DNA (mtDNA) heteroplasmy is commonly thought to be the product of either maternal inheritance [1] or rare, random somatic mutations that undergo subsequent expansion within an individual via genetic drift [2], [3]. Inherited heteroplasmy should be present in many, but perhaps not all, tissues while somatic mutations spread only as a result of cell division subsequent to the mutation event. Somatic mutations will be restricted to those cells or tissues derived from a common progenitor, and should follow patterns of development. Such intra-individual patterns should therefore differ among individuals as a function of where and when the initial mutation occurred.
Under these standard models, both inherited and somatic heteroplasmies should differ among unrelated individuals. However, until recent advances in sequencing technology it was impossible to assay low levels of heteroplasmy across the entire mitochondrial genome. We applied massively parallel sequencing technology to test these models by deeply sequencing the same ten tissues from two unrelated individuals. Unexpectedly we found that certain tissues, notably kidney, liver and skeletal muscle, have recurrent mtDNA mutations that were undetectable in other tissues in the same individuals. These mutations were found across unrelated individuals in these same tissues. Neither the maternal inheritance nor the random somatic mutation models explain the observed patterns of recurrent mtDNA heteroplasmy. The common recurrence of these tissue-specific mutations indicates a completely different model of mtDNA heteroplasmy, namely a decidedly non-random process that results in particular mutations, but only in specific tissues.
Next generation sequencing provides several advantages over previous methods in that it allows detection of very low heteroplasmy levels across the entire mtDNA genome without having to target specific sites. We sequenced mtDNA from 10 tissues (kidney, lung, liver, small bowel, large bowel, skeletal muscle, spleen, brain white matter, skin above belt and skin below belt) obtained at autopsy from two cancer-free individuals (Text S1). In brief, mtDNA sequences were generated as 100 nt paired-end reads on Illumina HiSeq 2000 machines, were aligned to the human reference genome rCRS [4] using BWA [5], and then were locally realigned and recalibrated using GATK [6]. Variants were reported as heteroplasmic if their frequencies were ≥1% on both strands from reads with a mapping quality score ≥30. We further eliminated variants with any of the following artifacts: strand bias, low average base quality score, or clustering at read ends. Strand bias was evaluated at both base pair and motif levels. We identified 20 heteroplasmic sites with mutation levels >1% in our two subjects (Table 1).
The coverage of mtDNA sequencing varied across tissues, ranging from over 7,000 to almost 80,000 (Figure 1A). By comparing mtDNA coverage to that for autosomal chromosomes, we estimated the mtDNA copy number per cell for each sample (Figure 1B). The two subjects had similar estimated mtDNA copy numbers in each tissue with values consistent with expectations based on previous data, ranging from a few hundred mtDNA per cell in spleen to a few thousand in the skeletal muscle [7], liver, and kidney.
Both subjects died of myocardial infarction and had no evidence of cancer or occult cancer. Subject 1 was a male, age 57 years, while subject 2 was a female, age 71 years (Table 2). This age and gender difference may explain the difference in mtDNA copy number in skeletal muscle between the two subjects.
Unexpectedly, we found eight sites to be heteroplasmic in the same tissues in both of our subjects (Figure 2). Contamination was unlikely since the pattern of these sites fit no known haplogroup, and the samples from the two subjects were collected at different times and sequenced at different facilities (see Materials and Methods). Most of these variants exist in the general population but are rare (Table S2) [8]. Some of these variants have been previously reported as being heteroplasmic in these same tissues but the recurrence and tissue-specificity of these mtDNA variations was not discussed [9]. In the four unrelated individuals combined from our study and that of He et al [9] there were 10 recurrent mutations. All of the recurrent mutations lay within the mtDNA control region, and nine of the ten recurrent mutations occurred in multiple individuals but only in specific tissues (site 16093 was the only exception and this mutation was found in a wide range of tissues). Six of the recurrent mutation sites were observed in both studies (Figure 2). Surprisingly, these recurrent tissue-specific mutations are all close to regulatory sites for mtDNA replication, indicating that these variations are likely to alter the replication dynamics of the mutated mtDNA molecules.
Three of the ten tissues studied (liver, kidney and skeletal muscle) harbored multiple mtDNA mutations that were shared across two or more of the four individuals combined from both studies (Figure 2). Mutations at sites 60 and 72 occurred in liver and kidney in both studies, while mutations at sites 94 and 203 were repeatedly detected in liver and/or kidney only in our subjects. Of these mutations, three sites (60, 72 and 94) when present in an individual, always occurred in both liver and kidney. Another three occurred in skeletal muscle of all four patients (sites 64, 189 and 408). Heteroplasmy at positions 189 and 408 has been found in skeletal muscle of the elderly [10]. A mutation at site 67 was found only in skeletal muscle in our two subjects. There was no observable linkage disequilibrium among the sites (all pairwise r2<0.007), indicating that these are independent mutations and are not due to contamination. Heteroplasmy for these tissue-specific recurrent mutations ranged from 1–21%, but was not detected in the other tissues. Strikingly, the level of heteroplasmy was similar across individuals in the same tissues (Figure 2). The depth of sequence coverage in these tissues (skeletal muscle, liver and kidney) ranged from 27,000–76,000× (Figure 1A), providing high confidence in our observation of heteroplasmy in these samples. Of the tissues we examined (with the notable exception of brain white matter), skeletal muscle, liver and kidney are the ones most often affected by mitochondrial disease [11].
Although the entire mtDNA genome was sequenced to high depth, we found no single base pair substitutions outside the control region that were repeated between individuals. The common 4977 deletion [12], [13], [14] was found in many tissues as expected. Here we have focused only on those heteroplasmic sites common to multiple subjects; the full list of identified heteroplasmic sites is given in Table 1. These data clearly indicate a non-random distribution (6.0e−4<p<6.0e−6) of recurrent heteroplasmic mutations that flank important regulatory elements for mtDNA replication (Figure 2). Several of the repeated variants were clustered (sites 60–72) near a recently reported origin of replication for the H strand [15], [16] at sites 54–57. Three of the other repeated mutations (189, 203 and 16093) occur very near the boundaries of the displacement loop [15]. Finally, site 408, which was heteroplasmic in the skeletal muscle of all four individuals, lies within the L strand promoter that initiates the RNA primer for mtDNA replication.
One variant, 16093, was unusual in that it was observed in all tested tissues in one subject from each study, with tissue-specific heteroplasmy levels that were strikingly similar across these two individuals (Figure 3; r = 0.93, p<0.003). These subjects were of similar age (59 and 57 years), so it is possible that the 16093 heteroplasmy increases with age at different rates in different tissues, leading to the similar heteroplasmy levels across tissues in these subjects. The 16093 site lies within a loop of a predicted large secondary structure of the mtDNA and is known to be hyper-mutable [17]. Individuals with the 16093C variant in blood tend to have C/T heteroplasmy in buccal cells [18]. In one of our subjects and in patient 11 from He et al. [9], the 16093C variant is the major allele in most tissues (except muscle), and these are the same two individuals who have widespread 16093T/C heteroplasmy across all tested tissues (Figure 2). The similar, tissue-dependent heteroplasmy levels at 16093 reinforce the observation that heteroplasmy levels at other sites are also comparable across individuals (Figures 2 and 3).
We also found two insertion/deletion (indel) somatic mutations that were repeated across the two sequenced subjects in a tissue-specific pattern. Both are length variations in polynucleotide tracts. The human mtDNA reference sequence (rCRS) has a stretch of six guanines, denoted by G6, from sites 66–71. Both of our subjects had measurable heteroplasmy for the G5 variant, decreasing the length of this poly-G tract by one nucleotide. This variant was found in the same four tissues in both subjects: kidney (0.9% and 3.6% in subjects 1 and 2 respectively), large bowel (0.7% and 2.4%), small bowel (1.0% and 4.3%) and the white matter of the brain (1.9% and 3.6%). This poly-G tract is located adjacent to one of the recurrent heteroplasmic SNPs (Figure 2). In kidney, the G5 variant did not occur on the same reads as the heteroplasmic variant at site 72, demonstrating that this is not a sequencing artifact and that the variants are on different mtDNA molecules. Considering its location, it is reasonable to hypothesize that this poly-G length variant also affects mtDNA replication.
The second repeated heteroplasmic indel was in an 8-nucleotide poly-A tract at positions 12418–12425. This is the longest poly-A tract in the rCRS. In both subjects, we found the shorter A7 variant in the kidney samples only (1.0% heteroplasmy in subject 1 and 1.6% in subject 2). We also found the longer A9 variant in both kidney samples, but at very low levels (<1%). This indel is the only repeated mutation that we found in the coding region of the mitochondrial genome. It is located near the start of the MT-ND5 gene (12337–14148) and causes a frameshift mutation, severely altering almost the entire length of the ND5 protein, an essential component of complex I of the electron transfer chain.
To confirm that the observed heteroplasmy was not due to sequencing artifact, we performed an alternative analysis for site G94A because it could be assayed using RFLP analysis. Specifically, the G allele at G94A permits digestion with the restriction enzyme BcoDI. Sensitivity of this assay was determined using a titration of plasmid constructs with and without the restriction site. This assay was specific to mtDNA because the primers did not amplify nuclear DNA (Figure S1A). We detected as low as 2.5% of the undigested variant (Figure S1B). RFLP analysis of all samples from the two sequenced individuals demonstrated that the tissues shown to be heteroplasmic from the sequencing analyses had both digested and undigested bands and were therefore heteroplasmic (Figure 4). As a negative control, we also analyzed this site in spleen DNA and found it to be consistent with the sequencing results (Figure 4). This result was statistically significant (p<0.001).
To test the generality of this heteroplasmy we examined two additional cancer-free individuals (Subjects 3 and 4 in Table 2) using the same RFLP assay. This analysis replicated heteroplasmy in liver in both additional subjects and in kidney in one individual. No significant heteroplasmy was detected in spleen in either subject, providing additional support for the tissue-specificity of this SNP (Figure 4). The result was also statistically significant (p<0.001).
Several of our observed heteroplasmic sites have been identified in studies of human disease. The T408A mutation, which was present in the muscle of all four sequenced individuals (two from the present study and two from He et al.), has been reported as an age-related somatic mutation in muscle [10], [19], [20], [21]. It has also been associated with disease in an investigation of a patient with a mitochondrial depletion syndrome [22] that was fatal at a young age (14 years), where the T408A mutation exhibited heteroplasmy at high levels (>70%) in all investigated maternal relatives, but was not detectible in the patient. The authors speculated that the T408A mutation interacted with a hypothesized nuclear DNA factor that affected mtDNA replication, thus leading to the mtDNA depletion in this patient. The A189G mutation, which was also present in the muscle of all four sequenced individuals, has also been reported in studies of aging muscle [10], [19], [20], [21], [23], [24], [25], and is often reported together with T408A. Both of these mutations increase in heteroplasmy level in muscle slowly with increasing age [19].
Heteroplasmy at site 16093 has often been reported in a range of tissues [18], [26], [27], [28], [29], [30], [31], [32], consistent with the observation of this variant being in all tissues in one of our two sequenced subjects and one of the subjects from He et al. The T414G mutation has been reported to accumulate with age in fibroblasts and skeletal muscle [33], [34] and we detected this mutation in one individual in a skin sample (Table 1). Our observation of heteroplasmy at 189, 408, 414, and 16093 and the previous studies reporting these same variants provide support for the validity of the next-generation sequencing data.
We found that four heteroplasmic somatic mutations (T60C, T72C, G94A, and G203A) recur, but only in liver and/or kidney. Given that liver and kidney arise from endoderm and mesoderm respectively, it is unlikely these mutations share a common developmental origin. These four sites are also global population polymorphisms, though at low frequencies [8] (Table S2). Some of these have been previously reported in the context of human disease. The T72C variant has been reported as a somatic mutation in the brain tissue of both Alzheimer's cases and controls [35], although it was not detected in any sequenced brain samples in this study. G94A has recently been reported in two Chinese pedigrees transmitting Lebers Hereditary Optic Neuropathy but in these families this variant was an inherited fixed polymorphism, not a heteroplasmic somatic mutation [36]. Despite the rarity of G203A in the global population (estimated as 0.3% in a survey of human mtDNA sequences deposited in GenBank) [8] it has been identified as a fixed variant in patients with deafness in two independent studies in different ethnicities [37], [38].
Our results indicate that mtDNA heteroplasmy due to somatic mutation is unexpectedly recurrent and tissue specific. By using a sensitive deep-sequencing technique across a wide range of tissues in multiple subjects we were able to test the hypothesis that specific mtDNA variations preferentially accumulate in particular tissues [34], [39]. One possible explanation for observing the same mtDNA heteroplasmic variants in two or more tissue types within the same person is that a mutation occurred early in embryonic development, before the tissues differentiated from their common progenitor. However, this hypothesis cannot explain the repeated observation of the same mutations in the same tissues in unrelated individuals. The occurrence of repeated mutations in the same tissues at sites that closely correspond to regulatory elements for mtDNA replication indicates somatic selection as the most likely mechanism driving the increase and maintenance of these heteroplasmic mutations. This inference is further supported by independent evidence that liver and kidney exhibit positive selection for mtDNA variants in a mouse model formed by artificially mixing the mtDNA of two different mouse strains [40], [41], [42]. In-vivo BrdU labeling in these mice over a time course of 50 hours did not detect any difference in labeling in liver samples between the two mtDNA haplotypes [41], leading the authors to conclude that replicative advantage was not the driving force for the segregation in these mice. However, we would argue that a 50 hour window is not comparable to the decades of replication advantage that would need to occur in our subjects. Recently, Sharpley et al [43] also generated a separate mouse model of heteroplasmy by mixing the naturally occurring NZB and 129S6 mtDNA sequences. This mouse model also showed that the segregation of the two mtDNA genomes varied in a tissue-specific manner, with liver and kidney having the strongest selection for the NZB version of the genome. Finally, liver has been argued to be under selection for nuclear aneuploidy and polyploidy, indicating that selection may have a special role in this tissue [44]. It is reasonable that the tissue specific selection of these mtDNA variants is due to regulation by nuclear-encoded mitochondrial genes with tissue-dependent expression, as has been shown in one of the mouse models in spleen [40]. In contrast to previous work documenting a wide variety of heteroplasmic sites and their functional implications, the unique value of this study is the comparison of mtDNA heteroplasmy across multiple tissues in several individuals and the demonstration that several somatic variants recur in a tissue-specific pattern.
The pattern of tissue-specific mutations we have found across multiple individuals could be explained by a few alternatives, including positive or negative selection. Under positive selection, mutations in certain tissues would increase in frequency due to their advantage. Under negative selection, mutations could occur at a high rate but would be removed from all tissues except for those where the recurrence is observed, where negative selection is presumably relaxed. Of these two alternatives, positive selection is the most likely explanation because under negative selection mutations should be scattered widely across the mtDNA control region, not just the recurrent ones at specific sites related to replication (Figure 2). In contrast, positive selection could simply be explained by a replication or other functional advantage in high copy number tissues due to increased mtDNA replication. It is important to note than any functional difference among the variant mtDNA molecules, if any, are due only to the sites we describe because all of our observed heteroplasmic sites are independent of each other, and there were no recurrent heteroplasmic sites in the coding regions.
Another alternative is that there are tissue-specific mutational hotspots within the mtDNA. For example, interferon-induced cytidine deaminases are capable of generating somatic mtDNA mutation in a tissue-specific fashion [45]. Although this alternative is not mutually exclusive to the selection argument, we still favor differential selection based on its simplicity, and on prior data suggesting that two of the tissues in which we observed recurrent mtDNA mutations, liver and kidney, also undergo selection in two separate heteroplasmic mouse models [40], [41], [42], [43]. The mouse models provide evidence suggesting that selection in the absence of any de-novo mutation generation can cause tissue specific heteroplasmy patterns because in both mouse models the two mtDNA haplotypes were artificially introduced through cell fusion and were not generated via a mutation process. Furthermore, even if the heteroplasmic sites we observed are mutational hotspots, their locations in the mtDNA genome are highly suggestive of functional roles (Figure 2). Homoplasmy in tumors has also been shown to possibly derive from random processes alone, but this computer simulation study is not directly analogous to recurrent heteroplasmy in normal tissues of multiple individuals [46]. Therefore, the most parsimonious and reasonable explanation for our data is positive selection in liver, kidney and skeletal muscle for certain mutations in and around the regions controlling mtDNA replication.
In a very different model system (i.e. a mouse strain with an abnormally high mutation rates due to a defective mtDNA polymerase) evidence for lower mutation load in the D loop was described [47]. Specifically, in this mouse model the accumulation of point mutations in the mtDNA was lower in the D loop region than the rest of the mtDNA. It is impossible to determine conclusively from these data whether this pattern is due to selection or to a variation in mutation rate, but it does demonstrate a non-random pattern in this part of the mtDNA, something we also observed but in a different direction.
Our data provide strong support for the conclusion that the current models of mtDNA variation are inadequate to explain what we now call “recurrent heteroplasmy”. The pattern of common, recurrent mutations we observed provides strong evidence that mtDNA heteroplasmy at several sites is non-random and is most likely the result of tissue-specific positive selection acting on the replication of mtDNA. The restriction of these mutations to liver, kidney and skeletal muscle indicates that the mtDNA replication process may vary across tissues, leading to tissue-specific selective forces, which correspond with high copy number tissues.
The protocols were approved by the Vanderbilt University Institutional Review Board. Samples were collected at autopsy within 48 hours post-mortem by a trained pathologist (RDH). Tissue samples for DNA extraction were collected using clean and sterile scalpels, placed in petri dishes, and transferred to 50 ml tubes containing ice-cold Dulbecco's phosphate-buffered saline (DPBS), rinsed again with DPBS and stored at −80°C until DNA extraction with exceptions as described below. Separate portions of the tissue sections were preserved in 10% formalin. Skin samples were collected from the ventral torso, from both above-belt (Skin-AB) and below-belt (Skin-BB) (e.g. above or below the waistline). Skeletal muscle was obtained from the diaphragm. The small and large bowel samples consisted of mucosal tissue that was collected by carefully scraping the loose mucosal layer from the internal surface of bowel sections. Bone marrow tissue was collected by flushing rib or vertebral body sections with DPBS and collecting the flushed material in a 50 ml tube on ice, which was then centrifuged to collect the cellular material. Splenocytes were isolated as previously described [48]. Gray and white brain samples were separated manually in a petri dish using a sterile scalpel. Demographic information for the four subjects is in Table 2.
For each tissue two DNA extractions were performed. The tissue was lysed using the DNeasy Blood and Tissue Kit (Qiagen 69504). Once the tissue was lysed and incubated at 56°C overnight, one set of DNA extractions was transferred into a 2.0 ml tube (Sarstedt 72.694.406) and put on the QIAsymphony for automated extraction (Qiagen). The protocols used on the QIAsymphony were Tissue_LC_200_V5 and Tissue_HC_200_V5 depending on the tissue type. The second set of DNA extractions was transferred to Autopure Qubes D (Qiagen 949022) and 3 ml of Cell Lysis Solution (Qiagen 949006) was added to each tissue sample. These samples were then placed on the Autopure LS (Qiagen) for automated extraction. The protocol used on the Autopure was Cell Lysate. The resulting DNA was stored in Nunc Cryotubes (Nunc 377267). DNA from each protocol was calibrated and samples combined prior to sequencing.
Sequences were generated as 100 nt paired-end reads on Illumina HiSeq 2000 machines. The two subjects were sequenced at different locations (subject 1 at Macrogen in South Korea and subject 2 at Illumina in California). Each sample was sequenced on 3–5 lanes, yielding 1.14–1.99 billion reads, which were aligned to the human reference genome hg19+rCRS (revised Cambridge Reference Sequence, NC_012920.1) using BWA [5] (ver. 0.5.9-r16). We performed local realignment and base quality score recalibration using GATK [6] (ver. 1.0.5974). The number of mapped reads ranged from 1.09–1.84 billion, with more than 90% of all reads being mapped except skeletal muscle from subject 2 (86.2%) (Table S1). Other programs were also used at various steps: samtools (ver. 0.1.16) for sorting and indexing bam files, bamtools (ver. 0.8.1025) for splitting and merging bam files, and picard (version 1.48) for marking duplicates and fixing mates after local realignment.
All samples were also genotyped on the Illumina Human Omni1 Quad chip with approximately one million SNPs in the nuclear genome. The consistency rate between the sequence- and chip-based SNP calls was >99.86% for all samples after standard quality control filtering. This indicates that the sequencing data were of high quality.
We screened for heteroplasmy in mtDNA using reads with MAPQ≥30. For each site, we calculated the fraction of bases A, C, G, T on the forward and reverse strands. A site was called heteroplasmic if it had ≥1% frequency for two or more bases on both strands and the variant did not have any of the following alignment artifacts: 1) strand bias, 2) clustering at read ends, and 3) low average base quality score. Due to the high read depth (Figure 1, Table S1) all duplicate reads were retained. Heteroplasmy estimates were assessed with and without the duplicate reads and heteroplasmy levels were not influenced by the duplicates.
The linear mtDNA reference genome (rCRS) was created by cutting the circular mtDNA at a fixed position. This may generate alignment artifacts near the linearization site (i.e., the ends of the mtDNA reference): 1) a read overlapping the linearization site may be unmapped or require heavy clipping to be aligned, 2) a read may be aligned but its paired-end mate may not be, and 3) a read may be aligned to one end of the reference but its mate to the other end. As a result, reads close to the linearization site may have low mapping quality scores and may be disproportionately filtered out. As our data are 100 nt reads with insert sizes mostly between 250 bp and 400 bp (Figure S2), these artifacts may influence the results hundreds of bases away from the mtDNA “ends”. To prevent artifacts due to mtDNA circularity we also created a new mtDNA reference genome by shifting the rCRS starting point to position 7002, and repeated the whole data processing steps as described above. Heteroplasmy was virtually identical between the two alignments, with less than 0.1% difference in heteroplasmy estimates.
In addition, the artificial N base at 3107 of the rCRS reference can lead to alignment artifacts. This N was removed before the alignment was made.
For all observed heteroplasmic sites, we checked for various sequencing artifacts. The mtDNA control region harbors multiple poly-nucleotide tracts that could lead to sequencing artifacts. Since these artifacts often have strand bias, we filtered out all sites with strand bias. In addition, for each heteroplasmic site we performed a motif analysis similar to that described in detail for site 310 in the supplementary material, to identify artifacts (Table S4). We also checked for the presence of artifacts due to sequence similarities between the nuclear and mtDNA genomes (NUMT), and none could be detected (see Text S1). The reported sites are free of any artifacts.
Alignment errors are known to cause artifacts that often show strand bias and excessive occurrence of mutant alleles at read ends. We found no cycle (i.e. position on the read) or strand bias for the mutant alleles at the heteroplasmic sites we identified. We extracted all bases with mapping quality score ≥30 and base quality score ≥20 at each heteroplasmic site and compared the distributions of cycle and strand for the major and mutant alleles. Figure S3 shows the distributions of C and A alleles for site 64 in the skeletal muscle of Subject 1. The mutant allele was uniformly distributed across the read length on both strands, showing no strand or cycle bias. Other heteroplasmic sites we identified had similar patterns.
For heteroplasmic sites close enough to be on the same DNA read or read pair, we assessed whether the minor alleles are on the same haplotype background, or in other words, if they are in linkage disequilibrium (LD). LD between the variants could be a sign of either contamination or sequencing artifacts. Specifically, for every pair of sites ≤100 bp apart (e.g., 60-72-94 in liver and kidney tissues and 64–67 in skeletal muscles), we extracted reads that covered both positions and had MAPQ≥30. We further required that the reads had all bases matched (i.e., CIGAR string “100M”) or had clipping at one end (i.e., CIGAR string matching the regular expression pattern “[0–9]*S[0–9]*M” or “[0–9]*M[0–9]*S”), and the bases at the two sites had base quality score ≥20. We then tallied haplotypes and calculated r2 between the two sites. All r2 values were very close to zero (<0.004), indicating no LD between any heteroplasmic sites.
For every pair of sites >100 bp apart (e.g. between 60-72-94 and 203 in liver tissues; among 64–67, 189, and 408 in skeletal muscles), we extracted read pairs that covered both positions and had MAPQ≥30 and then followed the above procedure. Again, all r2 values were very close to zero (<0.007), indicating no LD in the mtDNA.
The sequencing error rate is reflected in the recalibrated base quality scores. For example, a base quality (BQ) score of 25 means the error rate for that base is 0.32%, BQ = 27 means 0.2%, and BQ = 30 means 0.1%. These error rates are much lower than the 1% detection cutoff we used for the determination of heteroplasmic sites. Figure S4 shows the distribution of recalibrated base quality scores. For all our samples, 82.4% bases had recalibrated base quality score ≥30, 91.4% had scores ≥27, and 94.4% had scores ≥25. These results provide assurance that the bases we used for our inferences had high quality and an error rate much lower than our heteroplasmy detection threshold.
We calculated the depth of coverage for autosomes and mtDNA as:The multiplier 100 was used because we had 100 nt reads. The mtDNA depth ranged from 5651×–119203×, and the autosome depth ranged 37×–61× (Table S1). Assuming each cell carries a diploid (2×) nuclear genome, the mtDNA copy number was estimated as:The estimated mtDNA copy number ranged from 315 to 5880 (Table S1).
To test for non-randomness of the recurrent mutations, we calculated the probability that a mutation occurred at these 10 sites under two extreme scenarios. Using a model of constant mutation rate, c, along the whole mtDNA genome, the probability for a mutation to occur anywhere would be 16569c and the probability for it to occur at these 10 sites would be 10c. Thus the probability for an observed mutation in a DNA sample to occur only at any of these 10 sites is 10/16569 = 6.0e−4. Now suppose a mutation has been observed in a specific tissue of an individual. The probability to observe the same mutation in another individual only in the same tissue (out of 10 tissues) and on the same site (out of 10 sites) is further reduced to 6.0e−6. The recurrence patterns of the reported mutations fall between these two extreme scenarios, and therefore their p-values are between 6.0e−4 and 6.0e−6.
To test for correlation of 16093 heteroplasmy levels between the two individuals (Figure 3), we calculated the Pearson correlation coefficient and its associated p-value. The correlation was 0.93 and the p-value was 0.0028.
We calculated the p-value to evaluate the significance of our RFLP validation and replication results. For the validation part, we performed RFLP on six tissues (kidney, liver, spleen from Subjects 1–2). Let a = P(detect 94A|94A is absent), the probability of falsely detecting 94A in RFLP analysis while it was absent. Then the probability of seeing 94A in two kidneys, two livers but not the two spleens is a4(1−a)2. The value of a is probably lower than the false positive rate for sequencing analysis, which would be at most 0.2 (4 out of 20 tissues when the sequencing results were assumed to be false). Even at a = 0.2, the p-value will be a4(1−a)2 = 0.001. The p-value will be much smaller at a lower value of a; for example, p = 5.6e−6 if a = 0.05, and p = 9.8e−9 if a = 0.01. The p-value for the replication part can be similarly evaluated.
We determined the mtDNA haplogroups for our subjects: T2a1 for subject 1 and H1a1 for subject 2 (Table S3). Haplogrouping was performed using the H-Mito program (http://www.phylotree.org) supplied by Mannis van Oven [49].
To provide molecular validation of sequencing results, we performed RFLP analysis using control plasmids and patient DNA from suspected heteroplasmic and homoplasmic tissues. We focused on position 94, which sequencing results identified as heteroplasmic in kidney and liver. Control samples consisted of 100% wild-type plasmids at position 94 (G), 100% mutant plasmids (A), or 97.5% wild-type and 2.5% mutant plasmids. DNA samples from suspected heteroplasmic kidney and liver tissues, and from suspected homoplasmic spleen tissue were analyzed for subjects 1–4. Fifteen nanograms of the control plasmid or patient DNA was amplified with 10 µM D-loop-targeted forward (5′-GATCACAGGTCTATCACCCTATTAAC-3′) and reverse (5′-CAGATACTGCGACATAGGGTGCT-3′) primers (Operon) and Platinum PCR Supermix (Invitrogen) according to manufacturer's directions. Following amplification, PCR products were digested for 8 h at 37°C with the restriction enzyme BcoDI, which cuts the wild-type but not mutant PCR product at position 94. Successful digestion resulted in cutting of the 130-bp PCR product into 90- and 40-bp fragments. We added 5 µL 5× gel-loading dye (KD Medical/MediaTech) to each 20-uL reaction after restriction digest, and loaded 12 µL of digest products and gel-loading dye into each well of a 3% agarose (Sigma) gel in 1× TBE (Cellgro) with 0.0125% ethidium bromide (Bio-Rad). The gel was run at 150 V for 2 h, then UV imaged for 200 ms in a Syngene G:Box imager.
|
10.1371/journal.ppat.1004501 | Aminoterminal Amphipathic α-Helix AH1 of Hepatitis C Virus Nonstructural Protein 4B Possesses a Dual Role in RNA Replication and Virus Production | Nonstructural protein 4B (NS4B) is a key organizer of hepatitis C virus (HCV) replication complex formation. In concert with other nonstructural proteins, it induces a specific membrane rearrangement, designated as membranous web, which serves as a scaffold for the HCV replicase. The N-terminal part of NS4B comprises a predicted and a structurally resolved amphipathic α-helix, designated as AH1 and AH2, respectively. Here, we report a detailed structure-function analysis of NS4B AH1. Circular dichroism and nuclear magnetic resonance structural analyses revealed that AH1 folds into an amphipathic α-helix extending from NS4B amino acid 4 to 32, with positively charged residues flanking the helix. These residues are conserved among hepaciviruses. Mutagenesis and selection of pseudorevertants revealed an important role of these residues in RNA replication by affecting the biogenesis of double-membrane vesicles making up the membranous web. Moreover, alanine substitution of conserved acidic residues on the hydrophilic side of the helix reduced infectivity without significantly affecting RNA replication, indicating that AH1 is also involved in virus production. Selective membrane permeabilization and immunofluorescence microscopy analyses of a functional replicon harboring an epitope tag between NS4B AH1 and AH2 revealed a dual membrane topology of the N-terminal part of NS4B during HCV RNA replication. Luminal translocation was unaffected by the mutations introduced into AH1, but was abrogated by mutations introduced into AH2. In conclusion, our study reports the three-dimensional structure of AH1 from HCV NS4B, and highlights the importance of positively charged amino acid residues flanking this amphipathic α-helix in membranous web formation and RNA replication. In addition, we demonstrate that AH1 possesses a dual role in RNA replication and virus production, potentially governed by different topologies of the N-terminal part of NS4B.
| With an estimated 180 million chronically infected individuals, hepatitis C virus (HCV) is a leading cause of chronic hepatitis, liver cirrhosis and hepatocellular carcinoma worldwide. HCV is a positive-strand RNA virus that builds its replication complex on rearranged intracellular membranes, designated as membranous web. HCV nonstructural protein 4B (NS4B) is a key organizer of HCV membranous web and replication complex formation. Here, we provide a detailed structure-function analysis of an N-terminal amphipathic α-helix of NS4B, named AH1, and demonstrate that it plays key roles in shaping the membranous web as well as in virus production. We also show that the N-terminal part of NS4B adopts a dual membrane topology in a replicative context, possibly reflecting the different roles of this protein in the viral life cycle.
| Hepatitis C virus (HCV) infection is a leading cause of chronic hepatitis, liver cirrhosis and hepatocellular carcinoma worldwide, with a peak of the disease burden expected in around 10 years from now [1]. HCV and GB virus B have been classified in the Hepacivirus genus within the Flaviviridae family, which also includes the genera Flavivirus, Pestivirus and Pegivirus [2]. Additional closely related viruses have been identified recently in horses as well as other animal species, and have been classified in the Hepacivirus and Pegivirus genera, including nonprimate hepaciviruses (NPHV) [3], [4].
HCV contains a 9.6-kb positive-strand RNA genome encoding a polyprotein precursor that is co- and posttranslationally processed into ten structural and nonstructural proteins [2], [5]. As all positive-strand RNA viruses, HCV replicates its genome in a membrane-associated replication complex composed of viral proteins, replicating RNA, rearranged intracellular membranes and additional host factors [6], [7], [8], [9]. The specific membrane alteration induced during HCV RNA replication has been designated as membranous web [10], [11]. Nonstructural proteins 3 through 5B are essential for HCV RNA replication, and their functional complex is referred to as replicase.
Nonstructural protein 4B (NS4B) is the least characterized HCV protein. However, evidence from biochemical, structural and genetic studies as well as electron microscopy (EM) indicates that NS4B is a key organizer of HCV replication complex formation (reviewed in [12]). Indeed, NS4B has been shown to induce formation of the membranous web which serves as a scaffold for the viral replicase [10], [11]. More recent work has shown that the other nonstructural proteins, especially NS5A, contribute to the formation of double membrane vesicles (DMVs) which make up the membranous web [13] and are believed to represent sites of HCV RNA replication [14].
NS4B is a 27-kDa integral membrane protein comprising an N-terminal part (amino acids [aa] 1 to ∼69), a central part harboring four predicted transmembrane segments (aa ∼70 to ∼190), and a C-terminal part (aa ∼191 to 261). The N-terminal part comprises a predicted and a structurally resolved amphipathic α-helix, designated as AH1 and AH2, respectively. AH2 comprises aa 42–66 and has been shown to play an important role in HCV RNA replication [15]. Intriguingly, it has the potential to traverse the phospholipid bilayer as a transmembrane segment, likely upon oligomerization [15], [16], [17], [18]. Hence, the N-terminal part of NS4B may adopt a dual cytosolic and ER luminal topology. However, this has not been explored in a functional, replicative context.
AH1 was predicted as an amphipathic α-helix and reported to mediate membrane association and HCV RNA replication [19]. However, membrane association of AH1 is debated [12] and the actual structure as well as detailed functional analyses, covering the complete HCV life cycle, have not been reported.
Here, we describe the three-dimensional structure of AH1 and provide a detailed structure-function analysis, indicating that this structurally highly conserved segment of NS4B possesses a dual role in HCV RNA replication and virus production. In addition, we demonstrate that the N-terminal part of NS4B adopts a dual membrane topology in an authentic replication context, allowing to speculate that the different functions of NS4B may be governed by different topologies.
Sequence analyses and structure predictions were performed to assess the degree of conservation of the N-terminal part of NS4B and to identify potential structural determinants. The degree of aa conservation among different genotypes was investigated by ClustalW alignment of 27 reference sequences representative of the major HCV genotypes and subtypes. This alignment revealed that the segment comprising aa 50–70, including part of amphipathic α-helix AH2 (aa 42–69), is well conserved, whereas the aa 1–50 segment, including predicted amphipathic α-helix AH1, appears highly variable except for a few well-conserved positions (e.g., basic residues at positions 18 and 20 and a proline at position 38 (Fig. 1). However, the apparent variability is limited at most positions since the observed residues exhibit similar physicochemical properties, as indicated both by the similarity pattern (colons and dots) and the hydropathic pattern, where o, i, and n denote hydrophobic, hydrophilic, and neutral residues, respectively (see Legend to Figure 1 for details). Moreover, all secondary structure prediction methods indicate the presence of an α-helix in the segment comprising aa 5–35 in all genotypes. Hence, despite the apparent aa variability, conservation of the hydropathic pattern suggests that the overall structure of AH1 is conserved.
While AH2 was predicted and subsequently shown to associate with membranes, the N-terminal aa 1–40 segment does not show propensity to partition into a phospholipid bilayer [15]. This is in agreement with an analysis by Palomares-Jerez et al. [20] but in contrast to an earlier report by Elazar et al. [19] (see Discussion section).
To gain insight into the structure and potential lipotropic properties of the NS4B aa 1–40 segment, the corresponding peptide derived from the HCV JFH1 strain (genotype 2a), designated as NS4B[1–40], was chemically synthesized, purified, and analyzed by circular dichroism (CD) and nuclear magnetic resonance (NMR).
The secondary structure of NS4B[1–40] was examined by CD spectroscopy in various membrane mimetic media, including the lysolipid L-α-lysophosphatidylcholine (LPC), detergents (sodium dodecyl sulfate [SDS], n-dodecyl-β-D-maltoside [DDM], dodecylphosphocholine [DPC]), or co-solvents (2,2,2-trifluoroethanol [TFE]-water mixture) (Fig. 2A). In all cases, the CD spectra displayed the typical shape of an α-helix with two minima at 208 and 222 nm and one maximum around 192 nm. CD deconvolution indicated a similar α-helix content of 75±10%, although a weaker signal intensity was observed in the presence of DDM. These results indicate the high propensity of NS4B[1–40] to adopt an α-helical structure in a hydrophobic environment. Interestingly, the peptide solubilized in water displayed a complex spectrum with a weak maximum around 188 nm and two broad minima around 202 and 222 nm, indicating the presence of a mixture of α-helical and random structures. Accordingly, an α-helix content of approximately 22% together with 55% disordered structure was estimated with the various CD deconvolution methods used. Such a secondary structure content for a short peptide suggests that it possibly exists as soluble, micelle-like aggregates which stabilize some residual α-helical folding in aqueous solution through the formation of a hydrophobic core by the hydrophobic sides of several peptide monomers.
As samples of NS4B[1–40] prepared in deuterated micellar SDS and DPC displayed poorly resolved NMR spectra, the three-dimensional structure of the peptide was determined in 50% TFE-d2 which exhibits a CD spectrum comparable to those observed in LPC, SDS and DPC (Fig. 2A) and yielded well-resolved NMR spectra. Sequential assignment of all spin systems was completed and an overview of the sequential and medium-range nuclear Overhauser enhancement (NOE) connectivities is shown in Figure 2B. The NOE connectivity patterns demonstrated that the central part of the peptide, including residues 4–32, displays most characteristics of an α-helix, including strong dNN(i,i+1) and medium dαN(i,i+1) sequential connectivities as well as weak dαN(i,i+2), medium or strong dαN(i,i+3) and dαβ(i,i+3), as well as weak dαN(i,i+4) medium-range connectivities. Apart from this central α-helix, some rare connectivities of weaker intensity are present in both termini of the peptide as a sign of highly flexible unstructured ends. The NOE-based indications of an α-helical fold were supported by the deviation of the 1Hα and 13Cα chemical shifts from random coil values [21] (Fig. 2C). A series of continuous negative variations of 1Hα chemical shifts (Δδ1Hα <−0.1 ppm) and positive variations of 13Cα chemical shifts (Δδ13Cα>0.07 ppm) observed for residues 3 to 32 are indeed typical of an α-helical conformation. Based on the NOE-derived inter-proton distance and dihedral angle constraints deduced from chemical shifts, a set of 50 structures was calculated with X-PLOR, and a final set of 37 low-energy structures that fully satisfied the experimental NMR data were retained. The number and types of NOE constraints used for the structure calculations as well as the statistics for this final set of 37 structures are provided in Table S1. All structures show a regular α-helical conformation extending from residues 4 to 32. A superimposition of the calculated structures is shown in Figure S1. As illustrated in Figure 2D for the representative structure of NS4B[1–40], the central part of the α-helix, including residues Ile 13 to Leu 25, clearly exhibits an amphipathic character, with all hydrophobic residues exposed on one side and charged as well as polar residues on the opposite side. Interestingly, the α-helical folding of this central segment appears to be particularly stable, as indicated by the very slow amide proton exchange observed by NMR in this region (Fig. 2B, blue squares). Hence, this segment likely constitutes an important structural scaffold. In addition, the bulky hydrophobic residues Leu 6 and Ile 7 are also located on the hydrophobic side of the α-helix, suggesting that aa segment 6–25 may bind to a membrane hydrophobic core. These features suggest that AH1 may interact with the membrane interface in an in-plane topology, at least transiently. The C-terminal end (aa residues 26–32) of the α-helix does not include any hydrophobic residue but comprises an intriguingly large number of glutamine residues, which are, however, not conserved in all HCV genotypes (Fig. 1). Because of its purely hydrophilic nature, the folding of this C-terminal part most likely depends on interactions with other parts of NS4B and/or other interaction partners.
To investigate the structural conservation of AH1, we compared two distantly related hepacivirus species, HCV and NPHV. As shown in the upper part of Figure 3, there is only a low degree of aa conservation between the NS4B aa 1–40 segments from the HCV JFH1 strain and NPHV-B10-022 [3]. However, the latter was strongly predicted to adopt an α-helical fold by secondary structure analyses performed as described in the Legend to Figure 1. Based on these observations, α-helix projections generated by Heliquest [22] were compared, as shown in the lower part of Figure 3. Interestingly, despite divergent primary aa sequences, α-helix projections showed a conservation of the amphipathic character of AH1 and of a number of structurally conserved features, including the presence of two positively charged residues, i.e. arginine or lysine, flanking the helices at the borders between the hydrophilic and hydrophobic sides as well as the presence of two negatively charged glutamate residues aligned on the hydrophilic side of AH1 from HCV and NPHV (Fig. 3). Taken together, these observations highlight structurally conserved features of AH1 from phylogenetically distant hepaciviruses, suggesting similar functions of this segment of NS4B within the Hepacivirus genus.
The role of the structurally conserved features of AH1 in HCV RNA replication was investigated by the use of a subgenomic JFH1 replicon harboring a luciferase reporter gene. As shown in Figure 4A, alanine substitution of Lys 18 and Lys 20, either simultaneously (K18A/K20A) or individually (K18A and K20A), abrogated HCV RNA replication, as inferred by comparison with the non-replicative ΔGDD polymerase mutant. In addition, insertion of one alanine residue between Lys 18 and Ser 19 (KASK), resulting in a 110° twist of the α-helix and thus an altered positioning of the two lysine residues in positions 18 and 20, abrogated HCV RNA replication. Hence, the positively charged residues flanking AH1 on either side are required for HCV RNA replication. By contrast, alanine substitution of the two conserved glutamate residues highlighted above (E8A/E15A) did only slightly affect RNA replication at an early time point (about 5-fold reduction in relative light units [RLU] at 24 h, but no appreciable difference at 48 and 72 h).
To explore the mechanism by which substitution of the positively charged aa residues affects HCV RNA replication, the subcellular localization of NS4B harboring the different mutations was investigated in Huh-7 cells using a T7 RNA polymerase-based NS3-5B polyprotein expression system [23]. As shown in Figure 4B, all mutants displayed a similar cytoplasmic, reticular and dot-like distribution pattern, as described previously for NS4B and corresponding to the typical localization on membranes of the ER and of ER-derived modified membranes making up the membranous web [11], [15]. In addition, all mutants colocalized with NS5A to the same extent as the wild-type. These observations indicate that the replication defect of mutants K18A/K20A, K18A and K20A cannot be explained by an aberrant subcellular localization of NS4B or an obvious decrease in NS4B-NS5A colocalization.
To gain deeper insight into the consequences of removal of the positively charged residues flanking AH1, we investigated the ultrastructure of membrane rearrangements induced by the different NS4B mutants by EM. To this end, the mutants were expressed in Huh-7 cells using a T7 RNA polymerase-based NS3-5B polyprotein expression system, as above. Previous work had shown that DMVs are formed in this system and that these closely resemble the HCV-induced membrane rearrangements observed in the context of subgenomic RNA replication and HCV infection [13], [24]. As shown in Figure 5A, regular round-shaped DMVs were readily observed for mutants K18A/K20A, K18A and K20A but, strikingly, these exhibited a large increase in diameter as compared to the ones formed by the wild-type construct. By contrast, mutant E8A/E15A, which replicated almost as wild-type, formed DMVs that were indistinguishable from wild-type. Quantitation of the DMV diameter for each construct demonstrated that DMVs formed by mutants K18A/K20A, K18A and K20A were significantly larger than the ones formed by the wild-type as well as by mutant E8A/E15A (Fig. 5B; see Figure Legend for details). Taken together, these observations indicate that the loss of one of the conserved positively charged residues flanking AH1, i.e. Lys 18 and/or Lys 20, results in the formation of larger DMVs that do not support HCV RNA replication.
In order to further assess the importance of the positively charged residues flanking AH1, mutations K18A, K20A and K18A/K20A were introduced into a selectable subgenomic JFH1 replicon harboring a neomycin resistance cassette in the first cistron. The corresponding in vitro transcribed RNAs were electroporated into Huh-7.5 cells, followed by selection with 500 µg/ml G418 for 3 weeks. About 70 G418-resistant colonies per µg of electroporated RNA were obtained for mutant K18A whereas mutants K18A/K20A and K20A did not yield any viable colonies. Total RNA was extracted from pooled colonies and the NS3-5B region amplified by RT-PCR, followed by cloning of amplicons and sequencing of 10 DNA subclones. Sequence analyses revealed that the K18A mutation in AH1 was retained in all clones. Amino acid changes identified elsewhere in the NS3-5B region are listed in Table 1. Strikingly, mutation P189L in domain I of NS5A was found in 9 out of 10 clones. The only clone which did not harbor this aa change in NS5A showed a substitution of Gln 26 by arginine (Q26R) in NS4B. Interestingly, this mutation results in a new positively charged residue flanking the hydrophobic side of AH1 (Fig. 3). Selection of a compensatory positively charged residue flanking AH1 was also observed in clones harboring pseudoreversions Q22R and Q27R, which coexisted with the NS5A P189L change. The relevance of these compensatory mutations was tested by reintroducing them, alone or in combination with K18A, into a JFH1 subgenomic replicon harboring a luciferase reporter gene. As shown in Figure 6A, the NS5A P189L change alone did not confer any replication advantage to the wild-type replicon and did only slightly (about 2-fold) improve replication of the NS4B K18A mutant. However, the Q22R and Q26R changes in NS4B AH1 strongly enhanced RNA replication capacity of mutant K18A (about 1.5 log for Q22R and 1.0 log for Q26R). Remarkably, the addition of NS5A change P189L to the NS4B K18A/Q22R mutant completely rescued RNA replication, thereby supporting the notion of a functional interaction between NS4B AH1 and NS5A. Finally, the NS4B Q27R change, identified in one clone harboring a number of additional aa changes (Table 1), did not rescue RNA replication of mutant K18A.
Interestingly, modeling of pseudorevertants Q22R and Q26R in the NMR structure of AH1 revealed that the lateral chains of the two selected arginine residues are oriented to the same helix side as the mutated lysine residue in position 18, opposite to Lys 20, thereby restoring a positively charged residue flanking AH1 at distances of one (Q22R) or two (Q26R) α-helix turns (Fig. 6B). By contrast, the Q27R change, which was unable to rescue the replication defect of mutant K18A when reintroduced alone, results in a positively charged residue oriented to the opposite side of Lys 18, i.e. to the same side as Lys 20, thereby failing to restore a conserved feature of AH1. Taken together, selection of pseudoreversions confirms the requirement for two positively charged residues flanking AH1 on either side, as initially suggested by their conservation throughout the different HCV genotypes as well as the Hepacivirus genus (Fig. 3).
As mutant E8A/E15A did not show any significant replication defect, we introduced these aa substitutions into a full-length Jc1 (J6/JFH1 chimeric) HCV genome and assessed virus production by 50% tissue culture infective dose (TCID50) determination. As shown in Figure 7A, infectious virus production by this mutant was strongly reduced, with more than 100-fold lower extra- and intracellular TCID50 yields as compared to the wild-type virus. These results clearly indicate that NS4B AH1 possesses a role in HCV particle production. Quantitation of intra- vs. extracellular TCID50 suggests that the defect is primarily at the level of particle assembly, but an effect on release cannot be excluded (Fig. 7A). Quantitation of intra- vs. extracellular HCV RNA levels demonstrates that mutant E8A/E15A replicates in a full-length viral genome context and confirms the selective defect in virus production of this mutant (Fig. 7B).
As discussed in the Introduction section, the N terminus of NS4B has previously been proposed to adopt a dual topology, based on evidence from glycosylation acceptor site tagging experiments performed in an in vitro translation system as well as in transiently transfected mammalian cells [16], [17]. In addition, we had shown that AH2 has the potential to traverse the phospholipid bilayer and that oligomerization of AH2 is likely required for this process [15], [18]. However, the membrane topology of the N-terminal part of NS4B, comprising AH1 and AH2, has not been investigated in a functional, replicative context. Hence, we took advantage of a recently developed JFH1 subgenomic replicon harboring an HA epitope tag insertion after NS4B aa position 38, i.e. between AH1 and AH2 [14], and examined the topology of the epitope tag by selective membrane permeabilization and immunofluorescence analyses.
In a series of preliminary experiments, we found that incubation of fixed cells with 0.05% digitonin for 15 min at 4°C allowed for selective permeabilization of the plasma membrane but not the endoplasmic reticulum (ER) membrane of Huh7-Lunet cells which are highly permissive for HCV replication [25]. By contrast, 0.2% digitonin under the same experimental conditions permeabilized both membrane compartments. As shown in Figure 8A, the cytosolically oriented core and NS5A proteins could be detected at the same fluorescence intensity under both selective and total permeabilization conditions while the ER luminally oriented E1 glycoprotein could be detected only after total membrane permeabilization. Interestingly, the HA tag inserted between NS4B AH1 and AH2 was consistently detected at an approximately 50% reduced fluorescence intensity under selective as opposed to total membrane permeabilization conditions (Fig. 8A and 8B). These results, which were independent from the choice of anti-HA antibody, indicate a dual topology of the HA tag and, thereby, of the N-terminal part of NS4B in a functional, replicative context.
To investigate whether the AH1 mutations described and characterized above affect the membrane topology of the N-terminal part of NS4B, we introduced mutations K18A/K20A and E8A/E15A into a T7 RNA polymerase-driven construct allowing the expression of NS3-5B including NS4B harboring the HA tag between AH1 and AH2. As shown in Figure 8C and Figure S2, and consistent with the results obtained in a replicative context, the fluorescence intensity for HA tag detection was reduced by about 50% under selective as compared to total membrane permeabilization conditions. Mutations K18A/K20A and E8A/E15A did not affect this ratio. By contrast, a previously described AH2 mutant which is unable to associate with membranes, oligomerize and traverse the phospholipid bilayer, designated as AH2mut [15], [18], was detected at equal fluorescence intensity under both selective and total permeabilization conditions. This additional control validates our selective permeabilization analyses, confirming that the N-terminal part of NS4B adopts a dual topology in an authentic replication context as well as in the context of heterologous NS3-5B expression and is unaffected by mutations K18A/K20A and E8A/E15A which abrogate HCV RNA replication and strongly reduce virus production, respectively.
In the present study, we report the three-dimensional structure of NS4B AH1 and provide a detailed structure-function analysis of this N-terminal amphipathic α-helix. We show that AH1 folds as an α-helix extending from aa 4 to 32 and propose that the fold as well as a number of key structural features are conserved across hepaciviruses. Site-directed mutagenesis and reverse genetics revealed that the conserved positively charged aa residues 18 and 20 flanking AH1 on either helix side are essential for HCV RNA replication. EM analyses demonstrated that these residues play a crucial role in determining the correct size of the DMVs making up the membranous web. Furthermore, the conserved acidic residues 8 and 15 on the hydrophilic side of AH1 were found to be involved in virus production, likely at the level of assembly. Hence, NS4B AH1 possesses a dual role in HCV RNA replication and virus production. Finally, an HCV replicon harboring an epitope tag between AH1 and AH2 allowed to study the topology of the N-terminal part of NS4B in a functional context. Our findings obtained by selective membrane permeabilization and immunofluorescence analyses indicate that this part of NS4B adopts a dual membrane topology in an authentic replication context, likely determined by partial translocation of AH2 across the membrane.
Although AH1 of NS4B is globally amphipathic, NMR analyses revealed that the central amphipathic region comprising aa 13–25 is particularly stable and likely constitutes an important structural scaffold. Together with N-terminal hydrophobic residues Leu 6 and Ile 7, it might interact with the membrane interface in an in-plane topology, at least transiently. Such a binding might also allow stabilization of the α-helical fold of the hydrophilic C-terminal part comprising aa 26–32 upon interaction with the membrane. However, the highly hydrophilic character of this latter part suggests that it may adopt alternative conformations, possibly upon interaction with other parts of NS4B and/or other interaction partners. In addition, the overall hydrophilic character of AH1 is compatible with the absence of direct membrane association per se, as reported previously [15], and suggests that this α-helix may be involved in alternative intra- and/or inter-molecular interactions, e.g. with NS5A (see below).
The present work complements the structures of NS4B AH2 [15], H1 (RM and FP, unpublished) and H2 [26]. AH2 represents an amphipathic α-helix spanning aa 42 to 66 and H2 a “twisted” amphipathic α-helix extending from aa 229 to 253. Although AH1, AH2 and H2 share amphipathic character and play critical roles in HCV RNA replication, their structures are very different. AH1 is mainly hydrophilic, possesses a limited hydrophobic side lacking aromatic residues and is flanked by conserved positively charged residues, AH2 has a very hydrophobic side including 6 highly conserved aromatic residues, and H2 represents a “twisted” amphipathic α-helix. Future work will have to address the structure of the central part of NS4B believed to harbor four transmembrane segments and the ultimate goal will be to solve a complete structure of this integral membrane protein.
Our mutagenesis analyses show that AH1 contributes to RNA replication by affecting the biogenesis of DMVs making up the membranous web. While we previously did not observe a direct interaction of AH1 with cellular membranes [15], it is likely that AH1 interacts with membranes when brought into the appropriate context, e.g. in the presence of AH2 and/or upon oligomerization of NS4B. AH1 may thereby influence membrane curvature induction, resulting in proper DMV formation and assembly of a functional replication complex. Indeed, membrane-associated amphipathic α-helices flanked by positively charged residues have been described to play a role in membrane curvature sensing [27] (reviewed in [28]). As an example, Nath et al. recently demonstrated that the autophagy regulator Atg3 possess an N-terminal amphipathic α-helix which serves as membrane curvature sensor [29]. Comparably to AH1, the Atg3 amphipathic α-helix possesses two lysine residues (aa positions 9 and 11) bordering a hydrophobic face devoid of aromatic aa residues. As we observed DMVs of larger diameter in the AH1 mutant lacking at least one lysine, we may hypothesize that NS4B AH1 plays a similar role in membrane curvature sensing and induction. Hence, AH1 may interact with the surface of a “pre-curved” membrane (large DMVs), possibly induced by AH2 or another mechanism such as NS4B oligomerization, and then further bend the membrane to end up with smaller diameter DMVs competent for RNA replication. Supporting this hypothesis, disruption of the hydrophobic face by introduction of charged residues, similar to the Atg3 mutants described by Nath et al. [29], has previously been shown to abrogate RNA replication [19].
The size and morphology of DMVs have been reported to affect the replication of HCV and other positive-strand RNA viruses. In this context, pharmacologic inhibition or silencing of phosphatidylinositol-4 kinase IIIα or of its effector oxysterol-binding protein, both of which are required for HCV RNA replication, have been shown to reduce DMV diameter [30], [31]. A direct correlation between altered DMV morphology and impaired RNA replication has been demonstrated for mutations in the C-terminal part of NS4B [24]. Mutations in murine hepatitis virus nonstructural protein 4 have been reported to affect DMV morphology and RNA replication [32] and mutations in equine arteritis virus nonstructural protein 3 have been shown to impair DMV formation [33]. However, HCV mutants that increase the diameter of DMVs have to our knowledge not been described previously. Taken together, our observations and previous reports indicate that a defined size and morphology of DMVs is required for efficient viral RNA replication.
Selection for pseudoreversions and modeling in the three-dimensional structure revealed that the two positively charged residues flanking AH1 have to be located on opposite sides of the α-helix for NS4B to be functional in HCV RNA replication. It is possible that these residues stabilize electrostatic interactions with the negatively charged phospholipid head groups in-plane of a membrane surface. Analysis of pseudorevertants indicates that laterally oriented positively charged side chains on either side of the α-helix are required irrespective of their position along the helix. However, the fact that the two lysine residues in aa positions 18 and 20 are conserved across all HCV isolates indicates that viral fitness favors the presence of the positively charged residues at these two specific positions. In keeping with our observations, a mutagenesis study performed previously in a subgenomic replicon derived from the Con1 strain (genotype 1b) identified an important role for Lys 20, with pseudoreversion to positively charged aa residues, i.e. lysine or arginine, at positions 15 or 16 [34].
Selection of aa changes in NS5A, some of which we show here to partially rescue the defect of NS4B mutant K18A, point toward functional interactions and a concerted action of NS4B and NS5A in replication complex formation, as supported by recent EM analyses of the membrane rearrangements induced by NS4B and NS5A [13] as well as earlier work on the phosphorylation of NS5A [35], [36]. In our study, we have identified an NS5A P189L change in 9 out of 10 sequenced clones. This aa change is located in domain I of NS5A, in a surface-exposed position that has the potential to interact with cytosolically oriented NS4B residue Lys 18 (not illustrated). However, it confers only a minor replication advantage in the NS4B K18A mutant context. Similarly, in a previous mutagenesis analysis of the C-terminal region of NS4B, the NS5A K139E change has been identified in several clones without conferring a major advantage for RNA replication on its own but partially required in combination with other pseudoreversions in NS4B [24]. Further investigating the interaction between NS4B and NS5A will in all likelihood reveal important insight into the roles of NS4B in HCV RNA replication and likely also virus production.
While the best known function of NS4B is its role in inducing the membrane rearrangements required for HCV RNA replication, there is growing evidence that NS4B is also involved in virus production [24],[35],[36]. The NS4B mutants described so far either enhanced virus assembly [24], [35] or decreased RNA encapsidation [36] and were all localized in the C-terminal region of NS4B. E8A/E15A is one of the first mutants reported to strongly decrease virus production with almost unimpaired HCV RNA replication capacity. Since intracellular virus titers are reduced to the same extent as extracellular titers, assembly is likely affected in this mutant. Given the evidence for functional interactions between NS4B and NS5A as well as the critical role of NS5A in virus assembly [37], [38], we may hypothesize that AH1 mutant E8A/E15A affects NS4B-NS5A interplay, thereby influencing virion assembly. Future work investigating the role of NS4B in virus production would be facilitated by the development of a complementation system that allows to separate the functions of NS4B in RNA replication and virus production.
A dual topology of the N terminus of NS4B has been suggested earlier [16], [17]. In line with these observations, we had previously shown that AH2 has the potential to traverse the phospholipid bilayer, likely upon oligomerization [15], [18]. Here, we provide evidence for a dual topology of the N-terminal part of NS4B in a functional, replicative context. A working model for these two topologies is illustrated in Figure 9. Based on its physicochemical properties, we could envision a scenario where AH1 remains in the ER lumen, possibly associated with the inner side of the ER membrane, sensing its curvature, and bending the membrane during membranous web formation. However, we believe that it is unlikely that AH1 itself traverses the membrane so that an ER luminal loop be formed between AH1 and AH2 (scenario therefore not illustrated in Fig. 9). Indeed, AH1 appears not hydrophobic enough to achieve a transmembrane topology and there is no obvious interaction platform between multiple copies of AH1 or AH1 and AH2 to yield a transmembrane hydrophobic complex.
Given the limited coding capacity of viral genomes, many encoded proteins have evolved to exert multiple functions. One strategy to achieve multifunctionality may be topological changes with respect of the membrane. In this context, the hepatitis B virus large surface protein and the fusion protein of Newcastle disease virus have been shown to adopt dual membrane topologies with potentially different functions [39], [40]. Based on these observations, it is tempting to speculate that the different topologies of NS4B may serve distinct functions in the HCV life cycle. Clearly, future work will have to address this intriguing possibility.
Sequence analyses were performed using the web-based tools available at the European HCV Database (http://euhcvdb.ibcp.fr/) [41] and the Network Protein Sequence Analysis (NPSA) website of the Institut de Biologie et Chimie des Protéines (IBCP) (http://npsa-pbil.ibcp.fr) [42]. Multiple sequence alignment and analyses of aa conservation were carried out with the ClustalW program using default parameters [43]. HeliQuest was used for α-helix projections (www.heliquest.ipmc.cnrs.fr) [22].
Human hepatocellular carcinoma cell lines Huh-7.5 [44] (kindly provided by Charles M. Rice, The Rockefeller University, New York, NY) and Huh7-Lunet [25] were maintained in Dulbecco's modified Eagle medium supplemented with 10% fetal bovine serum. The Huh-7-derived cell line H7-T7-IZ, which stably expresses the T7 RNA polymerase, was maintained in the same medium supplemented with 5 µg/ml zeocin [13]. Transfections were performed by the use of polyethylenimine [45]. Monoclonal antibodies (mAbs) C7-50 against HCV core [46], A4 against HCV E1 [47] (kindly provided by Jean Dubuisson, University of Lille, France), as well as 9E10 [48] against HCV NS5A (kindly provided by Charles M. Rice) have been described. MAb AC-15 against β-actin was from Sigma-Aldrich. MAb HA-7 and polyclonal antibody Y-11 against the HA tag were from Sigma-Aldrich and Santa Cruz, respectively. Polyclonal antibody #86 against HCV NS4B has been described [37]. Secondary antibodies were anti-mouse-HRP (GE Healthcare), anti-rabbit-HRP (DAKO), Alexa Fluor 488- and Alexa Fluor 594-conjugated anti-mouse IgG, as well as Alexa Fluor 488-conjugated anti-rabbit IgG (Life Technologies).
Plasmids pUHD15-1 (allowing expression of the tetracycline-regulated transactivator tTA) [49] and pUHD-Cp7con (allowing expression from a tTA-responsive promotor) were cotransfected to express core-E1-E2-p7 derived from the HCV H77 (genotype 1a) consensus clone [50], as described [51].
pFK-based plasmids pFK_Jc1_δg, pFK_Jc1ΔE1E2_δg, pFK-i389-neo-sg-JFH1, pFK_i389LucNS3-3′-NS4BHA31R_JFH_δg, pFK_i389LucNS3-3′_JFH_δg and the non-replicating control construct pFK_i389LucNS3-3′_NS5BΔGDD_JFH_δg have been described [14], [52], [53]. Plasmid pTM-NS3-3′_JFH allows for the T7 RNA polymerase-based expression of the HCV replicase proteins (NS3 to NS5B) [23].
Subgenomic HCV JFH1 replicon constructs harboring a bicistronic firefly luciferase reporter gene were based on pFK_i389LucNS3-3′_JFH_δg. Mutations K18A, K20A, K18A/K20A, KASK and E8A/E15A were generated by two-step PCR amplification using primers JFH1-5042-fd and AH1-K18A-rv, AH1-K20A-rv, AH1-KKAA-rv, AH1-KASK-rv or AH1-EEAA-rv, respectively (Table S2), followed by primers AH1-K18A-fd, AH1-K20A-fd, AH1-KKAA-fd, AH1-KASK-fd or AH1-EEAA-fd, respectively, and JFH1-7730-rv (Table S2). Final products were amplified by overlap extension PCR using primers JFH1-5042-fd and JFH1-7730-rv (Table S2), followed by cloning into the NsiI-RsrII sites of pFK_i389LucNS3-3′_JFH_δg, yielding constructs Luc-JFH1-AH1_K18A, Luc-JFH1-AH1_K20A, Luc-JFH1-AH1_K18A/K20A, Luc-JFH1-AH1_KASK and Luc-JFH1-AH1_E8A/E15A.
Neomycin-selectable subgenomic HCV JFH1 replicon constructs were generated by subcloning of the NsiI/BsrGI fragments from the Luc-JFH1 constructs above into pFK_i389-neo-sg-JFH1, yielding constructs Neo-JFH1-AH1_K18A, Neo-JFH1-AH1_K20A, Neo-JFH1-AH1_K18A/K20A and Neo-JFH1-AH1_KASK.
Pseudorevertant constructs were generated in subgenomic HCV JFH1 replicon constructs harboring a bicistronic firefly luciferase reporter gene by two-step PCR amplification using primers JFH1-5042-fd and JFH4BQ22R-rv or JFH4BQ26R-rv (Table S2) with either pFK_i389LucNS3-3′_JFH_δg (wild-type) or Luc-JFH1-AH1_K18A as template, followed by amplification of pFK_i389LucNS3-3′_JFH_δg (wild-type) or the corresponding sequence harboring the NS5A K189L mutation using primers JFH4BQ22R-fd or JFH4BQ26R-fd and JFH1-7730-rv (Table S2). Final products were amplified by overlap extension PCR using primers JFH1-5042-fd and JFH1-7730-rv (Table S2), followed by cloning into the NsiI-RsrII sites of pFK_i389LucNS3-3′_JFH_δg, yielding constructs Luc-JFH1-AH1_Q22R, Luc-JFH1-AH1_Q26R, Luc-JFH1-AH1_K18A/Q22R, Luc-JFH1-AH1_K18A-5A_K189L, Luc-JFH1-AH1_K18A/Q22R-5A_K189L and Luc-JFH1-AH1_K18A/Q26R. Construct Luc-JFH1-AH1_K18A/Q27R was generated by PCR using primers JFH1-5042-fd and JFH1-7730-rv (Table S2), followed by the same cloning strategy.
Jc1 full-length constructs were generated by subcloning of the AvrII/RsrII fragments from the Luc-JFH1 constructs above into pFK-Jc1_δg, yielding constructs Jc1-AH1_E8A/E15A, Jc1-AH1_K18A, Jc1-AH1_K18A/Q22R, Jc1-AH1_Q26R and Jc1-AH1_K18A/Q26R. Construct Jc1-ΔGDD was generated by subcloning of the RsrII/SfiI from pFK_i389LucNS3-3′_NS5BΔGDD_JFH_δg into pFK-Jc1_δg.
pTM-based constructs harboring NS4B mutations were generated by subcloning of the NsiI/RsrII fragments from pFK_i389LucNS3-3′_JFH_δg or pFK_i389LucNS3-3′-NS4BHA31R_JFH_δg into pTM-NS3-3′, yielding constructs pTM-NS3-3′-AH1_K18A, pTM-NS3-3′-AH1_K20A, pTM-NS3-3′-AH1_K18A/K20A, pTM-NS3-3′-AH1_E8A/E15A, and pTM-NS3-3′-NS4BHA31R. Introduction of NS4B mutations into pTM-NS3-3′-NS4BHA31R was performed by two-step PCR using primers JFH1-5042-fd and AH1-KKAA-rv or AH1-EEAA-rv, followed by primers AH1-KKAA-fd or AH1-EEAA-fd and JFH1-7730-rv. Final overlap extension PCR was carried out with primers JFH1-5042-fd and JFH1-7730-rv, followed by cloning into the NsiI-RsrII sites of pTM-NS3-3′, yielding constructs pTM-NS3-3′-NS4BHA31R-AH1_K18A/K20A and pTM-NS3-3′-NS4BHA31R-AH1_E8A/E15A.
All contructs were verified by sequencing.
In vitro transcription of subgenomic replicon and full-length HCV RNA as well as electroporation were performed as described ([54] and references therein). RNA replication assay using a JFH1 subgenomic replicon harboring the firefly luciferase as reporter was performed as described previously [53], [55]. Jc1 virus was produced as described [56]. TCID50 was determined as described [48]. For the determination of intracellular infectivity, cells were harvested and subjected to three freeze and thaw cycles, followed by removal of debris by centrifugation for 2 min at 11,000× g before TCID50 determination.
Protein lysates were prepared and subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) followed by immunoblot analysis as described previously [57].
Peptide NS4B[1–40] from the HCV strain JFH-1 (accession number AB047639; aa sequence reported in Fig. 2) was synthesized on a Milligen 9050 apparatus, employing N-[9-fluorenyl]methoxycarbonyl (Fmoc) chemistry. The peptide was highly purified by reversed-phase high-performance liquid chromatography on a Nucleosil C18 column (120 Å, 5 µm, 250 mm) using a water-acetonitrile gradient containing 0.1% trifluoroacetic acid. The peptide was eluted as a single peak at 65% acetonitrile and identified by mass spectroscopy at its expected molecular mass.
Far UV CD spectra were recorded on a Chirascan spectrometer (Applied Photophysics) calibrated with 1S-(+)-10-camphorsulfonic acid. Measurements were carried out at 298 K in a 0.1-cm path length quartz cuvette. Spectra were measured in a 180–260 nm wavelength range with an increment of 0.2 nm, bandpass of 0.5 nm and integration time of 1 s. Spectra were processed, baseline corrected, smoothed and converted with the Chirascan software. Spectral units were expressed as the mean molar ellipticity per residue by using the peptide concentration determined knowing the weighted NS4B[1–40] peptide used to prepare the NMR sample. Estimation of the secondary structure content was carried out on the DICHROWEB server (http://www.cryst.bbk.ac.uk/cdweb/) [58].
Purified NS4B[1–40] peptide was dissolved in a mixture of 50% TFE-d2 (>99%) in H2O (v/v), and 2,2-dimethyl-2-silapentane-5-sulfonate was added to the NMR samples as an internal 1H chemical shift reference. Multidimensional experiments were performed at 25°C on a Bruker Avance 500 MHz spectrometer using standard homonuclear pulse sequences, including nuclear Overhauser enhancement spectroscopy (NOESY) (mixing times between 100 and 250 ms) and clean total correlation spectroscopy (TOCSY) (isotropic mixing time of 80 ms), as detailed previously [59], [60]. Water suppression was achieved by pre-saturation. Bruker Topspin software was used to process all data and Sparky was used for spectral analysis (http://www.cgl.ucsf.edu/home/sparky/). Intra-residue backbone resonances and aliphatic side chains were identified from homonuclear 1H TOCSY experiments and confirmed with 1H-13C heteronuclear single quantum correlation (HSQC) experiments in 13C natural abundance. Sequential assignments were determined by correlating intra-residue assignments with inter-residue cross peaks observed in bi-dimensional 1H NOESY. NMR-derived 1Hα and 13Cα chemical shifts are reported relative to the random coil chemical shifts in TFE [61].
NOE intensities used as input for structure calculations were obtained from the NOESY spectrum recorded with a 150 ms mixing time and checked for spin diffusion on spectra recorded at lower mixing times (50 ms). NOEs were partitioned into three categories of intensities that were converted into distances ranging from a common lower limit of 1.8 Å to upper limits of 2.8, 3.9 and 5.0 Å, respectively. Protons without stereospecific assignments were treated as pseudoatoms, and the correction factors were added to the upper distance constraints [62]. Additionally, dihedral angle constraints calculated with Talos [63] from 1H and 13C chemical shifts were introduced. Three-dimensional structures were generated from NOE distances with the standard torsion angle molecular dynamics protocol in the XPLOR-NIH 2.30 program [64] using the standard force fields and default parameter sets. A set of 50 structures was initially calculated to widely sample the conformational space, and the structures of low energy with no distance restraint violations (>0.5 Å) were retained. The selected structures were compared by pairwise root-mean square deviation (RMSD) over the backbone atom coordinates (N, Cα and C′). Local analogies were analyzed by calculating the local RMSD of a tripeptide window sliding along the sequence. Statistical analyses, superimposition of structures and structural analyses were performed with MOLMOL [65] and the PDB Protein Structure Validation Suite.
Huh-7.5 cells were electroporated with 1 µg in vitro transcribed RNA from neomycin-selectable subgenomic replicon constructs. Electroporated cells were resuspended in 10 ml of medium, followed by seeding of 10 µl, 100 µl or 1 ml into 10-cm tissue culture dishes. After 24 h, G418 was added at a concentration of 500 µg/ml and maintained until single cell clones became visible. Total RNA was extracted from pooled clones by using the RNeasy Mini Kit (Qiagen) according to the manufacturer's instructions. One µg total RNA was reverse transcribed with specific anchor primer JFH1-9442-rv, followed by PCR amplification using primers EMCV-fd and JFH1-9442-rv using PfuTurbo DNA Polymerase (Stratagene). Amplicons were cloned with Zero Blunt TOPO PCR Cloning Kit (Life Technologies), followed by sequencing of 10 clones.
Glyceraldehyde 3-phosphate dehydrogenase (GAPDH) and HCV RNA levels were measured by SYBR green real-time PCR, as described [55].
H7-T7-IZ or Huh7-Lunet replicon cells were seeded onto glass coverslips. H7-T7-IZ cells were transfected 24 h later with pTM-NS3-3′ plasmids or with pUHD15-1 and pUHD-Cp7, allowing expression of the HCV core-p7 region. Twenty-four h post-transfection cells were fixed with 2% paraformaldehyde (Sigma-Aldrich) for 10 min and then permeabilized for 15 min at 4°C either by 0.2% or 0.05% digitonin (Sigma-Aldrich). Indirect immunofluorescence was performed as described previously [57]. Slides were viewed on a Leica SP5 confocal laser scanning microscope.
Huh7-Lunet T7 cells were seeded onto glass coverslips. On the next day the cells were transfected with pTM-NS3-3′-based expression vectors by using the TransIT-LT1 transfection reagent (Mirus Bio). After 24 h the cells were washed three times with prewarmed PBS, fixed and processed for EM as described previously [24]. Specimens were examined with a Zeiss EM 10 transmission electron microscope at 60 kV.
Significance values were calculated by using the unpaired t test with the GraphPad Prism 6 software package (GraphPad Software).
The atomic coordinates for the NMR structure of synthetic peptide NS4B[1–40] and the corresponding NMR restraints in 50% TFE are available in the Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank under accession number 2LVG (RCSB identification code 102881). The chemical shifts of NS4B[1–40] residues have been deposited in the BioMagResBank (BMRB) under the accession number 18568.
|
10.1371/journal.pcbi.1003778 | Quantifying the Role of Population Subdivision in Evolution on Rugged Fitness Landscapes | Natural selection drives populations towards higher fitness, but crossing fitness valleys or plateaus may facilitate progress up a rugged fitness landscape involving epistasis. We investigate quantitatively the effect of subdividing an asexual population on the time it takes to cross a fitness valley or plateau. We focus on a generic and minimal model that includes only population subdivision into equivalent demes connected by global migration, and does not require significant size changes of the demes, environmental heterogeneity or specific geographic structure. We determine the optimal speedup of valley or plateau crossing that can be gained by subdivision, if the process is driven by the deme that crosses fastest. We show that isolated demes have to be in the sequential fixation regime for subdivision to significantly accelerate crossing. Using Markov chain theory, we obtain analytical expressions for the conditions under which optimal speedup is achieved: valley or plateau crossing by the subdivided population is then as fast as that of its fastest deme. We verify our analytical predictions through stochastic simulations. We demonstrate that subdivision can substantially accelerate the crossing of fitness valleys and plateaus in a wide range of parameters extending beyond the optimal window. We study the effect of varying the degree of subdivision of a population, and investigate the trade-off between the magnitude of the optimal speedup and the width of the parameter range over which it occurs. Our results, obtained for fitness valleys and plateaus, also hold for weakly beneficial intermediate mutations. Finally, we extend our work to the case of a population connected by migration to one or several smaller islands. Our results demonstrate that subdivision with migration alone can significantly accelerate the crossing of fitness valleys and plateaus, and shed light onto the quantitative conditions necessary for this to occur.
| Experimental evidence has recently been accumulating to suggest that fitness landscape ruggedness is common in a variety of organisms. Rugged landscapes arise from interactions between genetic variants, called epistasis, which can lead to fitness valleys or plateaus. The time needed to cross such fitness valleys or plateaus exhibits a rich dependence on population size, since stochastic effects have higher importance in small populations, increasing the probability of fixation of neutral or deleterious mutants. This may lead to an advantage of population subdivision, a possibility which has been strongly debated for nearly one hundred years. In this work, we quantitatively determine when, and to what extent, population subdivision accelerates valley and plateau crossing. Using the simple model of an asexual population subdivided into identical demes connected by gobal migration, we derive the conditions under which crossing by a subdivided population is driven by its fastest deme, thus giving rise to the maximal speedup. Our analytical predictions are verified using stochastic simulations. We investigate the effect of varying the degree of subdivision of a population. We generalize our results to weakly beneficial intermediates and to different population structures. We discuss the magnitude and robustness of the effect for realistic parameter values.
| Natural selection drives populations towards higher fitness (i.e. reproductive success), but crossing fitness valleys or plateaus may facilitate progress up a rugged fitness landscape. Rugged fitness landscapes arise from epistasis, i.e. interactions between genetic variants. For instance, two mutations together can yield a benefit while each of them alone is detrimental: such reciprocal sign epistasis can give rise to a fitness valley [1], [2]. While the high dimensionality of genotype space makes it challenging to probe the structure of fitness landscapes [3], [4], evidence has been accumulating for frequent landscape ruggedness, especially in recent years [1], [2], [4]–[15].
Population structure can play an important role in evolution [16]–[24]. In particular, the time taken to cross a fitness valley or plateau depends on population size since stochastic effects such as genetic drift have an increased importance in small populations, allowing neutral and deleterious mutations to fix with increased probability [25]–[28]. Population subdivision into demes can allow the maintenance of larger genetic diversity due to increased genetic drift as well as to the quasi-independent explorations of the fitness landscape that are run in parallel by each deme. Subdivision may thereby facilitate valley or plateau crossing locally and subsequent migration can then spread beneficial mutations throughout the entire subdivided population ("metapopulation''). This idea was first discussed by Wright in his shifting balance theory [29]–[32] and the importance of this effect has been the subject of a long debate [33]–[42]. In this work, we investigate the role of subdivision with global migration alone, without additional effects such as strong dependence of deme size on fitness, including extinction and refounding of demes, which played a crucial role in Wright's theory. Our generic and minimal model enables us to quantatively determine the conditions under which population subdivision accelerates fitness valley or plateau crossing.
Studying quantitatively the effect of subdivision on evolution may help in inferring fitness landscape structure from evolution experiments [43]. Work on structured populations has been used as qualitative proof of landscape ruggedness [16]. Current experiments investigating the evolution of subdivided populations at various migration rates have produced mixed results, some demonstrating faster adaptation of subdivided populations [44], and others not [45]. It is therefore important to determine under what conditions subdivision accelerates fitness valley or plateau crossing. Additionally, population subdivision is extremely common in natural systems. For instance, evidence has recently been found for compartmentalization of HIV in different organs of a single patient [46], [47].
Here we show that subdivision can significantly accelerate fitness valley or plateau crossing over a wide parameter range, both with respect to a non-subdivided population and with respect to a single deme. Intuitively, deleterious or neutral intermediate mutations may fix in individual demes, allowing for the maintenance of a larger proportion of these mutants in a metapopulation than in a well-mixed population. We first determine the optimal speedup of valley or plateau crossing by subdivision, in the best possible scenario where valley or plateau crossing by the metapopulation is driven by that of its fastest deme. This enables us to demonstrate that isolated demes must be in the sequential fixation regime for subdivision to significantly accelerate crossing. We then determine the conditions under which the best possible scenario can be realized. Using Markov chain theory, we obtain analytical expressions for the parameter range where valley or plateau crossing by a metapopulation is as fast as that of its fastest deme. Our analytical predictions are verified using stochastic simulations. Furthermore, we discuss the effect of varying the degree of subdivision of a population, and investigate the trade-off between the magnitude of the optimal speedup and the width of the parameter range over which it occurs. Finally, we extend our work to weakly beneficial mutations and to a population connected to smaller islands, and we discuss the magnitude and robustness of the effect for realistic parameter values.
Our results are organized as follows. First, we specify our model for the evolutionary dynamics of a subdivided population with migration. Then, we focus on the ‘best possible' scenario where the metapopulation is driven by its fastest deme. We calculate the ratio of the valley-crossing time for the metapopulation to the valley-crossing time for an equally-sized well-mixed population under this strong assumption. This yields the optimal speedup that may be obtained by subdivision, and enables us to demonstrate that sequential fixation in individual demes is necessary to achieve a significant speedup. Then, we determine the range of parameter values for which the best possible scenario is attained, i.e. the valley-crossing time for the metapopulation is indeed dominated by the valley-crossing time of its fastest deme. Qualitatively, migration has to be both rare enough to enable demes to cross the fitness valley or plateau quasi-independently and frequent enough to allow fast spreading of the final beneficial mutation to the whole metapopulation once it has fixed in the fastest deme: these conditions yield an optimal window of migration rates. Finally, we compare our analytical predictions with results from stochastic simulations.
We focus on asexual individuals, characterized by their genotype and associated fitness . Each individual has a division rate proportional to , and a death rate , which is the same for all. We consider an ensemble of identical demes, each with a constant number of individuals. The division rate averaged over the individuals of a deme is thus equal to the death rate . We treat migration as a random exchange of two individuals between two different demes, occurring at rate per individual. In our model, exchange between any two demes is equally likely, as in Wright's "island model'' [29]. This constitutes a generic and minimal model of subdivision with migration, without any dependence of migration rate on the average fitness of a deme (in contrast with models where demes containing beneficial mutants increase significantly in size and migrate more rapidly [30], [33]), or additional effects of extinction and re-founding of demes [30], [32], [33], specific geographic structure [16], [17], [19]–[21], or spatially heterogeneous environments [18], [22]–[24], on which previous studies focused.
We consider the simplest fitness valley or plateau, involving three successive genotypes denoted by ‘0', ‘1' and ‘2' (see Fig. 1A). The initial genotype is taken as reference for fitness: . We denote the fitnesses of the subsequent genotypes by and . The first mutation is assumed to be either neutral (), which yields a fitness plateau, or deleterious (), which corresponds to a fitness valley, while the second mutation is assumed to be beneficial (). We focus on first mutations that are not too strongly deleterious: . We only allow forward mutations, and note that including back mutations does not qualitatively affect crossing times [28]. Finally, we assume that all mutations have probability per division, but generalization to different mutation probabilities is straightforward.
In this paper, we focus on the average time required for the whole metapopulation to cross the fitness valley or plateau, i.e. to fix mutation ‘2' in all demes, starting from an initial state where all individuals have genotype ‘0’.
For small enough migration rates, each deme in the metapopulation performs a quasi-independent trial at crossing the valley or plateau. At best, the valley or plateau crossing time of the whole metapopulation is dominated by that, , of the “champion'' deme in the metapopulation, i.e. the deme that crosses the fitness valley or plateau fastest.
We now focus on this best possible scenario, which is illustrated schematically in Fig. 1B: first, the champion deme crosses the valley or plateau by sequential fixation, and then the beneficial mutation rapidly spreads by migration of through the whole metapopulation. Once this best possible scenario is characterized, the crucial question will be whether, and under what conditions, it can be attained: this point will be addressed in the following section.
The previous section was dedicated to the study of the best possible scenario, where the valley or plateau crossing time of the whole metapopulation is dominated by that, , of the champion deme in the metapopulation (i.e. the one that crosses fastest). We now determine analytically the conditions under which this best possible scenario is attained. For this, we focus on migration rates much smaller than division/death rates, , such that fixation or extinction of a mutant lineage in a deme is not perturbed by migration. In addition, we assume that isolated demes are in the sequential fixation regime, since we showed above that it is a necessary condition for subdivision to significantly accelerate crossing, and that it is a sufficient condition for subdivision to accelerate crossing in the best scenario.
In a nutshell, migration must be rare enough for demes to evolve quasi-independently, but frequent enough to spread the beneficial mutation rapidly. The analytical results below allow for predicting the range of migration rates such that subdivision maximally accelerates valley or plateau crossing.
We now present numerical simulations of the evolutionary dynamics described above, which enable us to test our analytical predictions, and to gain additional insight in the process beyond the optimal scenario. Our simulations are based on a Gillespie algorithm [48], [49], and described in detail in Methods, Sec. 1.
Let us first focus on the example presented in Fig. 1D, which shows an example plot of as a function of the ratio of migration to mutation rates, , obtained through our simulations when varying only the migration rate. With the parameter values used in this figure, the interval of Eq. 14 is . Note that here, and in the following examples, we use the general expressions of and given in Methods, Sec. 3, to compute the interval of Eq. 14. Fig. 1D features a minimum right at the center of this theoretically predicted optimal interval. Moreover, this minimum corresponds to , while : hence, the metapopulation crosses the valley on average 6.54 times faster than an isolated deme. This is very close to the limit of the best possible scenario, where the metapopulation would cross 7 times faster than an isolated deme (since here). This example illustrates that speedups tend towards those predicted in the best scenario, when the interval in Eq. 14 is sufficiently wide (here the ratio between its upper and its lower bound is 359). Besides, here: comparing it to the above-mentioned value of yields a 3.47-fold speedup of valley crossing by subdivision. The simulation results in Fig. 1D also show that significant (albeit smaller) speedups exist beyond the optimal parameter window.
Fig. 2 shows heatmaps of the valley crossing time of a metapopulation as a function of the migration-to-mutation rate ratio, (varied by varying ), and of the fitness valley depth, . Fig. 2A shows that the optimal interval of Eq. 14 (solid lines) describes well the region where the ratio of the crossing time of the metapopulation to that of an isolated deme is smallest and tends to the best-scenario limit . For migration rates lower than those in this interval, the ratio increases when decreases. This can be understood qualitatively by noting that if , is determined by the valley crossing time of the slowest among the independent demes. In the opposite case of migration rates larger than those in the optimal interval, increases with , and it tends to the non-subdivided case, , at high values of , as expected. Above a threshold value of (dashed line), becomes smaller than , in which case large values of , such that tends to , give a low (see Fig. 2A).
Fig. 2B plots the ratio of the crossing time of the metapopulation to that of the non-subdivided population, which directly yields the speedup obtained by subdividing a population. It shows that, for the parameter values chosen, subdivision accelerates valley crossing over a large range of valley depths and migration rates, extending far beyond the optimal range given by Eq. 14, and that the metapopulation can cross valleys orders of magnitude faster than a single large population. In addition, above a second, larger threshold value of (dotted line in Fig. 2), isolated demes enter the tunneling regime [28]: Fig. 2B shows that sufficiently above this threshold, the metapopulation no longer crosses the valley faster than the non-subdivided population, as predicted above. While having isolated demes in the sequential fixation regime is a necessary condition to obtain significant speedups by subdivision, the non-subdivided population is not required to be in the sequential fixation regime (see above, and Fig. 1C–D). The value of above which the non-subdivided population enters the tunneling regime is indicated by a dash-dotted line in Fig. 2: significant speedups are obtained both below and above this line. The highest speedups are actually obtained above it, i.e. when the non-subdivided population is in the tunneling regime. With the parameter values used, Eq. 8 predicts a minimum of for (solid line in Fig. 2B), which agrees very well with the results of our numerical simulations. (Note that this value of satisfies , and is such that the non-subdivided population is in the tunneling regime. These conditions were used in our derivation of Eq. 8.)
In the Results section, we have shown that having isolated demes in the sequential fixation regime is a necessary condition for subdivision to significantly accelerate crossing. This requirement limits the interval of the ratio over which the highest speedups by subdivision are obtained. The extent of this interval can be characterized by the ratio, , of the upper to lower bound in Eq. 14. Let us express the bound on imposed by the requirement of sequential fixation in isolated demes.
If , the threshold value below which an isolated deme is in the sequential fixation regime satisfies [28]. Let us also assume that , and that while , to be in the domain of validity of Eqs. 15 and 16. Combining the condition with the expression of in Eq. 16 yields(19)
For plateaus, isolated demes are in the sequential fixation regime if their size is smaller than [28]. In the regime of validity of Eqs. 17 and 18 ( while , and , ), this condition can be combined with Eq. 18, which yields(20)
Both Eq. 19 and Eq. 20 show that increasing the number of demes decreases the range where the highest speedup by subdivision is reached. This is because having more subpopulations makes the spreading of the beneficial mutation slower. In addition, we find that the bound on is proportional to . Hence, despite this bound, the interval where subdivision most accelerates plateau crossing can span several orders of magnitude, given the small values of the actual mutation probabilities in nature.
An interesting question raised by our results regards the optimal degree of subdivision. Given a certain total metapopulation size, into how many demes should it be subdivided in order to obtain the highest speedup possible? We first attack this question using our analytical results, and then we present simulation results, which allow for going beyond the best scenario and its associated parameter window.
Let us consider a metapopulation of given total size . Our analytical results show that increasing subdivision, i.e. increasing the number of subpopulations at constant , leads to stronger speedups of valley crossing (see Eqs. 4 and 7, with ). However, Eqs. 16 and 18, and the previous paragraph, show that when is increased, the parameter range where the speedup by subdivision tends to the best-scenario value becomes smaller and smaller. Eventually, this parameter range ceases to exist altogether: this occurs when becomes of order 1 and below. This sheds light on an interesting trade-off in the degree of subdivision , between the magnitude of the optimal speedup gained by subdivision and the width of the parameter range over which the actual speedup is close to this optimal value. This effect can be observed qualitatively in Fig. 3A, where the valley crossing time of a metapopulation with fixed total size is shown versus the migration-to-mutation rate ratio, , for different values of : when is increased, the minimum becomes deeper but less broad.
In addition, Eqs. 15 and 17 show that when is increased, the lower bound of the interval where the speedup by subdivision tends to the best-scenario value decreases, as for plateaus (Eq. 17) and even more rapidly for deep valleys (Eq. 15). Qualitatively, this is because spreading of the beneficial mutation gets longer when increases. Conversely, the upper bound of this parameter range is independent of for deep valleys (Eq. 15), and grows only logarithmically with for plateaus (Eq. 17). Hence, when is increased, the center of the interval where the actual speedup is close to the optimal value shifts towards higher migration rates. This effect, which can be observed in Fig. 3A, is studied more precisely in Fig. 3B: at fixed migration rate , the crossing time of a metapopulation exhibits a minimum at an intermediate value of . Indeed, the crossing time of the metapopulation first decreases when is increased because the minimum crossing time then decreases. But beyond a certain value of , the migration rate that yields the highest speedup becomes larger than the fixed migration rate , so increases when is increased further.
Next, we study the dependence on of the valley crossing time minimized over for each , again for a metapopulation with fixed total size . For values of small enough for the interval in Eq. 14 to be broad, we expect to be close to the optimal scenario value . But, as discussed above, as increases, this interval will become smaller and then vanish. In such a regime, our analytical results are no longer sufficient to predict the dependence of on , but our simulations can provide additional insight. Fig. 3C shows that, while (left of the dashed line), is close to the best-scenario value. When is increased beyond this point, decreases slower than the best-scenario value. Indeed, the interval in Eq. 14 is no longer wide enough for the best-scenario limit to be approached. Note also that when demes become small enough, verifying (right of the dotted line in Fig. 3C), mutation ‘1' becomes effectively neutral in individual demes, as tends to (see Eq. 1). For even higher values of , is observed to saturate rather than exhibiting a unique minimum. Interestingly, this occurs for such that the interval in Eq. 14 fully vanishes (i.e. when passes below 1, right of the solid line on Fig. 3C). While we do not have rigorous proof of the generic existence of this saturation, we have explored this point for other parameters, and found similar behavior (data not shown). Importantly, this indicates that there is a whole class of nearly optimal population structures.
Our work has focused on fitness valleys (), such that mutation ‘1' is deleterious, and on fitness plateaus (), such that mutation ‘1' is neutral. For , mutation ‘1' is effectively neutral, as far as valley crossing is concerned, in a population with individuals [28]. (This condition holds both in the sequential fixation regime and in the tunneling regime.) This implies that our arguments and our results obtained in the case of the fitness plateau also hold for weakly beneficial intermediates. This point is illustrated in Fig. 4A.
Thus far, we focused on demes of equal size for simplicity, but demes of different sizes are relevant in practice. As a step toward more general populations structures, we now consider a population connected by migration to smaller satellite populations of identical size, assumed to be in the sequential fixation regime. We only allow migration between the large population and each of the smaller islands, and the total migration rate is denoted by . The small island affected by migration is chosen randomly at each migration event. It is straightforward to adapt our work to this case (see Methods, Sec. 4). We obtain an interval of over which the crossing time for the large population is dominated by the crossing time of the champion island. This is corroborated by our simulations (see Fig. 4B).
Let us consider the example of Escherichia coli, for which the mutation probability per base pair per division is [50]. In order to gain a speedup of crossing by subdivision, we require demes to be in the sequential fixation regime. For plateaus, this condition reads . Let us consider deme sizes such that this condition is satisfied.
First, let us choose , which is within the smallest range of sizes used in current evolution experiments. For instance, it is the number of bacteria transferred at each dilution step for small populations in [27]. For this value of , all plateaus with are in the sequential fixation regime (from the condition ). Let us also consider , since 96-well plates are often used in these experiments [27], [51]. This yields a total population size of individuals, which is in the tunneling regime for all plateaus with . For , isolated demes are in the sequential fixation regime for . (Subdivision cannot significantly accelerate crossing for deeper valleys since isolated demes are then in tunneling, but those valleys take longer to cross than shallow ones and are thus probably less often crossed in practice.) The ratio of the bounds of the interval in Eq. 14 satisfies throughout this range of valleys, with for the plateau and for the deepest valleys in the range. Thus, actual speedups will approach the best-scenario one, and significant speedups will exist in a wide parameter window. Eq. 8 predicts that the highest speedup is obtained for , and Eq. 9 then yields a speedup factor by subdivision of . (Using instead the full expression of obtained from Eq. 23 (see Methods, Sec. 2) yields , i.e. a correction of 7%.) Moreover, for all valleys with , the best-scenario speedup ranges from 18 to . Thus, subdivision significantly accelerates crossing for this entire class of valleys.
It should be noted that the timescales obtained in this example are long compared to experimental ones. For instance, for the plateau, corresponds to divisions while is divisions. However, can become smaller if the number of subpopulations is increased, as discussed in our previous section. Besides, we have chosen to focus on standard Escherichia coli for simplicity. Organisms with a higher mutation rate, e.g. viruses such as HIV, or mutator strains, would have much shorter timescales, but smaller subpopulations would then be required for demes to be in the sequential fixation regime.
Our example thus far focused on a small but realistic deme size, . Experimentally more frequent values of are in the range – [27], [51]. Increasing at fixed decreases the range of for which demes are in the sequential fixation regime. For a plateau, this condition reads . For , this yields , and for , this yields . Hence, the range of plateaus (and similarly, of valleys) for which subdivision accelerates crossing becomes more restricted when is increased. Nevertheless, if these increasingly stringent conditions on are satisfied, significant speedups by subdivision are still expected. Indeed, Eq. 9 shows that the smallest value of the ratio is proportional to , so if one increases while decreasing as , the maximal speedup by subdivision will remain unchanged.
In this work, we have considered the crossing of one particular valley or plateau corresponding to a specific pair of two mutations. Given the complexity and high dimensionality of actual fitness landscapes, there may be a large number of parallel valleys or plateaus, so that one of these could be crossed quite frequently even though the crossing time for a single valley or plateau remains large. Our work shows that, under specific conditions, subdivision can significantly accelerate crossing for whole classes of valleys and plateaus. Furthermore, in a generic, high-dimensional fitness landscape that contains both valleys and/or plateaus and uphill paths, subdivision can provide an additional effect: it “shields'' some demes in the metapopulation from adaptation via the uphill paths, leaving them time to explore valley-crossing paths that may be better in the longer term. While this effect is outside the scope of the present paper, it could lead to additional advantages of subdivision in evolution on rugged fitness landscapes.
Our study of a generic and minimal model of population subdivision with migration demonstrates that subdividing a population into demes connected by migration can significantly accelerate the crossing of fitness plateaus and valleys, without the need for additional ingredients. We have derived quantitative conditions on the various parameters for subdivision to accelerate crossing, and for the resulting speedup to be maximal. In particular, isolated demes have to be in the sequential fixation regime for a significant speedup to occur. This condition is quite strong, but provided that it is met, significant speedups can be obtained in a wide range of migration rates, with the fastest deme driving the crossing of the whole metapopulation in the best scenario. We have derived the interval of migration rates for which this best scenario is reached. In addition, we have shown that increasing the degree of subdivision of a population enables higher speedups to be reached, but that this effect can saturate.
Our quantitative assessment of the conditions under which subdivision significantly speeds up valley or plateau crossing can aid in optimally designing future experiments, enabling one to choose the sizes and the number of demes, as well as the migration rates, such that subdivision can accelerate valley and plateau crossing.
Further directions include investigating the evolution of a metapopulation with a distribution of deme sizes on a more general rugged landscape, as well as assessing the impact of specific geographic structure. Our work could also be extended to sexual populations, where recombination plays an important role in valley or plateau crossing [52]. The interplay between recombination and subdivision, which respectively alleviate and exacerbate clonal interference, would be interesting to study.
Our simulations are based on a Gillespie algorithm [48], [49] that we coded in the C language. Here we will describe our algorithm for the case of a metapopulation of demes of identical size, which is the primary situation discussed in our work. In our simulations, each deme has a fixed carrying capacity –we discuss this choice further in this section.
In this section, we give more details on the calculation of the average valley or plateau crossing time by the champion deme amongst independent ones. We show in the Results section that, in the best scenario, the crossing time of the whole metapopulation is determined by this time.
is the average shortest crossing time of independent demes. This minimum crossing time, which we denote by , is also called the smallest (or first) order statistic of the deme crossing time amongst a sample of size [53].
Let us denote by the probability density function of valley or plateau crossing time for a single deme, and let us introduce (it satisfies where is the cumulative distribution function of valley or plateau crossing by a single deme). The probability that is larger than is equal to the probability that the crossing times of each of the independent demes are all larger than . By differentiating this expression, one obtains the probability density function of the crossing time by the champion deme (see e.g. [53]):(21)
We now express explicitly. Since demes are assumed to be in the sequential fixation regime, valley or plateau crossing involves two successive steps. The first step, fixation of a ‘1'-mutant, occurs with rate , and the second step, fixation of a ‘2'-mutant, occurs with rate (see the Results section for expressions of these rates). The total crossing time is thus a sum of two independent exponential random variables, with probability density function given by a two-parameter hypoexponential distribution [53]:(22)
Combining Eqs. 21 and 22, we obtain(23)with given by Eq. 22. can then be determined for any value of the parameters by computing the average value of over this distribution.
Since mutation ‘1' is deleterious or neutral while mutation ‘2' is beneficial, the first step of valley crossing is much longer than the second one over a broad range of parameter values. In this case, we can approximate with a simple exponential distribution,(24)
Eq. 21 then yields(25)i.e. is distributed exponentially with rate . In this case, we simply have , which can be written as , where is the average crossing time for an isolated deme. Hence, in this case, on which our analytical discussion focuses, the champion deme crosses the valley times faster on average than an isolated deme.
For this approximation to be valid, the second step of valley crossing must be negligible even for the champion deme, i.e., . For very large , the value of will not be as small as , since the second step will no longer be negligible (see [52] for a discussion of similar issues). The crossover to this regime can be determined by computing the average of the distribution in Eq. 23 and comparing it to .
In our Results section, we have derived an interval of the ratio of migration rate to mutation rate over which subdivision most reduces valley or plateau crossing time (see Eq. 14). The upper bound involves , the average number of migration events required for the ‘1'-mutants to be wiped out by migration, starting from a state where one deme has fixed genotype ‘1', while all other demes have genotype ‘0'. Similarly, the lower bound involves , the average number of migration events required for the ‘2'-mutants to spread by migration to the whole metapopulation, starting from a state where one deme has fixed genotype ‘2', while all other demes have genotype ‘0'. In our Results section, we have provided intuitive derivations of the simple expressions of and , valid for and , and (see Eq. 15). However, it is important to derive more general expressions, especially since subdivision generically most accelerates valley crossing in the intermediate regime where (see Results, Eq. 8).
Here, we derive general analytical expressions for and , both for fitness plateaus and for fitness valleys. These more general expressions are those used for numerical calculations of the bounds in our examples. Throughout this section, we consider a metapopulation of demes composed of individuals each, and we assume that individual demes are in the sequential fixation regime (see Results).
Let us consider a population of individuals connected by migration to smaller population islands with individuals each. These islands of identical size are assumed to be in the sequential fixation regime. For the sake of simplicity, we consider that migration only occurs between the large population and the islands: a migration step is a random exchange of two individuals between the large population and one of the islands (chosen at random at each migration event), and the total migration rate is denoted by . Here, we focus on the valley or plateau crossing time of the large population. We demonstrate that the evolution of a large population can be driven by that of satellite islands.
In the optimal case, the crossing time of the large population is determined by that of the champion island, i.e., that which crosses the fitness valley or plateau fastest. We now determine the conditions under which this optimum is achieved, focusing on migration rates much smaller than division/death rates, , such that fixation or extinction of a mutant lineage in either the large population or an island is not significantly perturbed by migration. Again, migration should be rare enough for islands to remain effectively shielded from migration events while they have fixed the intermediate mutation, until the final beneficial mutation arises. Second, migration should also be frequent enough for the spreading time of the final beneficial mutation from the champion island to the large population to be negligible with respect to the crossing time of the champion island. These two criteria again provide upper and lower bounds on .
The average time (with from Eq. 1) required for an island of ‘1'-mutants to fix the beneficial mutation ‘2' must be smaller than the average time, , for an island of ‘1'-mutants to be wiped out by migration from the large population, which still exhibits genotype ‘0'. The rate of migration events between the island of ‘1'-mutants and the large population is . Hence, , where is the probability of fixation of the lineage of a single migrant with genotype ‘0' in an island where all other individuals are ‘1'-mutants: for valleys, it is given by Eq. 1, while for plateaus, it is equal to . The first condition, , thus yields(47)
The second condition is that the average spreading time, , for the final beneficial mutation to fix in a large population after it has fixed in the champion island, must be smaller than the average valley or plateau crossing time, , of the champion island. Similar to previously, we obtain , where is the probability of fixation of a migrant with genotype ‘2' in the large population, which is assumed to exhibit genotype ‘0' before migration (see Eq. 1). is the average of the minimum crossing time among independent islands. We again focus, for simplicity, on the limit where the first step of valley or plateau crossing, which occurs at rate , is much longer than the second. Then, we simply have (see Results). In this expression, (with obtained from Eq. 1) is the average crossing time for an isolated island. Hence, the champion island crosses the valley times faster on average than a single isolated island. The second condition, , finally yields(48)
Together, Eqs. 47 and 48 yield the interval of over which we expect subdivision to maximally accelerate crossing:(49)
In this range, we expect the valley or plateau crossing time of the large population to be dominated by the crossing time of the champion island, so that . This prediction is confirmed by our simulations (see Fig. 4B).
|
10.1371/journal.pbio.0060034 | Caldendrin–Jacob: A Protein Liaison That Couples NMDA Receptor Signalling to the Nucleus | NMDA (N-methyl-D-aspartate) receptors and calcium can exert multiple and very divergent effects within neuronal cells, thereby impacting opposing occurrences such as synaptic plasticity and neuronal degeneration. The neuronal Ca2+ sensor Caldendrin is a postsynaptic density component with high similarity to calmodulin. Jacob, a recently identified Caldendrin binding partner, is a novel protein abundantly expressed in limbic brain and cerebral cortex. Strictly depending upon activation of NMDA-type glutamate receptors, Jacob is recruited to neuronal nuclei, resulting in a rapid stripping of synaptic contacts and in a drastically altered morphology of the dendritic tree. Jacob's nuclear trafficking from distal dendrites crucially requires the classical Importin pathway. Caldendrin binds to Jacob's nuclear localization signal in a Ca2+-dependent manner, thereby controlling Jacob's extranuclear localization by competing with the binding of Importin-α to Jacob's nuclear localization signal. This competition requires sustained synapto-dendritic Ca2+ levels, which presumably cannot be achieved by activation of extrasynaptic NMDA receptors, but are confined to Ca2+ microdomains such as postsynaptic spines. Extrasynaptic NMDA receptors, as opposed to their synaptic counterparts, trigger the cAMP response element-binding protein (CREB) shut-off pathway, and cell death. We found that nuclear knockdown of Jacob prevents CREB shut-off after extrasynaptic NMDA receptor activation, whereas its nuclear overexpression induces CREB shut-off without NMDA receptor stimulation. Importantly, nuclear knockdown of Jacob attenuates NMDA-induced loss of synaptic contacts, and neuronal degeneration. This defines a novel mechanism of synapse-to-nucleus communication via a synaptic Ca2+-sensor protein, which links the activity of NMDA receptors to nuclear signalling events involved in modelling synapto-dendritic input and NMDA receptor–induced cellular degeneration.
| Long-lasting changes in communication between nerve cells require the regulation of gene expression. The influx of calcium ions into the cell, particularly through membrane protein called NMDA receptors, plays a crucial role in this process by determining the type of gene expression induced. NMDA receptors can exert multiple and very divergent effects within neuronal cells by impacting opposing phenomena such as synaptic plasticity and neuronal degeneration. We identified a protein termed Jacob that appears to play a pivotal role in such processes by entering the nucleus in response to NMDA receptor activation and controlling gene expression that governs cell survival and the stability of synaptic cell contacts. Removal of Jacob from the nucleus protects neurons from NMDA receptor–induced cell death and increases phosphorylation of the transcription factor CREB, whereas the opposite occurs after targeting Jacob exclusively to the nucleus. The work defines a novel pathway of synapse-to-nucleus communication involved in modelling synapto-dendritic input and NMDA receptor–induced cellular degeneration.
| Ca2+ signals triggered by NMDA-type glutamate receptors can result in long-lasting changes of synaptic input and dendritic cytoarchitecture in phenomena commonly referred to as neuronal plasticity. On the contrary, NMDA receptors are also important players in neurodegenerative processes. Although both aspects require gene expression, our knowledge is still sparse concerning how these fundamental processes are regulated at the molecular level. The Janus face of neuronal NMDA receptor signalling is probably best reflected by the fact that the influx of Ca2+ ions is thought to act as one of the major mediators of synapto-nuclear signalling [1,2] and of excitotoxic cell death [3]. Within this scheme, a prevailing idea is the existence of Ca2+ microdomains coupled to the activation of synaptic and extrasynaptic NMDA receptors, and transducing incoming Ca2+ events to different downstream pathways [1–4]. In a series of elegant studies, Hardingham and colleagues [5–7] provided evidence that Ca2+ influx through synaptic NMDA receptors trigger nuclear cAMP response element-binding protein (CREB) phosphorylation via an extracellular signal-regulated kinase (ERK)-dependent pathway, whereas Ca2+ influx through extrasynaptic NMDA receptors leads via an ERK-independent pathway to a dephosphorylation of CREB termed CREB shut-off. As opposed to the synaptic pathway, the CREB shut-off signal is coupled to neuronal degeneration and cell death [7]. Thus, CREB-regulated gene expression appears to be a shared mechanism for both long-term plasticity and neuronal survival [1–3,8–10].
Although Ca2+ exerts its signalling functions via a variety of Ca2+ sensor proteins, pathways that result in a nuclear response to synaptic activity have primarily been based on signalling via calmodulin (CaM) [1,2]. In its Ca2+-bound state, CaM alters the properties of several other proteins and signalling cascades that have been implicated in diverse neuronal functions [11,12]. Whereas it is tacitly assumed that CaM is present in large excess in all cellular compartments, and therefore regulation of CaM signalling largely depends on binding of Ca2+ ions, a variety of additional EF-hand proteins have been identified in neurons, termed neuronal Ca2+ sensor (NCS) proteins. These NCS are believed to serve more-specific functions in neurons [13–15].
One of these NCS proteins is Caldendrin (also termed CaBP1) [16,17], a bipartite protein with a unique N-terminal half and a C-terminal half that contains four EF-hand motifs and qualifies Caldendrin as the closest relative of CaM in brain neurons. The second EF-hand is most likely cryptic [16,18]. Modelling of the C-terminal segment suggests that Caldendrin displays an altered surface-exposed amino acid residue distribution, especially at EF-hand 2 as compared to CaM [18]. Interestingly, the unique N-terminal half of Caldendrin exhibits no similarity to other known proteins [16]. Moreover, in contrast to the ubiquitously expressed CaM, Caldendrin is only present in a subset of synapses and seems to be exclusively and tightly associated with the somatodendritic cytoskeleton and the postsynaptic density (PSD) of mature principal neurons in brain regions with a laminar organization [16,19]. To test the hypothesis that Caldendrin might have specific functions in neurons that are distinct from those of CaM, we performed a yeast two-hybrid screen to identify specific interaction partners for the C-terminal half of Caldendrin.
This strategy disclosed a novel Caldendrin-binding partner, named Jacob, which exhibits a remarkably restricted expression in cortical and limbic brain regions of mammals. We report that Jacob displays a distribution similar to that of Caldendrin in the PSD, dendritic spines, and dendrites, but in contrast to Caldendrin, is also found in neuronal nuclei. Activation of NMDA receptors induces nuclear trafficking of Jacob that is under the control of Caldendrin and Importin-α. Our data imply that nuclear Jacob participates in the CREB shut-off pathway, which might play a physiological as well as pathophysiological role in the control of dendritic cytoarchitecture, synapse number, and neuronal survival under conditions of increased NMDA receptor activity.
Utilizing yeast two-hybrid screening to identify binding partners for Caldendrin, we obtained eight independent clones of a hitherto uncharacterized gene product, which we termed Jacob. The longest cDNA clone encompassed an open reading frame of 1,596 bp encoding a 532–amino acid (aa) protein with a calculated molecular weight of 60 kDa (Figures 1 and S1A). Several clones obtained from rat brain cDNA libraries reveal alternatively spliced Jacob transcripts (Figure S1B), giving rise to multiple Jacob isoforms with apparently different molecular weights.
Analysis of the primary structure of Jacob revealed a putative N-terminal myristoylation site and several potential phosphorylation sites for protein kinase C (PKC), cAMP-/cGMP-dependent protein kinases, and protein tyrosine kinases (Figures 1 and S1A). In addition, Jacob harbours a well-conserved bipartite nuclear localization signal (NLS). Interestingly, this NLS is part of an incomplete IQ motif (Figures 1 and S1A), a protein–protein interaction region characteristic for CaM binding [20–21].
In situ hybridization experiments revealed a strikingly restricted localization of Jacob transcripts in the limbic brain and cortical areas (unpublished data), showing extensive overlap with Caldendrin mRNA expression (see [19]). In accordance with in situ hybridization data, Jacob immunoreactivity (IR) was found predominantly in cortex and limbic brain structures, including the amygdala, the thalamus, and the hippocampus (Figure 2A). At the cellular level, particularly intense immunostaining was observed in the somatodendritic compartment of pyramidal cells in cortex (Figure 2B and 2C) and hippocampus, which closely resembles that seen for Caldendrin [16,19]. Moreover, both proteins extensively colocalize in hippocampal primary neurons (Figure S2).
At the ultrastructural level, Jacob IR was localized to a subset of asymmetric type I synapses on dendrites of cortical neurons (Figure 2F–2H). Apart from its synaptic localization, intense label was present in patches in dendrites (Figure 2E). In these patches, Jacob IR was mainly concentrated at the cortical cytoskeleton. In contrast to Caldendrin, intense Jacob immunolabelling was also seen in neuronal nuclei (Figure 2D). Patchy IR was found both at the nuclear envelope and in the nuclear matrix. Subcellular fractionation experiments confirmed that Jacob is a synaptic and a nuclear protein. Differential centrifugation of brain protein fractions demonstrated that Jacob IR is associated with particulate fractions, including light and heavy membranes, and is prominently present in synaptosomes, synaptic junctional membranes, and the PSD fraction (Figure 3A). Jacob, like Caldendrin [16], is tightly associated with the PSD, since extensive Triton X-100 extraction resulted in a further relative enrichment of Jacob IR in the detergent-extracted PSD fraction. Interestingly, immunoblots of the crude nuclear fraction demonstrated the presence of prominent Jacob IR bands in the range of 62–70 kDa, whereas the major bands detected in PSD preparations migrated at 72–80 kDa in SDS-PAGE (Figure 3A and 3B), a difference that most likely reflects posttranslational modification(s).
The architecture of the nucleus includes two overlapping nucleic acid–containing structures that are directly associated with the regulation of gene expression: the chromatin and the nuclear matrix. Therefore, we isolated highly pure nuclear matrix, heterochromatin, and euchromatin fractions. Interestingly, after chromatin fractionation, Jacob was found to be exclusively associated with the RNA polymerase II–containing euchromatin (Figure 3C). Moreover, the Jacob-containing protein complexes immunopurified from euchromatin also contained significant amounts of DNA (Figure 3D). In addition, the initial purification of the crude nuclear fraction showed an enrichment of the 62–70 kDa Jacob bands in the nuclear matrix (D. C. Dieterich and M. R. Kreutz, unpublished data). This subcellular localization could be confirmed by subsequent extraction of nuclear protein components, including chromatin from COS-7 cells transfected with a green fluorescent (GFP)-Jacob fusion protein (D. C. Dieterich and M. R. Kreutz, unpublished data). These findings suggest that Jacob is highly enriched at active sites of nuclear gene transcription and mRNA processing.
Jacob harbours an N-terminal myristoylation consensus motif. Transfection of HEK-293 cells cultivated in the presence of 3H-myristic acid with a wild-type (WT)-Jacob-GFP construct, subsequent immunoprecipitation with a monoclonal anti-GFP antibody and immunoblotting revealed incorporation of radioactivity at a band immunoreactive to both a polyclonal GFP antibody (Figure 3E) and Jacob antiserum (unpublished data). No incorporation was seen in controls transfected with GFP alone or with a myristoylation mutant, ΔMyr-Jacob-GFP, in which the crucial glycine at position 2 was mutated to alanine (Figure 3E). In contrast to the WT-Jacob-GFP construct, transient transfection of COS-7 cells with the ΔMyr-Jacob-GFP construct led to an exclusive nuclear localization of the mutant protein (Figure 3F).
Jacob's primary structure exhibits a well-conserved bipartite NLS between aa 250–265. To test for the functionality of this NLS, we generated a deletion mutant (ΔNLS-Jacob-GFP) lacking the six basic amino acid residues between 247–252. Transfection of this construct in COS-7 cells resulted in an extranuclear localization of the mutant protein (Figure 3F). Hence, the bipartite NLS seems to be necessary and sufficient for nuclear import of Jacob as the double mutant ΔNLS/ΔMyr-Jacob-GFP is extranuclear in transfected COS-7 cells.
To elucidate functional consequences of nuclear versus extranuclear localization of Jacob in terms of structural plasticity, we transfected hippocampal primary neurons with different mutant (ΔNLS-Jacob-GFP or ΔMyr-Jacob-GFP) or WT-Jacob constructs. Transfection of these different mutants had drastic effects on cell morphology. WT-Jacob-GFP–transfected neurons as compared to GFP controls exhibited more, but less-complex, dendritic processes (Figures 4A–4C and S3A). This effect was astonishingly rapid and was observed already after 6–12 h post-transfection. In sharp contrast to the WT-Jacob-GFP overexpression phenotype, ΔMyr-Jacob-GFP–transfected cells lost most of their dendritic processes within 12 to 24 h (Figure 4A–4C). In these cells, the construct exclusively accumulated in the nucleus (Figure 4A). Interestingly, the density of synaptic puncta was already reduced before the retraction of dendrites became visible, observed as early as 6 h following transfection (Figure 4D and 4E). No effect on synapse number of WT-Jacob-GFP was seen even 24 h after transfection (Figure 4E). This strongly suggests that the simplification of the postsynaptic receptive units precedes the retraction of the dendrite. The opposite was found after a lentiviral RNA interference (RNAi)-based knockdown of the nuclear Jacob isoforms harbouring exon 6 with the NLS (RNAi-NLS-GFP). Quantitative immunoblot analysis and immunostainings showed that viral infection of cortical primary cultures led to a specific reduction of these nuclear Jacob isoforms (Figure 4F, 4G, and 4I). Infected cells showed a slightly increased number of synapses and a more complex dendritic cytoarchitecture (Figure 4H, 4J, and 4K). Similarly, overexpression of ΔNLS-Jacob-GFP caused an increase in the number of dendrites, but had no effects on the number of synapses (unpublished data), providing further evidence that the reduction of synaptic contacts directly correlates with the presence of Jacob in the nucleus.
Alternative splicing generates splice isoforms like Δex9-Jacob that contain the NLS but lack large parts of the carboxy-terminus (Figure S1B). We generated a myristoylation-deficient construct of this isoform (ΔMyr-Δex9-Jacob-GFP), which accumulated in the nucleus after transfection of primary neurons, and its overexpression resulted in a comparable reduction of dendritic complexity and synapse number as those seen with ΔMyr-Jacob-GFP (Figure 4A–4E). On the other hand, overexpression of a construct lacking the first 235 amino acids (C-term-Jacob-GFP) had no morphological consequences despite the presence of the NLS and its exclusive nuclear localization (Figures 4E and S3). Taken together, these data provide strong evidence that the reduction of synaptic contacts directly correlates with the presence of Jacob in the nucleus and that the N-terminal half and the NLS are pivotal for Jacob's morphogenetic impact on dendritic architecture.
Since the NLS is not only essential, but also sufficient to target Jacob fragments to the nucleus, and the N-terminal half of the protein is crucial and sufficient to elicit the strong pleiomorphic negative effects on neurite and synapse number of nuclear Jacob, we next investigated the Caldendrin binding region in Jacob and vice versa, as well as the functional consequences of Caldendrin binding in more detail. Mapping of binding domains within both proteins was performed using deletion constructs for cotransformation in yeast two-hybrid assays. In Caldendrin, the region containing the first and the second, probably cryptic, EF-hand was found to be essential for Jacob binding (Figure 5A). Strikingly, in Jacob, we could map the Caldendrin binding region to the central α-helical region that harbours the bipartite NLS. Deletion of the first six basic residues of the NLS led to significantly reduced Caldendrin binding (Figure 5A).
To substantiate the yeast two-hybrid data, we verified the interaction between Jacob and Caldendrin, employing pull-down assays from brain tissue using a glutathione S-transferase (GST)-Caldendrin fusion protein. In this assay, the interaction of Caldendrin and Jacob was found to be Ca2+ dependent inasmuch as 1 μM free Ca2+ in the buffer was required to pull down recombinant Jacob (Figure 5B). Strikingly, CaM did not bind to Jacob at any Ca2+ concentration tested (Figure 5C), and CaM did not compete with GST-Caldendrin for binding to Jacob (Figure 5D). Further evidence for a bona fide interaction of the two proteins was provided by the binding of Jacob to an anti-Caldendrin antibody column and vice versa (Figure 5E). These findings are consistent with the colocalization of Jacob and Caldendrin in dendrites and dendritic spines in hippocampal primary neurons observed by confocal laser scans (Figure S2B). Importantly, the coimmunoprecipitation of Caldendrin from rat brain required the presence of Ca2+ and was not visible after addition of EGTA to the precipitation buffer (Figure 5F). Thus, the interaction of Caldendrin and Jacob in vitro and in vivo is strictly Ca2+ dependent.
Three-dimensional modelling substantiated the idea that structural differences in Jacob binding surfaces of Caldendrin and CaM [18] can explain the above findings. In crystallographic structures of CaM-peptide complexes, a helical peptide binds to one or both hydrophobic pockets formed by pairs of either EF-hands 1 and 2 or EF-hands 3 and 4 of CaM. The residues forming the two hydrophobic binding pockets are identical in CaM and Caldendrin. However, the binding affinities for the various peptides are given by the size and shape of the pocket, which has been classified as open, semi-open, or closed [22]. This size is dynamically regulated by the Ca2+ binding states of the EF-hands, which are not rigid structural units, but may function as hinges at low Ca2+ concentration [23]. Indeed, the second EF-hand of Caldendrin is incapable of binding Ca2+, which in turn results in different binding dynamics at the first hydrophobic pocket compared to CaM.
Interestingly, the Jacob sequence does not match any pattern for a typical CaM binding site, but has two incomplete IQ motifs (residues 237–247, aisvfRGyaeR, and residues 260–269, IQrnfRkhlr) within a central region (residues 229–272) containing the bipartite NLS (residues 247–266, PSORTII, [24]). Several CaM binding sequences within α-helices have been defined that contain an IQ motif [20–21,25], which was originally suggested to classify Ca2+-independent peptide binding according to the referential CaM/IQ complex. Modelling of the Caldendrin interaction site (Figure 5G) suggested that the phenylalanine at position 241 is absolutely essential for the protein interaction, and we therefore generated a point mutation at this position (substitution of F to E). This F241E mutant construct indeed showed no interaction with Caldendrin in yeast (Figure 5A), supporting the assumption that the first part of the central α-helix overlapping with the NLS is the Caldendrin binding region. This α-helical region fits neatly into the hydrophobic pocket generated by EF-hands 1 and 2 and resembles binding to the Ca2+ sensor protein in the manner of an “open” hydrophobic pocket as closely related to the binding of CaMKII to CaM, than the “semi-open” binding in MyosinI/CaM. Moreover, this largely excludes the second incomplete IQ motif as an active binding site for Caldendrin. Accordingly, mutations of residues 260 and 261, IQ to GG, yielded only minor differences in the binding properties of Jacob to Caldendrin in yeast (Figure 5A). In summary, it is predicted that Caldendrin will bind to comparable structures in the presence of a large excess of CaM, suggesting that the Caldendrin–Jacob interaction has evolved independently of CaM-signalling pathways. Even more interesting, binding of Caldendrin to Jacob can be predicted to reduce or inhibit the accessibility of the NLS.
The transport of proteins from the cytosol through the nuclear pore complex into the nucleus depends on the binding of Importins to a specific NLS within the cargo. Within this scheme, Importin-α functions as an adapter molecule by binding both the NLS-bearing protein and Importin-β. Structural modelling suggests that Caldendrin binding will potentially occupy Jacob's NLS, thereby masking this binding site for interaction partners that are likely involved in Jacob's nuclear localization. We tested this hypothesis first by confirming an interaction of Jacob with Importin-α. Coimmunoprecipitation of Importin-α1 from rat cortex indeed suggests a potential in vivo interaction of both proteins (Figure 5H). In pull-down experiments, we found specific binding of myc-his–tagged WT Jacob, but not of the ΔNLS-Jacob mutant to GST-Importin-α1 (Figure 5I). The binding of GST-Importin-α1 was not affected by the presence or absence of Ca2+ (unpublished data). We next investigated whether the binding of Importin-α1 can be competed by equimolar amounts of recombinant Caldendrin. Indeed, these studies revealed a competition between Caldendrin and Importin-α1 for binding to Jacob in the presence of Ca2+ (Figure 5J). Interestingly, no competition was seen in the presence of EGTA, suggesting that elevated Ca2+ levels are needed for Caldendrin to mask the NLS in Jacob.
An elegant recent study established a role of the classical Importin-mediated nuclear import for synapse-to-nucleus communication [26]. In this study, translocation of Importin-α1 and -α2 from distal dendrites to the nucleus was observed requiring NMDA receptor activity. Under resting conditions, however, dendritic Importins are largely immobile. Potential cargos associated with this translocation are at present unknown. Since Jacob is localized both to synapses and the nucleus, and harbours a bipartite NLS, which is bound by Importin-α1 and masked in a Ca2+-dependent manner by Caldendrin, we initially tested whether increased NMDA receptor activity will alter the intracellular localization of Jacob. For this purpose, we stimulated hippocampal primary cultures with NMDA for 3 min and quantified for endogenous Jacob the IR fluorescence signal intensity of propidium iodide–counterstained neuronal nuclei. Jacob IR increased significantly in neuronal nuclei within 30 min after NMDA receptor activation, with highest levels after 2 h (Figure S4A). Nuclear Jacob IR returned to control levels within 4 h (Figure S4A). As previously reported [26], Importin-α1 accumulates in the nucleus in a similar time frame. Interestingly, no recruitment of Caldendrin to the nucleus was observed (unpublished data). Bath application of glutamate led to a significantly increased nuclear accumulation of Jacob IR within a comparable time frame to NMDA receptor activation (Figure S4B). This accumulation could be completely blocked by coincubation of the competitive NMDA receptor antagonist DL-APV (DL-2-amino-5-phosphonopentanoic acid), indicating that activation of NMDA receptors is crucial for glutamate-induced recruitment of Jacob to neuronal nuclei (Figure S4B). To exclude the possibility that stimulation of primary neurons alters the accessibility of the nuclear Jacob antigen to the antibody, we performed quantitative western blot analysis on neuronal nuclei from organotypic hippocampal slice cultures stimulated with the same protocol. These experiments showed a significant increase in intensity of the two major Jacob nuclear isoforms (62 kDa/70 kDa) 2 h after stimulation (Figure S4C and S4D). Moreover, application of anisomycin after NMDA receptor activation did not affect the increase of nuclear Jacob IR, indicating that a recruitment of already existing extranuclear protein underlies the increased Jacob levels in hippocampal nuclei, but not de novo protein synthesis (Figure S4C and S4D).
Using quantitative fluorescence time-lapse microscopy of hippocampal primary neurons transfected with WT-Jacob-GFP or the ΔNLS mutant, we found that the presence of the NLS is essential for the nuclear translocation of Jacob. Glutamate stimulation of WT-Jacob-GFP–transfected cultures kept in the presence of anisomycin resulted in an increase of somatic and nuclear GFP fluorescence with a time course comparable to that of the endogenous protein (Figure 6A–6D). The nuclear accumulation of Jacob-GFP, however, was not seen in neurons transfected with the ΔNLS mutant Jacob-GFP construct (Figure 6E and 6F), suggesting that the presence of the binding site for Importin-α1 is a prerequisite for Jacob's nuclear accumulation. Importantly, concomitant to the nuclear accumulation of WT Jacob, the GFP fluorescence decreased in proximal and distal dendrites (Figure 6C), an effect that was absent in ΔNLS-Jacob-GFP–transfected neurons. This indicates that the presence of the NLS and the interaction with Importin-α1 are not only important for the nuclear import, but are also crucial for Jacob's transport from dendrites to the nucleus.
To learn more about the role of Caldendrin for the extranuclear retention of Jacob, and to understand the apparently contradictory findings (i.e., NMDA receptor activation with subsequent Ca2+ influx leading to Jacob's nuclear import and concomitantly Caldendrin binding preventing this process at high synapto-dendritic Ca2+ levels), we analyzed the transport process of Jacob in more detail using confocal laser scan microscopy. A brief depolarization of hippocampal neurons with KCl also induced a translocation of Jacob and Importin-α1 to the nucleus in the presence of anisomycin (Figure 7A and 7B). However, the nuclear accumulation of both proteins was less pronounced than after glutamate receptor stimulation. Importantly, this effect was completely abolished in the presence of the NMDA receptor antagonist DL-APV (Figure 7A and 7B), indicating that raising intracellular Ca2+ levels by other means than NMDA receptor activation is not sufficient to drive Jacob and Importin-α1 into the nucleus.
NMDA receptors are situated both at synaptic and extrasynaptic sites [27,28]. Bath application of NMDA is considered to affect preferentially, but not exclusively, extrasynaptic NMDA receptors [6,7]. To differentiate between these two populations, we indirectly stimulated hippocampal cultures by incubation with the GABAA receptor antagonist bicuculline. The blockade of inhibitory synapses leads to an increased release of glutamate at synaptic sites, and resulted as expected in an increased accumulation of Jacob and Importin-α1 in the nucleus (Figure 7C and 7D). This effect, however, was much less distinct as compared to the bath application of NMDA. A co-incubation with the noncompetitive NMDA receptor antagonist MK-801 attenuated the nuclear accumulation of Jacob and Importin-α1 to levels indistinguishable from control conditions (Figure 7C and 7D). Since MK-801 is an irreversible open channel blocker, we took advantage of this fact to differentiate between synaptic and extrasynaptic NMDA receptors. After washout of the drug following stimulation of synaptic glutamate receptors, we applied NMDA to the bath solution to exclusively activate extrasynaptic NMDA receptors. Interestingly, this regime induced a marked nuclear translocation of Jacob and Importin-α1 that was more prominent than the accumulation after stimulation of synaptic NMDA receptors (Figure 7E and 7F). Synaptic NMDA receptors contain predominantly the NR2A subunit, whereas their extrasynaptic counterparts contain mainly the NR2B subunit [29]. To test the hypothesis that the nuclear translocation of Jacob and Importin-α1 requires activation of NR2B-containing NMDA receptors, we repeated the experiments outlined above in the presence of the NR2B-specific antagonist ifenprodil. Intriguingly, in the presence of ifenprodil, the nuclear import of Jacob and Importin-α1 could be completely blocked after bath application of NMDA (Figure 7G and 7H). These results show that the nuclear import of these two proteins requires signalling via the largely extrasynaptically localized NR2B-containing NMDA receptors.
To follow up this hypothesis in more detail, we transfected a GFP-Caldendrin construct into hippocampal primary neurons. Expectedly, overexpression of Caldendrin blocked the increase of endogenous nuclear Jacob IR after synaptic stimulation at day in vitro (DIV) 16, indicating that the interaction with Caldendrin masks the bipartite NLS of Jacob (Figure 8A and 8B). However, after stimulation of extrasynaptic NMDA receptors, overexpression of Caldendrin attenuated Jacob's nuclear import much less efficiently (Figure 8A and 8B). We therefore checked whether RNAi knockdown of Caldendrin (Figure S5) affects the nuclear trafficking of Jacob differentially after synaptic and extrasynaptic NMDA receptor stimulation. We found that the nuclear immunofluorescence for Jacob was significantly increased in cells with reduced Caldendrin levels after enhancing synaptic activity with bicuculline at DIV 16 (Figure 8C–8E), whereas the Caldendrin knockdown had no effect on Jacob's nuclear import after activation of extrasynaptic NMDA receptors. This points to a regulatory function of this protein–protein interaction in nuclear trafficking of Jacob after enhanced synaptic activation that is related to the competitive accessibility of Jacob's NLS for either Caldendrin or Importin-α binding.
The predominant Ca2+- and NMDA receptor–activated signalling pathways to the nucleus in neurons funnel through the activation of the transcription factor CREB [1–2]. Previous work has shown that extrasynaptic NMDA receptor activation results in a dephosphorylation of CREB at Ser133 (pCREB) that renders it transcriptionally inactive and, therefore, constitutes a CREB shut-off signal [7,30]. Because Jacob was most efficiently targeted to neuronal nuclei after extrasynaptic NMDA receptor activation, we next addressed the question of whether the presence or absence of Jacob in the nucleus affects the phosphorylation of CREB at this crucial serine residue. As a first proof of principle, we explored whether nuclear overexpression of the ΔMyr-Jacob-GFP construct significantly reduced the levels of pCREB in hippocampal primary neurons as compared to untransfected or GFP-transfected controls under resting conditions (Figure 9A and 9B). Indeed, infection of cortical primary cultures with a Semliki Forest virus–expressing ΔMyr-Jacob-GFP led to drastically reduced pCREB levels as evidenced by quantitative immunoblotting, whereas total CREB levels were not affected (Figure 9C and 9D). To more rigorously test the hypothesis that Jacob is part of the CREB shut-off signalling pathway, we induced a knockdown of nuclear Jacob using plasmid-based RNAi constructs targeting exon 6–containing isoforms of the protein and subsequently stimulated extrasynaptic NMDA receptors with the protocol outlined above. We found that nuclear knockdown of Jacob completely abolished the reduction of pCREB observed after stimulation of extrasynaptic NMDA receptors (Figure 9E and 9F). These data point to a critical role of Jacob for survival of hippocampal primary neurons after triggering the CREB shut-off pathway. We therefore decided to assess next whether the absence of Jacob in the nucleus enhances neuronal survival after triggering CREB shut-off with the stimulation of extrasynaptic NMDA receptors. To this end, we chose in situ TdT-3′end labelling to visualize DNA fragmentation in hippocampal primary neurons as a measure of apoptotic cell death. Using this assay, we found that the number of neurons showing fragmented DNA after sustained extrasynaptic NMDA receptor activation was clearly reduced under conditions of nuclear knockdown of Jacob as compared to untransfected cells from the same plate or independent GFP-transfected controls from other plates (Figure 10A and 10B). Accordingly, the number of condensed propidium iodide–positive nuclei after nuclear knockdown of Jacob was reduced in the same manner as compared to controls (Figure 10C and 10D).
A prominent consequence of bath application of NMDA to primary cultures is the reduction of synaptic contacts within a few hours [34]. Importantly, we found that this reduction requires gene transcription. Coincubation of NMDA with actinomycin-D, an inhibitor of RNA Polymerase II, completely blocked the loss of synaptic contact sites in treated cultures 4 h after stimulation (Figure 11A and 11C). Thus, in line with previous work, loss of synapses appears to be an early event of structural breakdown in cultures treated with bath application of NMDA and requires gene transcription. To further strengthen the point that Jacob is upstream of a transcription-dependent cell death pathway following excessive extrasynaptic NMDA receptor activation, we performed a plasmid-based RNAi knockdown of nuclear Jacob isoforms. Using this approach, we found that the knockdown of Jacob in the nucleus not only prevented CREB shut-off, but also preserved the structural integrity of transfected neurons. As evidenced by immunostainings with the presynaptic marker bassoon quantified 4 h after stimulation, the number of synapses in cultures transfected with RNAi targeting of nuclear Jacob isoforms was essentially the same compared to cultures transfected with a control RNAi vector (scrRNA-GFP). Most importantly, however, after bath application of NMDA, the number of synapses dropped only in cultures transfected with the control vector but not in Jacob RNAi-transfected cultures (Figure 11B and 11D). Thus, the absence of nuclear Jacob prevents not only CREB shut-off and subsequent neuronal degeneration, but also early events of structural disintegration related to loss of synaptic input.
In this study, we identified a novel neuronal protein pathway that is well suited to couple NMDA receptor signalling to the cell nucleus and to trigger long-lasting changes in the cytoarchitecture of dendrites and the number of spine synapses. This novel pathway particularly couples activation of NR2B-containing NMDA receptors to morphogenetic signalling via the nuclear trafficking of Jacob. At resting conditions, Jacob is attached to extranuclear compartments in an either Importin-α bound or unbound state (see also Figure 12). Ca2+ influx through synaptic and extrasynaptic NMDA receptors is followed by a translocation of Importin-α from synapses and dendrites to the nucleus, and we propose that Importin-α–bound Jacob will be concomitantly recruited to the nucleus. Moreover, the presence of the NLS is essential for Jacob's translocation, indicating that trafficking from dendrites to the nucleus and not only nuclear import already requires the classical Importin pathway. This is reminiscent of previous data showing NMDA receptor-dependent Importin trafficking from dendrites to the nucleus [26], and establishes Jacob as the first identified cargo of this trafficking event. Accordingly, we always found a tight correlation between Jacob's and Importin-α1 nuclear translocation. Caldendrin binding can mask the bipartite NLS of Jacob in competition with Importin-α and thereby prevent its nuclear trafficking (Figure 12). However, in contrast to Importin-α binding, this requires high Ca2+ levels and not only NMDA receptor activation (Figure 12). We propose that Caldendrin will target Jacob to spine synapses after enhanced synaptic activation (Figure 12). In support of this hypothesis, we could provide evidence that activation of NR2B-containing NMDA receptors, which are mainly located at extrasynaptic sites, is crucial for the nuclear import of Jacob and Importin-α1. Interestingly, we found that blocking this receptor subtype did attenuate the nuclear accumulation of both proteins after stimulation of synaptic NMDA receptors that contain predominantly, but not exclusively, the NR2A subunit [29–31,35–38]. This suggests the intriguing possibility that the nuclear Jacob-Importin pathway is physically coupled to NR2B-containing NMDA receptors and that the presence or absence of Ca2+-bound Caldendrin in the respective synapto-dendritic compartment will decide whether local Jacob shuttles to the nucleus or not.
On the basis of the characteristics and consequences of its nuclear import, we found conclusive evidence that Jacob is part of the CREB shut-off pathway. The most prominent nuclear target of neuronal NMDA receptor signalling is the transcription factor CREB [1–2,8–9]. Subsequent to its phosphorylation at serine 133, pCREB triggers gene expression crucially involved in processes of synaptic plasticity and neuronal survival [8–9]. Analysis of this pathway has demonstrated that synaptic NMDA receptors strongly activate CREB-dependent gene expression, whereas extrasynaptic NMDA receptors trigger a CREB shut-off [7]. A most intriguing finding in recent years has been that the antagonistic signalling of extrasynaptic versus synaptic NMDA receptors resembles their opposing actions on the activation of ERK kinase [6,39–41]. Activation of synaptic NMDA receptors is coupled to the Ras-ERK pathway and subsequent CREB phosphorylation, whereas extrasynaptic NR2B-containing receptors promote dephosphorylation and inactivation of the Ras-ERK-pathway [6,39–41]. One caveat of this scenario, however, is that shutting down Ras-ERK alone cannot explain the shut-off of CREB since other mechanisms, and here prominently nuclear CaMK-IV, should be in principal sufficient to phosphorylate CREB in the absence of ERK activity [8–9]. Thus, the opposing influence of both types of NMDA receptors after bath application of NMDA requires another mechanism that will actively trigger CREB shut-off. Our data suggest that the same conditions that trigger shut-off of CREB and the Ras-ERK pathway drive Jacob into the nucleus. Overexpression of Jacob in the nucleus—without activating these pathways—is sufficient to attenuate CREB phosphorylation, and a nuclear knockdown of Jacob prevents CREB shut-off as well as neuronal cell death after triggering the pathway. Finally, the rapid loss of synaptic contacts, one of the hallmarks of bath application of NMDA in hippocampal primary cultures, was prevented by reducing the amount of nuclear Jacob. Noteworthy in this regard is the observation that CREB shut-off cannot be induced in young cultures (<DIV 7) [30,42], a developmental stage at which Jacob protein levels are very low (unpublished data). We therefore propose that nuclear Jacob is an essential component of CREB shut-off that might be actively involved in rendering CREB in a dephosphorylated state.
What is Jacob's physiological role in the nucleus? In initial experiments, we could not establish a direct binding of Jacob to CREB although both proteins are found in the overlapping fractions after gel filtration of nuclear protein complexes (unpublished data). Therefore, it is conceivable that Jacob is indirectly coupled via CREB-binding proteins to the CREB signalosome. To further support a role in gene expression, we provided substantial evidence that Jacob is highly enriched in two nuclear compartments associated with gene transcription and pre-mRNA processing. Jacob is abundant in euchromatin fractions and therefore present at active sites of gene transcription. The protein harbours long stretches of basic amino acid residues, which are well suited for DNA binding, although no known DNA binding motif was identified in its primary structure. Particularly with regard to the phenotype of its nuclear overexpression that involves a rapid destabilization of synaptic contacts and a retraction of dendrites, and which cannot be explained entirely by CREB shut-off, it is reasonable to assume that Jacob will be part of additional nuclear signalling events.
The nature of such signalling events will be obviously related to the circumstances of Jacob's nuclear trafficking. CREB shut-off has been largely assigned so far to pathophysiological insults, including spill-over of glutamate after excessive stimulation or reversal of glutamate transporters in the context of epileptic seizures or brain ischemia [3]. This view, however, probably has to be extended because in recent years, a number of observations raise the possibility that the activation of extrasynaptic NR2B-containing NMDA receptors can occur in a physiological context. It was shown that in several brain regions, sustained synaptic activation causes spillover of synaptically released glutamate to nonsynaptic sites [43–49]. In addition, sustained synaptic activation favours nonsynaptic release of glutamate from astrocytes [50–52], and it has been suggested that this glia-neuron transmission via extrasynaptic NMDA receptors has profound effects on non–Hebbian types of neuronal plasticity [53]. Moreover it was also claimed that activation of extrasynaptic NMDA receptors might directly induce heterosynaptic long-term depression at certain synapses in close proximity [54]. The evolving concept behind these studies is the idea of homeostatic scaling of synaptic input. Homeostatic plasticity refers to a process by which principal neurons in particular constantly adjust the integration of synaptic input to optimize the contribution of a single synapse with reference to its location in the dendrite and the synchronized activity in a given neuronal network [55,56]. A major aspect of homeostatic plasticity is the fact that uncontrolled potentiation of synapses will induce a ceiling effect characterized by epileptic activity and a decoupling of a given neuron from the dynamics of presynaptic input. Homeostatic plasticity reflects the necessity to either remove certain synapses that contribute less efficiently to the optimal activity within a neuronal network or to reduce the level of potentiation of synapses in this network. Jacob's nuclear accumulation and its rapid morphogenetic effects are in favour for a role in the regulation of plasticity-related gene expression related to homeostatic synaptic plasticity. Interestingly, this role includes a stripping of synaptic contacts that precedes the simplification and regression of dendritic processes. It is therefore conceivable that the loss of synapses is the initial trigger for the retraction of dendritic arbors. Moreover, this process is surprisingly rapid, indicating that synapses are actively destabilized. This in turn suggests that Jacob either blocks an essential nuclear signalling event required to prevent the removal of synaptic input or regulates the expression of genes that will actively destabilize synapses. It is likely that the CREB shut-off pathway will be part of this mechanism, but it is unclear whether it is sufficient to trigger solely the course of events following Jacob's nuclear import.
A further intriguing aspect of this study is that it provides the first demonstration that an EF-hand CaM-like Ca2+ sensor protein regulates the nuclear localization of a protein by competitive binding to its NLS in a Ca2+-dependent manner. The significance of this novel mechanism of neuronal Ca2+ signalling is further underscored by the fact that binding of Caldendrin is specific in that its ancestor and closest relative in brain, CaM, did not bind to Jacob at any Ca2+ concentration tested. This is of importance since CaM levels are probably more than a magnitude higher in neurons than those of Caldendrin [18], and Ca2+ binding affinities are comparable between both proteins [57]. Computer modelling based on templates from crystallized structures shows that the outer surface of solvent-exposed amino acids, particularly EF-hand 2, which seems to be crucial for binding to Jacob, and another recently identified binding partner light chain 3 (LC3) [58] are covered by residues that clearly differ between CaM and Caldendrin [18]. Accordingly, LC3, a component of the microtubular cytoskeleton, apparently does not bind to CaM [58]. The principal specificity of Caldendrin protein interactions is further supported by the observation that very few mutations occurred in this region during vertebrate development and that none of these mutations affected the solvent-exposed amino acids of EF-hand 2 [18]. Thus, the singularity of the Caldendrin surface is intrinsic and independent from insertions or deletions, and we therefore suggest that this is probably due to adaptations of its surface to a specific localization and function in neurons of higher vertebrates.
How could this singularity with respect to other Ca2+ binding proteins relate to Caldendrin's neuronal function? In contrast to the interaction with Jacob, Caldendrin binding to most of its interaction partners is Ca2+ independent, as already described above for the LC3 interaction [57]. For instance, Ca2+-, CaM-, and ATP-independent interaction of the C-terminal half of Caldendrin/CaBP1 was demonstrated for the inositol trisphosphate receptor ((InsP3R) [59–60]. The functional consequence of Caldendrin binding is a reduction of InsP3-triggered intracellular Ca2+ release [59,60]. At the synapse, a Ca2+-independent binding was reported for L-type voltage-dependent CaV1.2 Ca2+ channels [61,62]. This interaction will probably lead to increased Ca2+ currents following synaptic activation and thereby indirectly via increased synaptic activity could promote Caldendrin's and possibly Jacob's synaptic localization. Low synaptic activity and, hence, low synapto-dendritic Ca2+ levels will instead favour Caldendrin's binding to the InsP3R. It is therefore conceivable that Caldendrin can thereby directly lower Ca2+ levels in dendritic microdomains, and in consequence, negatively regulate its own association with Jacob. Therefore, a switch of binding partners could directly relate to Caldendrin's role in regulating Jacob's nuclear transition. Along these lines, it can be predicted that keeping the delicate balance between Jacob's nuclear and extranuclear localization via Caldendrin binding will provide a powerful regulatory mechanism in the transformation of dendritic Ca2+ signals into morphogenetic signals for the dendritic cytoarchitecture of principal neurons under pathophysiological and probably also under physiological conditions.
Yeast two-hybrid screening was performed as described previously [63]. Library screening was done with a rat brain cDNA library in pACT2 (Matchmaker-GAL4 Two-Hybrid II; Clontech). The bait construct consisted of the entire open reading frame of Caldendrin cloned in frame into the pAS2–1 vector. A total of 3.5 × 106 cotransformants were screened, and 108 clones were picked, which turned blue within 6 h in the initial test and after retransformation. Eight of these clones were found to encode a novel protein. Interactions were scored for ß-galactosidase activity by a colony lift assay. Binding activity of different constructs after retransformation was evaluated in three independent experiments.
A rat brain hippocampus cDNA Lambda ZAPII library (Stratagene) was screened with a cDNA probe encompassing the first 400 bp of Jacob's open reading frame. cDNA labelling, filter hybridization, and subcloning were done using standard procedures [16]. Cloning of full-length murine Jacob was done by reverse transcriptase PCR (RT-PCR) from mouse brain with primers encompassing the start and stop codon of rat Jacob. The PCR product was cloned into a TOPO TA vector (Invitrogen) and sequenced. A list of the constructs employed in this study is provided in Text S1.
Two peptides (aa 285–299 and aa 300–314), the GST fusion proteins GST-J1–230 and GST-J253–404 were used to immunize two rabbits and one guinea pig each. Specificity of the antibodies was tested on immunoblots of crude rat brain homogenate by preabsorption of the antibodies with corresponding N-terminal or C-terminal (J262–532) MBP fusion proteins or with affinity-purified antiserum.
Immunohistochemistry and immunocytochemistry were performed essentially as described previously [19] (see Text S1 for more details). Details of confocal laser scan microscopy and time-lapse imaging experiments can also be found under Supplementary Materials and Methods in Text S1.
See Supplementary Materials and Methods in Text S1 for details.
For RNAi treatment, oligonucleotides with the sense/antisense sequence (19–21 bp) linked by a 9- or 10-bp–long stemloop sequence were obtained from Biomers. Sequences were as follows: nuclear Jacob knockdown (RNAi-NLS: 5′ AGA ATG ATT CCG CGT CTG TAA 3′/bp 892–912 of the Jacob cDNA); nuclear Jacob scrambled control (scrRNA: 5′ AGA TAT AGT CGC CGT CTG TAA 3′); all Jacob isoforms' knockdown (PAN-Jacob: 5′ TGC TAC TAG TTA CAG TGT AGA 3′/bp 390–410 of the Jacob cDNA); all Jacob isoforms' scrambled control (5′ TGA TAG GTC TAT ACG AGT TCA 3′), Caldendrin sRNAi (5′ TCC TGG CGG AGA CAG CAG ATA 3′/bp 665–685 of Caldendrin cDNA); and Caldendrin scrambled (5′ AGA ATC CTA AGA CAA GTG CAG 3′). Forward and reverse oligos were annealed, phosphorylated, and cloned BamHI, HindIII into the pRNAT-H1.1/Neo vector (Genscript) for plasmid-based RNAi knockdown. COS-7 cells were cotransfected with Flag-Jacob or Flag-Jacob-Myc/His and the RNAi expression vector. Cells were harvested 2 d after transfection and the samples solubilised for SDS-PAGE.
For lentiviral transfections, double-stranded, phosphorylated oligos were cloned BamHI/BglII, HindIII into the pZ-off vector and further subcloned EcoRI, AccI/BstBI into the FUGW H1(+) vector. HEK-293T cells were grown on polyD lysine-coated 10-cm3 plates to 90% confluence and cotransfected with the shRNA-FUGW H1(+) (10 μg), the VSVg (5 μg), and Δ8.9 (7.5 μg) vectors using Lipofectamine 2000 according to the manufacturer to produce competent virus particles. For virus production, cells were grown in Neurobasal medium supplemented with GlutaMAX and B27 at 32 °C and 5% CO2 overnight; the medium was changed and virus harvested 48–60 h after transfection. Sterile-filtered virus was directly added to primary cortical neurons at DIV0. After 3 wk, cells were either fixed with PFA for immunostaining or harvested and prepared for SDS-PAGE.
For the preparation of Semliki Forest particles and infection of primary cortical neurons, pSFV-Helper2, pSFV-ΔMyr-Jacob-EGFP, or pSFV-EGFP after in vitro transcription were cotransfected into CHO-K1 cells with Lipofectamine2000 (Invitrogen) according to the supplier's manual. After 24 and 48 h, the culture medium containing the budded particles was harvested. Viral particles were concentrated by ultracentrifugation through 10% sucrose, the pellet was resolved in Tris-buffered solution overnight at 4 °C. Aliquots of the particles were stored at −80 °C after shock freezing. For infection of primary cortical neurons, the particles were activated by chymotrypsin and further diluted with OptiMEM. High-density cortical cultures were infected at DIV16 and harvested 24 h later. Neurons were homogenized in 20 mM Tris buffer containing protease and phosphatase inhibitors, and solubilised in SDS buffer. The protein concentration was determined by amido black test, and equal amounts were loaded for SDS-PAGE.
Statistical analysis was performed with ANOVA and subsequent Bonferroni's Multiple Comparison test. Data are presented as mean ± standard error of the mean (SEM). A level of p < 0.05 was considered statistically significant.
Details about the computer models can be found under Supplementary Materials and Methods in Text S1.
The Protein Data Bank (http://www.pdb.org/pdb/home/home.do) accession number for the structural model discussed in this paper is 1wdc.
|
10.1371/journal.pmed.1002655 | Safety and pharmacokinetics of single, dual, and triple antiretroviral drug formulations delivered by pod-intravaginal rings designed for HIV-1 prevention: A Phase I trial | Intravaginal rings (IVRs) for HIV pre-exposure prophylaxis (PrEP) theoretically overcome some adherence concerns associated with frequent dosing that can occur with oral or vaginal film/gel regimens. An innovative pod-IVR, composed of an elastomer scaffold that can hold up to 10 polymer-coated drug cores (or “pods”), is distinct from other IVR designs as drug release from each pod can be controlled independently. A pod-IVR has been developed for the delivery of tenofovir (TFV) disoproxil fumarate (TDF) in combination with emtricitabine (FTC), as daily oral TDF-FTC is the only Food and Drug Administration (FDA)-approved regimen for HIV PrEP. A triple combination IVR building on this platform and delivering TDF-FTC along with the antiretroviral (ARV) agent maraviroc (MVC) also is under development.
This pilot Phase I trial conducted between June 23, 2015, and July 15, 2016, evaluated the safety, pharmacokinetics (PKs), and acceptability of pod-IVRs delivering 3 different ARV regimens: 1) TDF only, 2) TDF-FTC, and 3) TDF-FTC-MVC over 7 d. The crossover, open-label portion of the trial (N = 6) consisted of 7 d of continuous TDF pod-IVR use, a wash-out phase, and 7 d of continuous TDF-FTC pod-IVR use. After a 3-mo pause to evaluate safety and PK of the TDF and TDF-FTC pod-IVRs, TDF-FTC-MVC pod-IVRs (N = 6) were evaluated over 7 d of continuous use. Safety was assessed by adverse events (AEs), colposcopy, and culture-independent analysis of the vaginal microbiome (VMB). Drug and drug metabolite concentrations in plasma, cervicovaginal fluids (CVFs), cervicovaginal lavages (CVLs), and vaginal tissue (VT) biopsies were determined via liquid chromatographic-tandem mass spectrometry (LC-MS/MS). Perceptibility and acceptability were assessed by surveys and interviews. Median participant age was as follows: TDF/TDF-FTC group, 26 y (range 24–35 y), 2 White, 2 Hispanic, and 2 African American; TDF-FTC-MVC group, 24.5 y (range 21–41 y), 3 White, 1 Hispanic, and 2 African American. Reported acceptability was high for all 3 products, and pod-IVR use was confirmed by residual drug levels in used IVRs. There were no serious adverse events (SAEs) during the study. There were 26 AEs reported during TDF/TDF-FTC IVR use (itching, discharge, discomfort), with no differences between TDF alone or in combination with FTC observed. In the TDF-FTC-MVC IVR group, there were 12 AEs (itching, discharge, discomfort) during IVR use regardless of attribution to study product. No epithelial disruption/thinning was seen by colposcopy, and no systematic VMB shifts were observed. Median (IQR) tenofovir diphosphate (TFV-DP) tissue concentrations of 303 (277–938) fmol/106 cells (TDF), 289 (110–603) fmol/106 cells (TDF-FTC), and 302 (177.1–823.8) fmol/106 cells (TDF-FTC-MVC) were sustained for 7 d, exceeding theoretical target concentrations for vaginal HIV prevention. The study’s main limitations include the small sample size, short duration (7 d versus 28 d), and the lack of FTC triphosphate measurements in VT biopsies.
An innovative pod-IVR delivery device with 3 different formulations delivering different regimens of ARV drugs vaginally appeared to be safe and acceptable and provided drug concentrations in CVFs and tissues exceeding concentrations achieved by highly protective oral dosing, suggesting that efficacy for vaginal HIV PrEP is achievable. These results show that an alternate, more adherence-independent, longer-acting prevention device based on the only FDA-approved PrEP combination regimen can be advanced to safety and efficacy testing.
ClinicalTrials.gov NCT02431273
| Pre-exposure Prophylaxis (PrEP) for HIV prevention has been shown to be effective in highly adherent users.
Preclinical studies of an innovative pod-intravaginal ring (IVR) design delivering antiretroviral (ARV) regimens have shown protection against Simian Human Immunodeficiency Virus (SHIV) in nonhuman primates.
The goal of this clinical study was to assess safety, pharmacokinetics (PKs), and acceptability of pod-IVRs delivering 3 different ARV regimens: 1) tenofovir disoproxil fumarate (TDF) only, 2) TDF and emtricitabine (FTC), and 3) TDF, FTC, and maraviroc (MVC) via the vaginal route.
A 7-d vaginal ring study was performed, sequentially delivering TDF, TDF-FTC, and TDF-FTC-MVC in 6 women to determine safety and PKs.
There were no concerning safety findings that would inhibit further development of the vaginal rings.
Vaginal secretion and tissue levels were detected in ranges that have been associated with protection against HIV/SHIV infection in preclinical and clinical studies.
Vaginal delivery of combinations of ARV drugs, including the only Centers for Disease Control (CDC)-recommended regimen for HIV PrEP, was achieved in this early Phase I clinical study.
The interventions appeared safe and acceptable and therefore merit further study.
| Scale-up of prevention and treatment efforts to curb the HIV epidemic have resulted in decreasing the number of new infections by half per year in 2012 since the peak in 1996 [1]. In Fast-Track, the Joint United Nations Programme on HIV/AIDS (UNAIDS) has set aggressive goals, including 500,000 (or fewer) new annual infections by 2020, a 75% reduction from 2010 numbers, and an end to the AIDS epidemic by 2030 [2]. Unfortunately, the number of annual, new HIV infections has stalled around 1.9 million since 2010, suggesting that a prevention gap has been reached [3]. To meet these ambitious UNAIDS Fast-Track goals, further work is needed to develop new highly effective strategies for HIV prevention.
Pre-exposure prophylaxis (PrEP) using Food and Drug Administration (FDA)-approved antiretroviral (ARV) drugs holds significant promise as a strategy for preventing HIV infection. Multiple HIV PrEP clinical trials have demonstrated that vaginal and oral ARV regimens based on the nucleoside reverse transcriptase inhibitor (NRTI) tenofovir (TFV), alone or in combination with the NRTI emtricitabine (FTC), can be effective in susceptible men, women, and partners of HIV-infected individuals [4–11]. Oral tenofovir disoproxil fumarate (TDF)-FTC (Truvada, Gilead Sciences, Foster City, CA) is the only FDA-approved regimen for HIV PrEP [12]. Centre for the AIDS Programme of Research in South Africa (CAPRISA)-004 demonstrated that topical dosing—in this case pericoitally with a 1% TFV gel—can be effective in preventing vaginal HIV transmission [4]. In post hoc analyses of both Vaginal and Oral Interventions to Control the Epidemic (VOICE) gel and Follow-on African Consortium for Tenofovir Studies (FACTS) 001 trials, TFV gel was effective in highly adherent women [13]. A critical factor driving success in these trials appears to involve sustaining high adherence to frequent dosing [14].
Treatment of any medical condition that requires long-term use of medications must address the challenge to maintain adherence to therapy; adherence has been shown to be inversely related to dosing frequency regardless of dosing formulation [15–18]. Use of extended-release or long-acting formulations, including intravaginal rings (IVRs) [19], implantable devices, and injectable formulations [20], can decrease the burden of frequent dosing and potentially improve adherence. Two Phase 3, randomized, double-blind, placebo-controlled trials (MTN-020–A Study to Prevent Infection with a Ring for Extended Use (ASPIRE) and IPM 027–The Ring Study) recently evaluated a monthly IVR delivering the non-nucleoside HIV reverse-transcriptase inhibitor dapivirine (DPV) [21,22]. The trials enrolled 2,629 and 1,959 women, respectively, between the ages of 18 and 45 years in Malawi, South Africa, Uganda, and Zimbabwe and demonstrated that an IVR delivering an ARV agent can reduce the risk of HIV acquisition by as much as 56% in highly adherent users, as determined through quarterly plasma and residual IVR DPV concentrations.
User adherence to IVRs depends on a number of factors including user sensory perceptions and experiences of product characteristics and impact such as dimensionality and perceptions of increased discharge [23]. Users ascribe meaning to product characteristics, and those meanings can influence perceptions of product efficacy and the user’s willingness to engage in consistent and correct product use [23,24].
The primary purpose of this exploratory, open label clinical trial was to evaluate the safety, pharmacokinetics (PKs), and user perceptibility of an innovative IVR design (termed “pod-IVR”) [25–27] delivering three formulations: first, TDF alone over 7 d; second, TDF-FTC over 7 d in a crossover design; and third, TDF-FTC-maraviroc (MVC) over 7 d. The pod-IVR consists of a silicone scaffold that holds up to 10 individual “pods” of polymer-coated drug cores, allowing independently controlled drug release from each pod through delivery channels.
All human research was approved by the University of Texas Medical Branch Institutional Review Board (IRB #14–0479), conducted according to the Declaration of Helsinki, and registered in clinicaltrials.gov (NCT02431273). All participants provided written informed consent.
Women were recruited through announcements to use IVRs releasing TDF, TDF-FTC, and TDF-FTC-MVC for 7 d each between June 23, 2015, and July 15, 2016. Women were prescreened to confirm basic eligibility and then scheduled for a screening visit where inclusion/exclusion criteria were reviewed, medical and sexual history were obtained, and baseline labs with blood counts, liver and kidney function, and sexually transmitted infection (STI) and HIV screening were collected. Inclusion criteria were women aged 18–45 with regular or suppressed menses, on contraception, and willing to abstain from sexual intercourse during IVR use. Women with STIs, HIV, Hepatitis B, abnormal screening labs, allergy to study product, currently using an IVR, or at high risk for HIV were excluded.
The trial consisted of an open-label, crossover design of a pod-IVR with 3 different ARV regimens. There were 3 separate 7-d treatment periods (Treatment Period 1, TDF-only pod-IVR; Treatment Period 2, TDF-FTC pod-IVR; and Treatment Period 3, TDF-FTC-MVC pod-IVR) with 6 participants per treatment period. Each participant used each pod-IVR for 7 d with a washout period of at least 14 d between each treatment period. Progression to the next treatment period was contingent upon the absence of grade 3 or 4 genitourinary adverse events (AEs) considered to be drug related by the investigator or other investigator-assessed drug-related serious adverse event (SAE). There was a 3-mo pause and re-consent between the second and third treatment periods to allow for determination of safety and PKs of the combination IVR before proceeding to the use of the TDF-FTC-MVC pod-IVR in the third treatment period. The study design was chosen with the ultimate goal of developing a TDF-FTC and TDF-FTC-MVC pod-IVR; however, the FDA required assessment of a TDF-only pod-IVR prior to use of the TDF-FTC and TDF-FTC-MVC formulations.
Following the screening visit, women who were clinically deemed eligible returned to the clinic for IVR insertion (Visit 1, TDF IVR; Visit 5, TDF-FTC IVR) after cessation of menses. Women were instructed on vaginal ring placement, which was performed during this visit, and were asked to be abstinent during IVR use. They returned for Visit 2 (TDF IVR) or Visit 6 (TDF-FTC IVR) on Day 2 (± 1 d) after IVR insertion and for Visit 3 (TDF IVR) or Visit 7 (TDF-FTC IVR) on Day 7 (± 1 d) when the IVRs were removed. Women returned for follow-up 1–2 wk after IVR removal for Visit 4 (TDF IVR) or Visit 8 (TDF-FTC IVR).
The TDF-FTC-MVC pod-IVR arm was carried out separately from the TDF and TDF-FTC IVR portion of the study. Women who were deemed clinically eligible returned to the clinic for IVR insertion (Visit 1) after cessation of menses, instructed on vaginal ring placement (which was performed during this visit), and asked to be abstinent during IVR use. Women returned for Visit 2 on Day 2 (± 1 d) after IVR insertion and for Visit 3 on Day 7 (± 1 d) when the IVRs were removed. Women returned for follow-up 1–2 wk after IVR removal for Visit 4.
During all visits, AEs were reviewed and colposcopy was performed to evaluate the vagina and cervix for safety assessments. For PK assessments, blood (plasma) and cervicovaginal samples (cervicovaginal fluid [CVF], Dacron swabs; cervicovaginal lavage [CVL], 2.5 mL sterile phosphate-buffered saline solution containing 1 mM LiCl) were collected. In summary, for PK assessments, two samples were collected during pod-IVR use (Visit 2/6/2 and Visit 3/7/3) and one sample after pod-IVR removal (Visit 4/8/4). Additionally, at Visit 3/7/3, the IVRs were collected and vaginal biopsies were obtained. One biopsy was flash-frozen in liquid nitrogen for the analysis of ARV drug concentrations. A second vaginal biopsy was fixed in formalin and hematoxylin–eosin stained. The slides were reviewed by a trained pathologist with expertise in reproductive tract pathology.
Behavioral assessments of perceptibility, acceptability, and adherence were obtained through daily diaries during IVR use, surveys completed using computer-assisted self-interview (CASI) format at the time of IVR removal, and qualitative interviews conducted at Visit 4/Visit 8/Visit 4. Baseline demographics and sexual history surveys were completed prior to product initiation. A user sensory perception and experience (USPE) survey, acceptability and adherence measures, and an in-depth interview were completed following IVR use.
TDF and FTC labeled for human use were purchased from commercial vendors with a Drug Master File (DMF) registered with the FDA. MVC was isolated from the commercial formulation (Pfizer, Inc., New York, NY), which consists of film-coated tablets for oral administration containing 300 mg of MVC and inactive ingredients, as described previously [26]. All other reagents were obtained from Sigma-Aldrich, unless otherwise noted.
Polydimethylsiloxane ([PDMS], silicone) pod-IVRs were fabricated in a multistep process that has been described in detail elsewhere [25,26,28,29], and only a brief description is provided here. Each pod contained a single drug. The drug powder was compacted into cores of 3.2-mm outer diameter in a manual tablet press (MTCM-I; Globe Pharma, New Brunswick, NJ). Drug cores were coated with poly(vinyl alcohol) to yield pods, which were placed in the corresponding IVR cavities and sealed in place by back-filling with room-temperature-cure silicone. Each pod was matched with the appropriate configuration of mechanically punched delivery channels. The IVR drug loadings were as follows: TDF, 180 mg; FTC, 125 mg; MVC, 90 mg. Residual drug content of used IVRs was measured by high-performance liquid chromatography (HPLC) according to published methods [26,28] and used to calculate in vivo release rates.
AEs were reviewed at all study visits along with colposcopy. Histology was performed on VT samples collected on Day 7 (Visit 3/7/3). Vaginal pH and Nugent scores [30] were measured at each study visit. Vaginal microbial community profiles also were measured at each study visit by a custom quantitative polymerase chain reaction (qPCR) array described previously [31]. The array targets 46 distinct key vaginal microbiota to the species level along with several housekeeping genes.
Concentrations of TDF, TFV, tenofovir diphosphate (TFV-DP), FTC, and MVC were determined via previously described liquid chromatographic-tandem mass spectrometric (LC-MS/MS) assays [32–35]. All assays were developed and validated following the FDA Guidance for Industry, Bioanalytical Method Validation recommendations and met all acceptability criteria [36]. Isotopically labeled internal standards were used for all compounds and the determination of drug concentrations in all specimen sources.
The lower limits of quantification for these methods were as follows: CVFs, TDF (0.0625 ng/sample), TFV (0.25 ng/sample), FTC (1.0 ng/sample), MVC (0.05 ng/sample); CVL, TDF (0.5 ng/mL), TFV (5.0 ng/mL), FTC (20 ng/ mL), MVC (1 ng/mL); VT homogenate, TFV (0.05 ng/sample), TFV-DP (50 fmol/sample), FTC (0.25 ng/sample), MVC (0.05 ng/sample); plasma, TFV (0.31 ng/mL), FTC (0.31 ng/mL), MVC (0.1 ng/mL). CVF and tissue samples were ultimately reported as ng/mg or fmol/mg, respectively, following normalization to net biopsy or CVF weight. VT TDF concentrations were not measured.
CVLs were performed by instilling a sterile phosphate-buffered saline solution (2.5 mL) containing LiCl (1 mM). The added LiCl allowed dilution of the collected CVF to be calculated in the TDF-FTC-MVC group using an established method [37]. CVL samples were not compensated for dilution in the TDF and TDF-FTC groups.
PK parameter values for CVF were determined by noncompartmental analysis (NCA) using Phoenix WinNonlin 6.4 (Pharsight Corporation, Sunnyvale, CA). The NCA was run using the linear trapezoidal rule for increasing concentration data and the logarithmic trapezoidal rule for decreasing concentration data (linear up and log down) as the calculation method. Post-dose concentrations below the corresponding lower limit of quantification (LLOQ) (CLLQ) were imputed as follows:
CLLQ=LLOQofassay2×(mediansamplemass)
(1)
Qualitative data were analyzed by applying study-specific thematic identification and summarizing verbatim transcript data per theme, with illustrative quotes retained as conventional [38]. Quantitative data were analyzed using GraphPad Prism (version 7.00, GraphPad Software, Inc., La Jolla, CA). Statistical significance is defined as two-sided P < 0.05. The unpaired, nonparametric Mann–Whitney test was used to compare two groups, including vaginal pH, Nugent scores, in vivo TDF release rates, CVF FTC concentrations in the TDF-FTC and TDF-FTC-MVC pod-IVR treatment periods and terminal half-lives of ARV drug elimination from CVF within one IVR treatment period and across IVR treatment periods. The nonparametric (i.e., do not assume a Gaussian distribution) Kruskal–Wallis tests with no matching/pairing of the data was used to compare three IVRs, including CVF TDF and TFV concentrations and CVF:VT concentration ratios and terminal half-lives of ARV drug elimination from CVF within one IVR arm.
Six participants were enrolled and completed the first 2 treatment periods using the TDF-only and TDF-FTC pod-IVRs. Due to the delay between treatment periods 2 and 3, 4 participants dropped out and were replaced for treatment period 3. Demographics for the initial 6 study participants who completed the first 2 treatment periods (i.e., TDF-only pod-IVR and TDF-FTC pod-IVR) can be found in Table 1, and demographics for the final treatment period (i.e., TDF-FTC-MVC pod-IVR) can be found in Table 2. There were no missed study visits, and one participant (ID 479–12) had an additional visit (Visit 2A, TDF pod-IVR arm) with visits on Day 1 (Visit 2) and Day 2 (Visit 2A) for evaluation of AEs (see Safety measures below).
The pod-IVRs were safe and generally well tolerated in all three treatment periods. AEs were recorded with daily diaries and during study visits. There were no concerning safety findings by participant report, examination with colposcopy, evaluation of vaginal microbiome (VMB), or histology from vaginal biopsy. There were no SAEs or grade 3 or 4 genitourinary AEs. Fifty-eight AEs occurred during the study period (54 Grade 1, 4 Grade 2; Tables 3–5). All AEs were Grade 1 except for four Grade 2 findings. Two of the Grade 2 events were pelvic pain/cramping, for which the subjects took over-the counter medication and were determined to be "possibly related." One woman had a Grade 2 candida vaginitis determined to be "possibly related," which did not recur after treatment. One woman reported having Grade 2 diarrhea, which was determined to be "not related" and attributed to “food poisoning”.
The following is a description of findings during IVR use. Thirty-eight AEs occurred during use of the IVRs, 17 during TDF IVR use (Table 3), 9 during TDF-FTC use (Table 4), and 12 during TDF-FTC-MVC use (Table 5). On colposcopy, cervicovaginal erythema was found in 3 women (6 findings) during IVR use, 2 of whom had erythema at baseline. During IVR use, 6 women had pelvic pain/cramping (13 findings). Four women reported vaginal discharge (7 findings); one episode was a mucus-like discharge that appeared peri-ovulatory that was deemed "possibly related." Three women had metrorrhagia/intermenstrual bleeding during IVR use (3 findings), one during TDF IVR use and two during TDF-FTC-MVC use. Three women had vulvovaginal itching (3 findings). Two women were diagnosed with candida vaginitis (one Grade 1 and one Grade 2) and were treated. One woman reported an odor (1 finding) on Day 7 of TDF-FTC-MVC IVR use. One woman reported nausea (1 finding) associated with cramping which lasted only 10 min during TDF-FTC-MVC IVR use. One woman had Grade 1 diarrhea during TDF-only IVR use, which was "possibly related." One woman had malaise after influenza vaccine during TDF-only IVR use, and this was considered "not related."
With two exceptions, all biopsies were reported as “consistent with a normal vaginal epithelium”. For one subject in the TDF pod-IVR treatment period, there was one section of the vaginal biopsy with minimal to mild inflammatory infiltrates and necrosis, which may be consistent with an infection (she had asymptomatic Candida infection determined by qPCR); a recut from the same specimen showed normal tissue. In the TDF-FTC-MVC pod-IVR treatment period, there was a finding of minimal inflammatory infiltrate and hemorrhage in the subject who had the Grade 2 symptomatic Candida infection (Table 5).
In all 3 treatment periods, the impact of the pod-IVRs on the vaginal microbial community profiles was analyzed using custom qPCR arrays targeting the most common 46 vaginal bacteria at genus or species levels [31]. A second array was used to quantify the levels of 16 additional pathogen targets, including bacteria and viruses common to the vagina. Collectively, the results show no clear, systematic impact of the pod-IVRs on the stability of the microbial community profiles for the trial participants (S1–S3 Figs). One subject (ID 479–16) had asymptomatic Candida albicans noted only by qPCR (saline microscopy with potassium hydroxide [KOH] was negative) during the study.
The median vaginal pH prior to IVR placement (Visits 0 and 1) was 4.0 (IQR, 4.0–4.5), and with the TDF and TDF-FTC pod-IVRs in place was 4.5 (IQR, 4.1–4.5) and 4.0 (IQR, 4.0–4.5), respectively. There was no statistically significant difference in vaginal pH prior to pod-IVR placement compared to during pod-IVR use: TDF pod-IVR, P = 0.3405; TDF-FTC pod-IVR, P = 0.5882. The median Nugent score prior to IVR placement (Visits 0 and 1) was 3.0 (IQR, 3.0–3.3) and with the TDF and TDF-FTC pod-IVRs in place was 2.5 (IQR, 2.0–4.0) and 4.5 (IQR, 2.8–5.3), respectively. There was no statistically significant difference in Nugent score prior to pod-IVR placement compared to pod-IVR: TDF pod-IVR, P = 0.6068; TDF-FTC pod-IVR, P = 0.5611.
The median vaginal pH prior to IVR placement (Visits 0 and 1) was 4.1 (IQR, 4.1.–4.5) and with the TDF-FTC-MVC pod-IVRs in place was 4.4 (IQR, 4.1–4.5). There was no statistically significant difference (0.3775) in vaginal pH prior to pod-IVR placement compared to during pod-IVR use. The median Nugent score prior to IVR placement (Visits 0 and 1) was 4.0 (IQR, 2.8–5.3) and with pod-IVRs in place was 5.0 (IQR, 4.0–5.0). These findings could be attributed to the prevalence of Gardnerella vaginalis in some of the participants’ VMBs (S3 Fig). There was no statistically significant difference (P = 0.5858) in Nugent score prior to pod-IVR placement compared to during pod-IVR.
All participants in the TDF and TDF-FTC treatment periods were willing to recommend the ring to others, and 4 of 6 would “probably” or “definitely” use the pod-IVR for HIV prevention. The remaining 2 expressed no perceived risk of HIV acquisition. Participants’ confidence in their ability to insert and remove the IVR either held constant between using the first and second study rings or improved. Participants’ willingness to use the pod-IVR for 28 d and access the ring for “real world” use was high. For example, after using the first ring, participants averaged 3.5 (on a 1–5 Likert scale from “not at all confident” to “completely confident”) that she could use the ring for 28 d, while the average increased to 4.0 after using the second ring. Qualitative interview data revealed that, overall, the pod-IVRs were well tolerated by participants. Prior to insertion of the pod-IVR, clinicians reviewed verbal instructions with the participants, which reportedly allowed women to feel more confident in their ability to insert the pod-IVR on their own: clinician’s checking pod-IVR placement on first insertion also led to greater confidence in some. While 2 of 6 women felt the TDF-FTC pod-IVR was more difficult to insert than the TDF pod-IVR, the majority of participants felt both pod-IVRs were flexible and easy to insert. The majority of women noted that they did not feel the pod-IVR in their vaginas at all during the use period. Of note, psychological awareness of the pod-IVR was reported by one woman, which subsided after a few days. Lack of physical awareness of the pod-IVRs allowed most participants to continue their typical physical activity routines during the weeks of use, with one exception (i.e., an elite runner). All participants felt that the pod-IVRs had no impact on their typical hygiene practices. All participants were able to successfully remove the pod-IVRs, with 3 of 6 women reporting difficulty initially locating the TDF pod-IVRs in their vaginas.
All 6 women from the TDF-FTC-MVC pod-IVR study group reported that they would use the device if it were a 7-d HIV prevention ring; 5 of 6 said they would use it if it were a 28-d device. All 6 were willing to recommend the IVR to others. With respect to use behaviors, all reported being highly confident that they could insert the IVR correctly. The two participants who previously used the TDF and TDF-FTC pod-IVRs noted that inserting the TDF-FTC-MVC pod-IVR was much easier than the first two pod-IVRs they had used, because they were familiar with the insertion process and confident they could insert the device correctly. Of the 6 women who used the TDF-FTC-MVC pod-IVR, none reported concerns with being able to remove the device; ultimately, all thought removing it was easy and were able to do so successfully. When asked about preferred access should the IVR become available, 3 of the women would use the pod-IVR if accessed by prescription, and 4 would use the pod-IVR if accessed over the counter.
Qualitative interview data revealed that, overall, the TDF-FTC-MVC pod-IVR was well tolerated by participants. Prior to insertion of the pod-IVR, clinicians provided verbal instructions. Participants reported that their confidence to insert the pod-IVR was increased by the clinician’s instructions, as well as when the clinician checked for proper pod-IVR placement on first insertion. The majority (5 of 6) of the women noted that they did not feel the pod-IVR in their vaginas at all during the use period. None of the participants felt that the pod-IVRs had an impact on their typical hygiene practices. All 6 expressed the desire for the pod-IVR to have multiple purposes, such as for HIV prevention and contraception. Detailed data of user experiences are reported elsewhere [39].
In vivo release rate measurements are based on the residual drug mass remaining in the used IVRs and the assumption, supported by in vitro data [26,28,29], that drug release is linear over the period of IVR use. The mean daily in vivo IVR drug release rates in used IVRs were as follows: median (IQR); TDF pod-IVR; TDF, 0.81 (0.66–1.07) mg/d; TDF-FTC pod-IVR; TDF, 0.98 (0.88–1.93) mg/d; FTC, 1.99 (1.70–2.31) mg/d (S1 and S2 Tables). Importantly, >95% of the residual TDF in the used IVR pods was present as the prodrug by HPLC, i.e., no significant prodrug hydrolysis was observed following 1 wk of use in vivo. The in vivo release rates of TDF from the TDF and TDF-FTC pod-IVR were found to be not significantly different (P = 0.4286).
For in vivo release rates under 1 mg/d TDF, less than 4% of the IVR drug content would have been released over 1 wk of IVR use, making differential residual drug measurements challenging. The amount of TDF released in the TDF-FTC-MVC pod-IVR treatment period could not be quantified accurately, although it was higher than in the TDF and TDF-FTC pod-IVR treatment periods, based on median TDF CVF concentrations (S3 Table). The mean daily in vivo IVR drug release rates were as follows: median (IQR); FTC, 2.37 (1.94–2.57) mg/d; MVC, 2.07 (1.77–2.09) mg/d.
Drug and drug metabolite concentrations in key anatomic compartments with the IVRs in place are summarized in S1–S3 Tables. These data show that the IVRs maintained high ARV drug exposure in CVFs and vaginal tissues (VTs) relative to lower concomitant plasma concentrations. FTC and MVC drug concentrations in all matrices were higher than corresponding TDF/TFV concentrations as expected based on the residual drug levels in used IVRs.
CVF ARV drug concentrations all exhibited low variability during IVR use in all treatment periods (Figs 1–3), with the exception of one low data point at Visit 3 for MVC (Fig 3D). TDF CVF concentrations in the three groups (i.e., TDF, TDF-FTC, and TDF-FTC-MVC) were not statistically significantly different (P = 0.4233); however, TFV concentrations were different (P = 0.0118) with this test. Total TFV CVF concentrations, defined as the molar sum of paired TFV and TDF concentrations reported as TFV, across all three groups also were different (P = 0.0155), as detailed below, with higher median concentrations in the TDF-FTC-MVC pod-IVR group. FTC CVF concentrations in the TDF-FTC and TDF-FTC-MVC groups were not significantly different (P = 0.9323).
A comparison of the medians and IQRs shows that the corresponding TDF (S1 Table) and TDF-FTC (S2 Table) pod-IVR datasets are similar, but the TDF-FTC-MVC (S3 Table) pod-IVR leads a total (i.e., TDF + TFV, on a molar basis) TFV exposure that is ca. two times higher: median (IQR); TDF; TDF pod-IVR, 58.1 (43.9–97.4) ng/mg; TDF-FTC pod-IVR, 43.1 (31.0–65.0) ng/mg; TDF-FTC-MVC pod-IVR, 96.9 (14.5–137.1) ng/mg; TFV; TDF pod-IVR, 13.9 (6.2–19.3) ng/mg; TDF-FTC pod-IVR, 15.9 (7.1–20.0) ng/mg; TDF-FTC-MVC pod-IVR, 28.0 (24.3–31.9) ng/mg; total TFV; TDF pod-IVR, 36.2 (31.3–60.6) ng/mg; TDF-FTC pod-IVR, 34.4 (26.9–48.7) ng/mg; TDF-FTC-MVC pod-IVR, 70.0 (59.2–80.6) ng/mg.
The 2-wk window for the follow-up visit after IVR removal (Visit 4/8/4) allowed ARV drug washout kinetics to be measured and the corresponding terminal half-lives of elimination to be calculated (Figs 1–3): median (IQR); TDF pod-IVR; TDF, 11.8 (10.6–14.4) h; TFV, 39.5 (30.0–58.1) h; TDF-FTC pod-IVR; TDF, 14.2 (11.7–15.4) h; TFV, 31.4 (25.9–36.2) h; FTC, 19.1 (13.5–20.1) h; TDF-FTC-MVC pod-IVR; TDF, 18.3 (11.8–24.2) h; TFV, 24.8 (24.4–31.9) h; FTC, 17.0 (15.5–17.8) h; MVC, 16.0 (15.8–18.3) h. Comparison of the TDF and TFV half-lives in the TDF pod-IVR group showed that the datasets were different (P = 0.0043). The TDF and TFV half-lives in the TDF-FTC pod-IVR group also were different (P = 0.0022), but the TDF and FTC half-lives were not different (P = 0.2403). The ARV drug half-lives across IVR groups were not significantly different (TDF, P = 0.4848; TFV, P = 0.4286). Comparison of the TDF, FTC, and MVC groups (Fig 3) showed that they were not significantly different (P = 0.9320). However, comparison of just the TDF and TFV groups showed that they were significantly different (P = 0.0909).
The measurement of Li+, an exogenous tracer added to the naïve CVL fluid, was used to correct the CVL drug concentration analyses for dilution in the TDF-FTC-MVC pod-IVR group and, hence, affords the corresponding drug concentrations in undiluted CVF [37]. The corrected CVL drug concentrations exhibited a moderate (total TFV and MVC) to weak (FTC) correlation with paired CVF (Dacron swab) drug concentrations (S4 Fig): TFV (total TFV, reported as the molar sum of TDF and TFV concentrations); slope, 7.53 ± 0.84; R2, 0.889; FTC; slope, 2.40 ± 1.43; R2, 0.220; MVC; slope, 4.37 ± 1.11; R2, 0.609. The CVF volume collected during the lavage procedure (S4 Fig) with the TDF-FTC-MVC pod-IVRs in place was: median (IQR), 85.6 (28.0–116) μL.
Molar antiviral drug concentrations in VT biopsy homogenate are described in Fig 4A and Fig 5A to allow direct comparison with the pharmacologically active metabolite of TFV against HIV, TFV-DP. The CVF (ng/mg) to VT (ng/mg) drug concentration ratio (Fig 4B and Fig 5B) provides a simple measure of xenobiotic partitioning between the two anatomic compartments: the lower the ratio, the more the antiviral agent distributes into the vaginal mucosa and the higher the vaginal bioavailability. Drug CVF:VT median (IQR) ratios are as follows: TDF and TDF-FTC pod-IVRs (data combined); TFV (molar sum of TDF and TFV, as TFV), 6.6 (4.1–28.6); FTC, 17.2 (8.5–66.5). The concentration ratio for TDF is 2.6 times lower than for FTC; TDF-FTC-MVC pod-IVR; TFV, 14.2 (6.0–16.5); FTC, 11.5 (5.8–17.0); MVC 3.5 (0.9–6.6). There was no statistically significant difference between the three groups (P = 0.1591).
To our knowledge, the clinical trial described here is the first to evaluate long-acting vaginal delivery of TDF-FTC, the only FDA-approved drug regimen (in the oral formulation Truvada) for HIV PrEP [12]. The innovative open-label, crossover trial format allowed two IVR formulations—TDF alone and in combination with FTC—to be evaluated sequentially with each participant acting as her own control. To our knowledge, the trial also involved the first triple ARV combination IVR to be evaluated clinically. The pod-IVRs were safe and generally well tolerated, despite high VT ARV concentrations especially for FTC and MVC (Fig 5A). The implications of our findings are discussed below in the context of developing a viable HIV PrEP candidate targeted at resource-poor regions.
The 3 pod-IVRs maintained high, controlled ARV drug concentrations in CVF over the period of product use, while leading to low systemic exposures (S1–S3 Tables). The low plasma ARV drug concentrations are a benefit of topical dosing as the risks of systemic toxicity and the emergence of drug resistance are reduced. The CVF drug levels decreased sharply post-IVR removal on Day 7 (Figs 1–3). The drug washout profiles allowed the CVF half-lives of the three drugs as well as TFV, the hydrolysis product of TDF, to be measured for the first time (Fig 3E). In all three groups, TFV had a significantly longer half-life (median, 24.8–39.5 h) than the other agents. We previously have described a mechanism in sheep [40] in which TDF, delivered via IVR, distributed efficiently from the CVF into the VT, where it hydrolyzed enzymatically to TFV. The accumulation of TFV in the VT formed a depot that could release TFV back into the lumen. A similar mechanism may operate here in women, accounting for the longer TFV half-life in CVF relative to the other agents.
The systematically higher (2.4–7.5-fold) CVF drug concentrations collected via lavage versus Dacron swab in the TDF-FTC-MVC pod-IVR group was unexpected and remains largely unexplained. It is possible that the lavage is more efficient at extracting drug that is associated with the surface of the vaginal epithelium. While we have used the Li+ tracer technique previously to measure CVF dilution in CVL in sheep [29], this is the first example of its application in a clinical setting for PK analysis. During IVR use, the median CVF volumes (0.086 mL; IQR, 0.028–0.12 mL) in the 6 participants were lower than the volumes collected previously in women at different phases of the menstrual cycle, 0.30 ± 0.22 mL (follicular phase) and 0.45 ± 0.21 mL (luteal phase), using a different assay [41]. This was expected based on the collection of multiple samples (one swab for VMB analysis, one pH swab, two Dacron swabs for CVF drug analyses, and one vaginal sidewall scraping for microbial biofilm imaging) prior to the lavage. As a result, a significant fraction of the available CVF was removed before CVL collection.
In HIV PrEP, there is no biomarker of ARV drug effect in susceptible, uninfected individuals to guide product development, unlike treatment of HIV-1/AIDS. The choice of ARV agent and target drug levels in key pharmacological compartments largely is based on theoretical arguments and results from preclinical studies. The strategy employed here was based on using FDA-approved ARV drugs that have shown clinical efficacy in HIV PrEP using oral or topical regimens [4,5,7–11,42]. Combined with PK data from oral dosing randomized clinical trials demonstrating efficacy [7,42], one can bracket the concentration targets associated with vaginal protection. A randomized, PK, cross-over study (MTN-001) compared TFV vaginal gel and oral TDF tablets by measuring drug and drug metabolite levels in VT from 144 HIV-uninfected women [32]. The TFV concentrations in VT homogenate at end-of-period visit were as follows: median (IQR); vaginal TFV gel, 113 (27–265) ng/mg; oral TDF, 0.15 (0.15–0.27) ng/mg. Assuming that the range of drug concentrations obtained from these dosing modalities is a key determinant for efficacy in preventing vaginal acquisition of HIV, the TFV concentrations in vaginal biopsies collected on Day 7 at IVR removal in this study, 8.4 (4.7–11.2) ng/mg (S1 Table), are suggestive of positive pharmacodynamic outcomes. However, this analysis assumes oral dosing is not fundamentally different than topical dosing in terms of tissue concentrations required for high PrEP efficacy.
A previous clinical trial evaluating a different TDF IVR platform—a reservoir IVR, in which a solid TDF formulation is contained in a hollow hydrophilic polyether urethane tube and the drug is delivered through the ring elastomer—reported median CVF TDF and TFV concentrations during ring use of 110 ng/mg and 70 ng/mg, respectively [43]. These levels are higher than the corresponding drug concentrations measured here (S1–S3 Tables), although the TDF:TFV ratio was considerably higher in our study (2.7–4.2-fold, depending on the formulation, compared to 1.5) and the TDF CVF concentrations were more stable (Figs 1–2). Together, these results suggest less TDF hydrolysis either in the IVRs during use or sample handling in our trial. Sample bioanalysis in both trials was carried out by the same laboratory using the same methods. The observation may be significant since the VT bioavailability in sheep of the TDF prodrug is nearly 100 times higher than for parent TFV [40].
Immune cells in the vaginal mucosa are believed to be the key pharmacologic compartment determining the efficacy of vaginal HIV PrEP. It is not feasible in early-stage clinical trials to collect sufficient VT to extract immune cells for intracellular drug analysis. VT biopsy homogenate has been shown to correlate well with CD4+ cells extracted from VT [44], and the relevant analytes are TFV, TFV-DP, the pharmacologically active metabolite of TFV against HIV, and FTC (Figs 4A and 5A). Due to the small amount of VT collected, the active, triphosphorylated metabolite of FTC could not be measured here. Median VT homogenate TFV-DP concentrations of 303 (TDF pod-IVR), 289 (TDF-FTC pod-IVR), and 302 (TDF-FTC-MVC pod-IVR) fmol/mg were measured post-IVR removal. These concentrations are 2–3-fold higher than the 120 fmol/mg median ectocervical biopsy levels obtained with a TDF reservoir IVR in a recent clinical trial [43]. Importantly, the median TFV and FTC VT concentrations (TDF pod-IVR; TFV, 8.4 ng/mg; TDF-FTC pod-IVR; TFV, 5.1 ng/mg; FTC, 75 ng/mg) were lower than those obtained in pigtailed macaques (TFV 28–35 and FTC 460–650 ng/mg, depending on the sampling location) with TDF-FTC pod-IVRs [26], which provided complete protection from Simian Human Immunodeficiency Virus (SHIV) infection in pigtailed macaques using the rigorous, repeat low-dose challenge model [45]. Smith and colleagues used the same model to evaluate a reservoir TDF IVR and also obtained complete protection from SHIV infection [46]. The median VT TFV concentration around 10 ng/mg observed with these reservoir TDF IVRs is comparable to the VT TFV values obtained in our study. When complemented by ca. 10 times higher tissue FTC levels, as observed here, these results suggest that the devices may be effective in HIV PrEP.
No clinical efficacy data for HIV PrEP using intravaginal MVC currently are available, and the levels in the pharmacologically relevant compartments required to afford protection are, therefore, unknown. The published in vitro antiviral potencies of MVC against HIV-1 primary and laboratory-adapted isolates in peripheral blood mononuclear cells (PBMCs) span a wide range of inhibitory concentrations: IC50, 0.1–4.5 nM; IC90, 0.5–13.4 nM [47]. The observed median MVC concentrations of 142 ng/mg (276 μM) in VTs on Day 7 were more than 20,000 times higher than the highest IC90, suggesting favorable pharmacodynamic outcomes. A randomized clinical trial (MTN-013/IPM 026) in 48 HIV-negative women evaluating matrix-IVRs delivering MVC alone or in combination with DPV measured MVC CVF concentrations of 2.5 and 1.1 ng/mg, respectively, at Day 28 when the IVRs were removed [48]. These concentrations are 170–390 times lower than those obtained in the current study (S3 Table). It should be noted that matrix-IVRs tend to have a drug release burst in the first week. VT MVC levels in MTN013 were below the lower limit of quantification for all subjects using the DPV-MVC IVR and were only quantifiable in 4 of the 12 MVC IVR users, with a 0.13–4.4 ng/mg concentration range, much lower than the median concentration (142 ng/mg) measured following pod-IVR delivery reported here (S3 Table).
To our knowledge, the ASPIRE study and The Ring Study were the first published Phase 3 studies of an ARV (i.e., DPV) delivered via IVR for HIV PrEP [49,50]. In the ASPIRE study, the incidence of HIV infection in the DPV group was 27% lower than in the placebo group; this improved to 56% protection in women over the age of 21 years; the difference was attributed to better adherence in women over the age of 21 years [49]. Both studies found low rates of adherence, particularly in young women raising concerns about the viability of IVRs for HIV PrEP in resource-limited regions like sub-Saharan Africa, where contraceptive IVRs are not as commonly used as in the developed world. It is believed by some that with increased familiarity with IVRs, adherence to IVR use will improve [51–55] and make IVR-based ARV regimens viable. This belief is based on previous experience with oral HIV PrEP regimens that demonstrated a dramatic increase in adherence when moving from the initial, blinded, placebo-controlled trials to the open-label follow-on trials [11,56–58].
In terms of the user perception and acceptability data, the pod-IVR delivery device was both easy to use and well tolerated. Women were willing to use it within the context of this cross-over study and anticipated being confident in their ability to use it for longer periods of time in “real-world” settings for HIV prevention. While, on average, participant confidence in their ability to access and use the IVR stayed the same or increased between uses, one participant’s data suggest that her confidence in her ability to insert and remove the IVR decreased with her second experience, indicating that further study is necessary to determine the factors related to confidence and what is needed to become skilled in IVR insertion and removal. Women were willing to use the TDF-FTC-MVC pod-IVRs effectively within the context of the study and anticipated feeling confident in their ability to use it for HIV prevention for sustained periods of use. Additionally, given the design of the current study, future work will need to assess user perceptions and experiences of pod-IVRs for longer periods of use, as well as in relation to sexual activity and use during menstruation.
Limitations of this study are as follows. Women at low risk for acquisition of HIV were recruited for this study since it was a first in humans and was primarily focused on PK and safety; these women may have different perceptions about HIV prevention than women at high risk for HIV, who will be the targeted population for these prevention products. The small sample size was also a limitation; however, the primary goal of the study was to show drug release and initial safety, as well as to get initial understanding of women’s perceptions and acceptability of the IVRs. The short duration was also a limitation since the IVRs will be used for 28 d; however, this initial study of 7 d duration was required by the FDA prior to a 28-d study. A 28-d study is planned as a follow-on to this study. Lastly, as described above, due to the small size of the vaginal biopsies, we were unable to measure FTC triphosphate in VT.
In conclusion, a crossover Phase I clinical trial sequentially evaluated TDF, TDF-FTC, and TDF-FTC-MVC pod-IVRs and demonstrated that the devices exhibited favorable safety and PK profiles across all three treatment periods. The results justify longer and larger follow-on clinical trials in the future.
|
10.1371/journal.pmed.1002341 | Community and health system intervention to reduce disrespect and abuse during childbirth in Tanga Region, Tanzania: A comparative before-and-after study | Abusive treatment of women during childbirth has been documented in low-resource countries and is a deterrent to facility utilization for delivery. Evidence for interventions to address women’s poor experience is scant. We assessed a participatory community and health system intervention to reduce the prevalence of disrespect and abuse during childbirth in Tanzania.
We used a comparative before-and-after evaluation design to test the combined intervention to reduce disrespect and abuse. Two hospitals in Tanga Region, Tanzania were included in the study, 1 randomly assigned to receive the intervention. Women who delivered at the study facilities were eligible to participate and were recruited upon discharge. Surveys were conducted at baseline (December 2011 through May 2012) and after the intervention (March through September 2015). The intervention consisted of a client service charter and a facility-based, quality-improvement process aimed to redefine norms and practices for respectful maternity care. The primary outcome was any self-reported experiences of disrespect and abuse during childbirth. We used multivariable logistic regression to estimate a difference-in-difference model. At baseline, 2,085 women at the 2 study hospitals who had been discharged from the maternity ward after delivery were invited to participate in the survey. Of these, 1,388 (66.57%) agreed to participate. At endline, 1,680 women participated in the survey (72.29% of those approached). The intervention was associated with a 66% reduced odds of a woman experiencing disrespect and abuse during childbirth (odds ratio [OR]: 0.34, 95% CI: 0.21–0.58, p < 0.0001). The biggest reductions were for physical abuse (OR: 0.22, 95% CI: 0.05–0.97, p = 0.045) and neglect (OR: 0.36, 95% CI: 0.19–0.71, p = 0.003). The study involved only 2 hospitals in Tanzania and is thus a proof-of-concept study. Future, larger-scale research should be undertaken to evaluate the applicability of this approach to other settings.
After implementation of the combined intervention, the likelihood of women’s reports of disrespectful treatment during childbirth was substantially reduced. These results were observed nearly 1 year after the end of the project’s facilitation of implementation, indicating the potential for sustainability. The results indicate that a participatory community and health system intervention designed to tackle disrespect and abuse by changing the norms and standards of care is a potential strategy to improve the treatment of women during childbirth at health facilities. The trial is registered on the ISRCTN Registry, ISRCTN 48258486.
ISRCTN Registry, ISRCTN 48258486
| There is growing evidence of the mistreatment of women during childbirth at health facilities, particularly in low-resource settings.
Few studies have designed and evaluated interventions to directly address disrespect and abuse.
We performed a comparative before-and-after evaluation of an intervention to address disrespect and abuse during childbirth.
This study used a participatory community and health system intervention in 2 hospitals in Tanga Region, Tanzania.
After nearly 1 year, the intervention was associated with a 66% reduced odds of a woman experiencing any disrespect or abuse during childbirth, with the biggest reductions observed for physical abuse and neglect.
This study suggests that supporting frontline community- and facility-based actors to identify and act on the symptoms and causes of disrespect and abuse in their own settings may be associated with a reduction of disrespectful treatment.
The results should be interpreted as a proof-of-concept study, as the intervention was tested in only 2 facilities.
| Maternal health in the Millennium Development Goal (MDG) era (2000 through 2015) was dominated by a focus on increasing skilled birth attendance, typically through facility delivery, as a means to reducing maternal mortality [1]. Countries with high maternal mortality ratios (MMR) worked to remove barriers to delivery in health facilities by eliminating user fees, providing conditional cash transfers, improving transport, and scaling up emergency obstetric care [2,3]. In sub-Saharan Africa, the MMR dropped by 45% between 1990 and 2015, which was short of the 75% MDG target, and the region still accounts for 66% of all maternal deaths as of 2015 [4].
As the MDG era came to a close, new evidence called into question the prevailing strategy that focused so narrowly on increasing intervention coverage. The World Health Organization’s multi-country survey examined records from more than 300,000 deliveries in hospitals in 29 countries and found that coverage of key clinical interventions did not imply reduced mortality [5]. In India, a massive conditional cash transfer program dramatically increased facility delivery, but it appeared to have little effect on the MMR [6]. These and other studies raised alarm in global circles about the poor quality of clinical care in facilities, its deterrent effect on the utilization of facilities for childbirth, and its impact on maternal and newborn health [5,7–10].
Meanwhile, a parallel development in the human rights field was drawing attention to other aspects of quality in delivery care. Investigative reports by human rights organizations documented abusive and discriminatory treatment in labor and delivery rooms in Kenya [11] and the United States [12], in clear violation of human rights standards. This gave new urgency to an old phenomenon of routine childbirth that is medicalized and then managed in facilities in ways that undercut women’s efforts to maintain control over their birth experience, preserve their dignity, and safeguard their physical and emotional wellbeing [13–15]. Subsequent literature reviews described a range of disrespect and abuse during childbirth in health facilities, including nondignified care (e.g., shouting/scolding, threatening comments), neglect, lack of physical privacy, physical abuse, inappropriate demands for payment, and nonconsented care, and confirmed that this treatment is a worldwide phenomenon [16,17].
In the wake of these studies, civil society organizations have created the core of a newly energized global movement for respectful maternity care (RMC) [18]. The movement’s hundreds of members include organizations from high-, middle-, and low-income countries, representing patients, professional associations, academicians, activists, and other stakeholders. These parallel developments—in public health, human rights, and civil society advocacy—have created the foundation for action.
The Staha study (meaning “respect” in Swahili) was designed to build a conceptual and evidentiary basis to address disrespect and abuse in the United Republic of Tanzania and to inform the global RMC movement. In 2 districts of the Tanga Region, we conducted a baseline study to measure prevalence of disrespect and abuse. It found that approximately 20% of women reported at least 1 incident of disrespect and abuse during their delivery in these facilities [19]. Subsequent discussions with community members, health workers, and managers led to the design of a multicomponent intervention to reduce disrespect.
This paper reports on an intervention to reduce the prevalence of disrespect and abuse during childbirth in 2 districts of the Tanga Region of Tanzania.
The study protocol was approved by the IRBs of Columbia University, Ifakara Health Institute, and the National Institute for Medical Research in Tanzania.
The intervention was developed through an iterative participatory process with local community and health system stakeholders that enabled them to analyze their own experience of disrespect and abuse in light of the baseline data and to consider potential actions to reduce it. Through this process, the Staha study identified a set of community and health system interventions that were intended to promote mutuality of respect between patients and providers. (See Fig 1 for the Staha intervention theory of change framework).
The intervention consisted of 2 main components (Fig 2). First, community and facility stakeholders together adapted a client service charter that had been promulgated by the Tanzanian government through the Ministry of Health and Social Welfare in 2005 but had lain dormant since that time. The local adaptation of the charter was drafted by a committee composed of the District Legal Officer, the Chairperson of the Council Health Service Board (CHSB), the Chairperson of the Social Welfare Committee of the District Council, the District Medical Officer, the District Hospital Medical Officer in-charge, the Chairperson of the District Hospital Governing Committee (HGC), the District Health Secretary, and a health center in-charge, as well as a village executive officer. This committee was selected based on recommendations from the district stakeholders and those involved in the participatory process. The adapted charter was then reviewed by 70 local stakeholders for feedback, including village executive officers, ward executive officers, district council authorities, health facility in-charges at intervention facilities, representatives from nongovernmental organizations (NGOs) in the district, and political leaders. Eighty-six percent of the stakeholders provided feedback. The main focus of the charter was to build consensus on norms and standards to foster mutual respect and respectful care. The charter was then disseminated to communities and posted in health facilities within the intervention district.
Second, the norms and standards articulated in the client charter were activated through a maternity ward quality-improvement process at 1 intervention facility, using tools from the Institute for Healthcare Improvement [20], to address disrespectful and abusive treatment as a system-level problem. Facilitated by members of our study team, maternity ward and hospital staff was presented with findings from the baseline and identified drivers of disrespect and abuse and proposed and prioritized interventions for change based on importance and feasibility. A quality-improvement team in the intervention hospital, consisting of staff from the maternity ward, reproductive and child health unit, pharmacy, and facility management, then facilitated the implementation of the interventions in the maternity ward and were responsible for tracking progress weekly. Maternity ward interventions were implemented 1 at a time and included moving the admissions area to a private room, using curtains for delivery and for physical examinations, posting supply stock outs to ensure transparency and build trust, and continuous customer satisfaction exit surveys anonymously filled by women who delivered in the ward. The latter were also used to monitor progress with the quality-improvement interventions and to decide when to implement the next intervention. The satisfaction surveys were analyzed by the quality-improvement team and discussed with the maternity ward staff. Staff in the maternity ward encouraged each other to treat women more respectfully. Quality-improvement interventions on the facility management level included tea provided to maternity ward staff, counseling of staff who continued to exhibit disrespectful behaviors, and best practice sharing with other wards and the regional hospital.
Although supported and facilitated by our study team, the implementation of the intervention was carried out by district, facility, and community stakeholders. The charter process took place over 6 months, after which our study team assisted in the facilitation of the quality-improvement process for 11 months. The intervention was then managed independently by facility managers and the regional quality improvement focal person for the following 10 months, at which point the endline survey commenced. Despite funding delays and turnover of key district and facility staff, the major components of the intervention were initiated within the intervention time frame and sustained beyond the endline survey.
We used a comparative before-and-after evaluation design to test the intervention to reduce disrespectful and abusive treatment of women during labor and delivery. As an implementation science study, this evaluation also collected qualitative and process data to identify potential mechanisms of change. Two districts in the Tanga Region of Tanzania, a rural region in the northeast corner of the country, were purposively chosen for the study. Tanzania has a maternal mortality ratio (MMR) of 398 per 100,000 live births [4], with 50.2% of deliveries occurring at health facilities [21]. The intervention was randomly assigned to Korogwe District with Muheza District as the comparison group. The intervention was measured at the facility level, with Korogwe District Hospital representing the intervention and Muheza District Hospital the comparison. The facilities are located approximately 60 kilometers apart.
At each facility, 2 surveys were performed, at baseline (December 2011 through May 2012) and 10 months after support for the intervention’s implementation ended (March through September 2015). Women aged 15 and over who delivered at the study facilities were eligible. Trained interviewers, unaffiliated with the facility, invited women discharged from the maternity ward to participate in an exit interview. All participants provided written informed consent. Interviews with eligible participants were conducted in tents outside of the maternity ward to ensure privacy. Women were given a bar of soap and a bottle of water in appreciation of their time and participation. Patients requiring support following disclosure of abuse were provided with a referral for mental health services at the regional hospital. The planned sample size for the study was 2,936 women: 734 women per district per time period (baseline and endline). Assuming the sample size, a 2-sided alpha of 0.05, and a 30% baseline prevalence of disrespect and abuse, there would be 80% power to detect a 15% decline in reported disrespect and abuse in the intervention facility compared to the comparison facility.
The primary outcome of interest was self-reported experience of disrespectful or abusive treatment during labor and delivery. A woman was labeled as having experienced disrespect or abusive treatment if she answered “experienced” to any of the 14 questions about whether the specific disrespectful or abusive actions listed in Table 1 occurred during her labor and delivery. Secondary outcomes included affirmative responses for each of the questions in the categories of disrespect and abuse (Table 1). Individual questions and categories were based on a landscape analysis by Bowser and Hill and were further adapted and validated for the Tanzanian context with focus group discussions and in-depth interviews with recently delivered women and health system stakeholders [17]. We also explored the association between the intervention and delivery satisfaction and quality of care. Women were asked to rate their satisfaction with delivery, the respect providers showed them during delivery, and the quality of care during delivery. For satisfaction, responses were dichotomized from a 4-point Likert scale into very satisfied, somewhat satisfied, somewhat dissatisfied, or very dissatisfied. Quality measures were dichotomized into excellent, very good, fair, or poor.
The literature suggests that a range of respondent and delivery experience factors are associated with report of disrespectful and abusive treatment during labor and delivery. Factors that were identified in past research included the following: respondent characteristics such as age, education, marital status, socioeconomic status, parity, reported low mood or depression in the last 12 months, and reported past physical abuse or rape [16,19,22,23]. Delivery experience factors included length of stay for delivery, Caesarean section, if the woman came directly to the facility for delivery, and any reported complications during childbirth. These same factors were included in a recent paper from the Staha study on prevalence and correlates of disrespect and abuse [19]. To measure socioeconomic status, we used a principal component analysis (PCA), developed by Filmer and Prichett, of 18 survey questions about household assets [19,24]. PCA index results were categorized into quintiles, with the lowest 2 quintiles (40%) classified as poor.
We assessed several process measures in the endline survey that related to the fidelity of the intervention, including questions on women’s experience regarding a range of respectful practices that providers were encouraged to adopt. Respectful maternity care questions were adapted from the Maternal and Child Health Integrated Program (MCHIP) Maternal and Newborn Quality of Care Survey [25].
We first compared means and frequencies of baseline participant characteristics, including factors that may influence reporting of disrespect and abuse, by district using chi-square tests and t tests. Monthly baseline trends in reporting of disrespect and abuse were compared between intervention and comparison districts to confirm parallel trends, a key assumption of difference-in-difference analysis. This was done by regressing the main outcome on an interaction term between month of baseline data collection and district. Second, as per our prespecified analytic plan, we performed unadjusted logistic regression using a difference-in-difference approach for all primary and secondary outcomes using a dummy variable for facility, time (pre-post), and the interaction term of the 2 as a measure of the intervention impact. Finally, to test whether differences in women’s or facility characteristics influenced our estimates, we used a multivariable logistic regression to estimate a difference-in-difference model that included all variables in our conceptual model in addition to the above dummy variables. We followed the analysis plan as set forth in the IRB protocol except that we elected to adjust the final analysis for demographic variables to account for observed differences between the intervention and comparison group that might otherwise bias the association between the intervention and the outcome. For quality of care and satisfaction outcomes, we estimated relative risks using generalized linear models with a Poisson distribution, a log link, and robust standard errors to account for the high prevalence of these outcomes. Complete case analysis was used to permit comparability across models and avoid bias due to missing data. For the fidelity and process indicator measures, endline data from the 2 districts were compared using chi-square tests.
To address potential biases due to selection and contamination, sensitivity analyses were conducted by excluding participants who reported that they were aware of the intervention and by restricting analysis to those who lived in the nearby vicinity of the study facilities and thus were not drawn to the intervention hospital from the control catchment. All statistical analyses were performed with STATA (version 13). The trial is registered on the ISRCTN Registry (www.controlled-trials.com), number ISRCTN48258486.
At baseline, 2,085 women at the 2 study hospitals who had been discharged from the maternity ward after delivery were invited to participate in the survey. Of these, 1,388 (66.57%) agreed to participate. At endline, 1,680 women participated in the survey (72.29% of those approached). Women did not participate in the study largely due to the required wait time postdischarge for the administration of the interview. At baseline, there were some statistical differences between women delivering in the intervention hospital versus the comparison hospital (Table 2). A higher proportion of participants in the intervention facility than the comparison facility were married and of higher socioeconomic status, and a smaller proportion reported low mood or depression in the last 12 months or ever being physically abused or raped. Higher proportions of participants in the comparison facility had shorter lengths of stay for delivery and were more likely to come directly to the facility for delivery compared to women in the intervention facility. Other baseline characteristics were not statistically different. Preintervention trends in the main outcome between the 2 facilities did not differ significantly, with the exception of the first month of data collection, which was likely due to a small sample of surveys collected in that month.
Table 3 presents crude difference-in-difference estimates for all primary and secondary outcomes. There was a 3.39% (p < 0.0001) decrease in the percent of all women who experienced disrespect and abuse between the intervention and comparison facilities. Table 4 presents results from the multivariable logistic regression difference-in-difference analysis for the main outcome of interest. Complete data from baseline and endline were available for 2,983 women (97.23%) for this analysis. The intervention was associated with a 66% reduced odds (95% CI: 0.21–0.58, p < 0.0001) of a woman experiencing disrespect and abuse when adjusted for all covariates in the model. Women in the intervention facility were also significantly less likely to report events of neglect (OR: 0.36, 95% CI: 0.19–0.71, p = 0.045) and physical abuse (OR: 0.22, 95% CI: 0.05–0.97, p = 0.003) when adjusted for all variables in the conceptual model (Table 5). Finally, the intervention was associated with an increased likelihood of rating the respect providers showed them during their stay at the facility for delivery as excellent or very good (RR: 3.44, 95% CI: 2.45–4.84, p < 0.0001) and rating the overall quality of care for delivery as excellent or very good (RR: 6.19, 95% CI: 4.29–8.94, p < 0.0001) (Table 5). The process indicators show that women delivering in the intervention facility rated most elements of quality of care and respectful maternity care significantly higher than those in the comparison facility (Table 6).
This study found that after a participatory community-health system intervention in Tanga Region, Tanzania, the likelihood of self-reported disrespectful and abusive care during labor and delivery was significantly reduced (66% reduced odds). The largest reduction was for physical abuse, followed by neglect. Process indicators showing better patient-reported quality of care in the intervention facility, including respectful treatment from providers, support the likelihood that the intervention was responsible for the reduction in disrespect and abuse. Importantly, these effects were still observed nearly one year after the end of Staha’s facilitation of implementation in the intervention district, indicating the potential for sustainability.
While the size of the absolute reduction in disrespectful treatment in the intervention facility was large (from 13.10% to 3.20%), there were reductions in the comparison facility as well (from 21.27% to 15.76%, Table 3). Although this represents a substantial reduction in the risk of experiencing disrespect and abuse for an individual woman, there was a relatively small difference between the intervention and comparison facilities in prevalence reduction (3.39%). This was likely due to quality changes that occurred in the comparison facility over the study period, including the posting of patients’ rights in the maternity ward and a pharmacy price list, and delivery room renovations to include cubicles for privacy.
There have been several other efforts to design, implement, and assess interventions specifically aimed at reducing disrespect and abuse. In Kenya, where the early human rights reports had garnered significant public attention, the Heshima project used multiple interventions, including maternity open days for prospective patients to visit the facility, values-clarification workshops for providers, and dispute-resolution training for community leaders. They reported a 7% reduction in the proportion of women reporting disrespect and abuse [26]. However, the study did not include a comparison group and coincided with broader health system reforms that may have affected the results. In Tanzania, an intervention conducted in a high-volume referral hospital in Dar es Salaam used maternity open days to improve patient knowledge and awareness, combined with a respectful care workshop to sensitize and empower providers. These interventions, developed through a participatory process [27], were shown to have a positive effect on multiple proximate indicators, including patient knowledge of rights and birth preparedness, provider attitudes, and patient-provider communications [28]. However, the study was not designed or powered to measure impact on prevalence of disrespect and abuse, nor was a comparison group included. To our knowledge, the Staha study reported here is the first comparative before-and-after evaluation of an intervention to address disrespect and abuse during childbirth that includes a comparison group to reduce the likelihood that secular trends or other factors that could explain the observed change.
In Staha, as in the few other projects that have tried to tackle disrespect and abuse explicitly, we adapted tools from several related fields, including quality improvement, social accountability, and behavior change. Each of these tools and techniques has a mixed record in the literature [29,30]. Systematic reviews generally show that their effect depends on whether, in context, their implementation activates deeper change mechanisms [31,32]. Thus, the mechanism of action is never the tool itself but the entire process by which the problem is identified and analyzed, the intervention chosen, its use negotiated and practiced, and its effects assessed and understood [33]. And this implementation process likely works in a sustained way only when it engages the organizational culture and power dynamics at the heart of disrespect and abuse [34–36].
This study had some limitations. First, we found several differences between the characteristics of women in the 2 groups at baseline. We controlled for these factors in our regression analyses, and our finding of parallel trends for reports of disrespect and abuse in the pretreatment data supported the notion that the comparison district was a plausible counterfactual for the intervention district. Second, the response rate for both baseline and endline introduces the possibility of selection bias. However, women in our sample were comparable in terms of age and parity to the larger delivery population in the facilities, providing confidence that participants are representative of the sample population. Our sample had a lower proportion of women who had Caesarean sections than the facility delivery population. In previous study analyses, having a Caesarean section was not associated with self-report of disrespect and abuse upon discharge after childbirth [19]. Third, it is possible that women may have self-selected into the intervention facility or that there was contamination from the comparison district, due to exposure to the client charter and/or word of mouth about the quality-improvement intervention. These women may have had a more positive expectation and subsequently more positive reporting of their care. However, elimination of women who reported that they had heard of the intervention did not alter our findings (S1 Table), nor did restricting our sample to participants who lived in the nearby vicinity of the hospitals and were therefore not likely to be bypassing other facilities. Fourth, it is possible that other unmeasured changes in the facilities during the study period, such as turnover in staff or facility management, may have influenced study results. We did not identify any such changes that were major enough to explain the large effect we saw. Fifth, even though we observed a smaller baseline prevalence and reduction in disrespect and abuse than assumed in the power calculation, because we still found a significant result, our study was appropriately powered. Sixth, we used logistic regression in instances in which we believe the odds ratio was a good approximation of the risk ratio due to the rare disease assumption. In the instances in which the outcome was common (i.e., >20%), we opted to use relative risk regression to obtain our estimates of interest. Lastly, this is a proof-of-concept study, as only 2 hospitals in Tanzania were included. Future, larger-scale research should be undertaken to evaluate the applicability of this approach to other settings.
Our findings have important implications moving forward as the RMC field evolves. It is tempting in global- and national-level discussions to conceptualize the next challenge as the translation of global standards of respectful care into practice by identifying the most-effective intervention through studies such as Staha. However, insights from implementation science, behavioral science, and organizational science would all warn against a simple translational approach [37]. Eliminating disrespect and abuse requires individual behavior change, organizational and systems change, and, ultimately, deeper societal transformation. These are complex, multidimensional challenges that do not evaporate just by order of a court or mandate of a minister [38]. There will never be a simple, single technical fix to identify and prescribe. Thus, unlike with a new drug or biomedical procedure, our goal in testing these interventions to mitigate disrespect and abuse should not be to assert definitively which tool works best, so that it can be mandated, funded, and promoted widely. Instead, initiatives such as Staha contribute to the field by demonstrating promising strategies for enabling and supporting frontline community- and facility-based actors to identify, confront, and act on both the symptoms and causes of disrespect and abuse in their own settings.
This study provides evidence that a participatory community-health system intervention that articulates new norms, standards, and practices for mutual respect between patients and providers and supports their implementation through facility-based management and health provider reflection is a potential strategy to reduce the prevalence of disrespect and abuse during childbirth. The magnitude of the effects observed here suggests that this is a promising direction for future efforts to reduce disrespect and abuse. Future initiatives to build on the Staha findings should carefully adapt the intervention to local context, retain the active participation of key stakeholders, and explore efficient means for scaling it both geographically and institutionally by identifying the particular changes needed at higher levels of the health system to sustain such practices at the frontline [39,40].
Finally, improvements in technical quality of care, and in human resource and commodities availability, should accompany efforts to humanize care to address persistently high rates of maternal and newborn deaths in health facilities in many low-resource settings. Developing quality-improvement strategies that can tackle both clinical competence and compassion, supported through community accountability mechanisms, should be a global priority.
|
10.1371/journal.pntd.0007076 | Topical treatment with gallium maltolate reduces Treponema pallidum subsp. pertenue burden in primary experimental lesions in a rabbit model of yaws | Gallium is a semi-metallic element known since the 1930s to have antimicrobial activity. This activity stems primarily from gallium's ability to mimic trivalent iron and disrupt specific Fe(III)-dependent pathways, particularly DNA synthesis (due to inhibition of ribonucleotide reductase). Because of its novel mechanism of action, gallium is currently being investigated as a new antibacterial agent, particularly in light of the increasing resistance of many pathogenic bacteria to existing antibiotics. Gallium maltolate (GaM) is being developed as an orally and topically administrable form of gallium. Yaws is a neglected tropical disease affecting mainly the skin and skeletal system of children in underprivileged settings. It is currently the object of a WHO-promoted eradication campaign using mass administration of the macrolide azithromycin, an antibiotic to which the yaws agent Treponema pallidum subsp. pertenue has slowly begun to develop genetic resistance.
Because yaws transmission is mainly due to direct skin contact with an infectious skin lesion, we evaluated the treponemicidal activity of GaM applied topically to skin lesions in a rabbit model of yaws. Treatment efficacy was evaluated by measuring lesion diameter, treponemal burden in lesion aspirates as determined by dark field microscopy and amplification of treponemal RNA, serology, and immunohistochemistry of biopsied tissue samples.
Our results show that topical GaM was effective in reducing treponemal burden in yaws experimental lesions, particularly when applied at the first sign of lesion appearance but, as expected, did not prevent pathogen dissemination.
Early administration of GaM to yaws lesions could reduce the infectivity of the lesions and thus yaws transmission, potentially contributing to current and future yaws control campaigns.
| Yaws is a neglected tropical disease affecting children in underprivileged countries, transmitted through direct skin contact with an active lesion. This infection, although rarely fatal, can lead to disfigurement and serious disability. The World Health Organization is currently conducting a yaws eradication effort that employs mass administration of azithromycin, an antibiotic against which the yaws pathogen has slowly begun to develop genetic resistance. Because this phenomenon has the potential to undermine the eradication effort, we investigated the antimicrobial activity of gallium maltolate, which has a novel mechanism of action, against the yaws pathogen. Our initial results show that topical application of gallium maltolate has significant treponemicidal activity, and suggest that this compound might find an application in the effort to eradicate yaws. Future studies will evaluate whether oral administration of gallium maltolate is as effective as the antibiotics currently approved for yaws treatment to clear systemic infection.
| Yaws is a neglected tropical disease (NTD) [1,2] caused by the spirochete bacterium Treponema pallidum subsp. pertenue (Tp pertenue), a pathogen closely related to the syphilis agent, Treponema pallidum subsp. pallidum (Tp pallidum) [3]. Yaws mainly affects children less than 15 years of age, among whom the disease is spread via skin contact with an infectious early lesion [4,5], although a marginal role in transmission may be played by vector flies and non-human primates [6,7]. Similar to syphilis, untreated yaws becomes a multistage chronic disease that mainly affects the skin and skeletal system of infected individuals [2,8–11]. In contrast to syphilis, yaws is believed not to affect either the cardiovascular system or the central nervous system (CNS), and not to be vertically transmitted, even though several studies suggest that CNS, cardiovascular, and fetal involvement cannot be ruled out [12–15]. A detailed review of the early and late clinical manifestations of this disease is available elsewhere [2,8–11,16–20].
Yaws is currently reported from 14 countries in the Western Pacific, South-East Asia, and African WHO regions where, collectively, about 65,000 new cases of yaws per year have occurred since 2008 [21]. The highest disease incidence is reported in Papua New Guinea, the Solomon Islands, and Ghana [22–26]. A major anti-yaws campaign in the 1950s and 1960s by the WHO and UNICEF eradicated about 95% of the disease in 46 developing countries, causing its prevalence to drop from 50 million cases (reported in 1952) to 2.5 million cases (reported in 1964) [27]. This success induced the WHO to gradually eliminate its eradication programs, confident that the primary healthcare facilities established during the campaign would identify and eliminate the remaining cases. Lack of commitment and resources, however, led to disease resurgence in several countries [17]. In 1995, the yaws global prevalence in children was estimated to be of approximately 500,000 cases [21]. In 2013, a new campaign to achieve global eradication of yaws by 2020 was initiated by the WHO [28]. This new effort was warranted by the evidence that a single oral dose of azithromycin proved as effective as injected benzathine penicillin in curing yaws [29]. Using azithromycin could avoid the intrinsic difficulties associated with the use of penicillin, which requires an efficient cold chain and personnel able to perform injections. The use of azithromycin, however, has induced the insurgence of macrolide-resistant yaws strains whose spread could undermine the success of the ongoing campaign [30]. Additionally, even when treated with systemic antibiotics like azithromycin or penicillin, lesions might remain contagious for several hours to days post-treatment, based on studies of drug administration in the rabbit model of syphilis [31]. In this context, the application of a topical anti-treponemal agent unable to induce genetic resistance could be useful to reduce transmissibility.
Gallium (Ga) is a semi-metal that has been extensively studied as an anticancer agent, and is currently being evaluated for repurposing as a novel antimicrobial agent due to its demonstrated activity against pathogenic bacteria and its very low human toxicity [32–38]. In the early 1930s, prior to the discovery of penicillin, experiments conducted at the Pasteur Institute in Paris, France, supported the efficacy of some gallium compounds, particularly “gallium tartrate” (GaT), against Tp pallidum and trypanosomes [39]. These studies claimed that administration of GaT eradicated treponemes from several infected rabbits within three to four days after a single intravenous or intramuscular injection and caused the then-used Meinike reaction (a serum-induced precipitation of cholesterolized organ extracts performed to diagnose an active Tp infection as an alternative to the modern non-treponemal tests) to become negative [39]. The antimicrobial activity of Ga is due primarily to it acting as a non-functional mimic of Fe(III) [34,40]. Unlike Fe, which readily cycles between trivalent and divalent states, Ga is not reducible under physiologic conditions, remaining as Ga(III). By competing with Fe(III), Ga(III) can inhibit many Fe(III)-dependent biochemical activities, the most prominent being the activity of ribonucleotide reductase to synthesize DNA [34]. Recent interest in Ga compounds as antimicrobial agents [41,42] has been motivated by the need for new approaches to fight antibiotic-resistant bacteria and by the shortage of new antibiotics in the pharmaceutical pipeline.
Gallium maltolate (GaM) is currently under investigation as an orally and topically administrable form of Ga [38,40]. GaM is pH and charge neutral, and is moderately soluble in both water and lipids, making it well suited for pharmaceutical administration [40]. Locally administered GaM was effective against Pseudomonas aeruginosa in a mouse burn/infection model [43], it was also effective against Staphylococcus aureus and methicillin-resistant S. aureus (MRSA) [44] and several veterinary pathogens [45–49]. Additionally, topical GaM provided in a water/hydrophilic petrolatum emulsion was shown to have anti-inflammatory and analgesic activity in people with neuropathic pain and inflammatory conditions [38,50–53]. Here, we investigated the efficacy of topical GaM against Tp pertenue in a rabbit model of yaws.
New Zealand White (NZW) rabbits were used for propagation of Tp pertenue and intradermal (ID) experimental infections to assess efficacy of topical GaM. Animal care was provided in accordance with the procedures described in the Guide for the Care and Use of Laboratory Animals [54] under protocols approved by the University of Washington Institutional Animal Care and Use Committee (IACUC). The protocol number assigned by the IACUC committee that approved this study is 4142–01. No investigations using human samples or humans were conducted in this study.
Outbred adult male NZW rabbits ranging from 3.5–4.5 Kg were purchased from Western Oregon Rabbit Co. (Philomath, OR). Rabbits were housed at 16°C to 18°C in individual cages and fed antibiotic-free food and water. Prior to entry into the study, to rule out previous infection with the rabbit syphilis agent Treponema paraluiscuniculi, each animal was bled and heat-inactivated sera were tested individually with both the fluorescent treponemal antibody absorption (FTA-ABS) and Venereal Disease Research Laboratory (VDRL; BD, Franklin Lakes, NJ) tests according to the manufacturer’s instructions. Only rabbits seronegative to both tests were used for either treponemal propagation or experimental ID inoculation. The Tp pertenue strain (Gauthier) used in these experiments was isolated in the early 1960`s in Brazzaville, Congo, from a patient’s skin lesion and provided to us by Dr. Sheila Lukehart (University of Washington), who previously received it from Dr. Peter Perine (CDC, Atlanta, GA). A 2012 frozen glycerol stock of the Gauthier strain containing 4x106 Tp cells/ml was inoculated into the testicles of a NZW rabbit as previously described [55] and treponemes were allowed to proliferate until the animal developed an orchitis and presence of treponemal cells within testicular tissue could be assessed by dark-field microscopy (DFM) from a needle aspirate. An aliquot of the glycerol stock (100 μl) was saved for DNA extraction using the DNA Mini Kit (Qiagen, Germantown, MD) to confirm strain identity by PCR using the tprL gene (tp1031) as amplification target as previously described [56]. Briefly, the amplicon generated by Tp pertenue DNA (209 bp) in the tprL PCR assay differs in size from that originated by Tp pallidum and endemicum subspecies DNA (588 bp) due to a deletion that encompasses part of the tprL 5’-flanking region and ORF in the pertenue subspecies. Bacteria for ID inoculations of test animals were extracted from the testes of the euthanized rabbit in sterile saline supplemented with 10% normal rabbit serum (NRS). Testicular extract was collected in a sterile 15 ml tube and spun twice at 1,000 rpm (180 x g) for 10 minutes in an Eppendorf 5810R centrifuge (Eppendorf, Hauppauge, NY) to remove rabbit cellular debris. Treponemes were enumerated using DFM and percentage of motile organisms was also recorded. Extract was diluted in 10% NRS-saline to obtain approximately 7 ml of treponemal suspension at the desired concentration (107 cells/ml). Test (n = 4) and control rabbits (n = 2) were injected ID with 100 μl of treponemal suspension (containing 106 treponemes) in 10 sites on their clipped backs. The skin was marked with permanent ink one inch below each injection site to facilitate location of the lesions. Following ID injection, treponemal motility was assessed again using DFM to ensure that the time elapsed between harvest and ID inoculation did not affect pathogen viability. After ID inoculation, rabbit backs were clipped daily to allow monitoring of lesion development and surgical procedures to collect lesion biopsies and aspirates. For the purpose of antimicrobial testing, ID infection is preferable to IT infection because skin lesions can be readily aspirated for DFM examination to assess the presence of Tp cells and evaluate treatment efficacy.
A 0.5% w/v GaM cream was provided by Gallixa (Menlo Park, CA) along with the carrier alone (an emulsion of water and hydrophilic petrolatum). Test animals were divided into three groups (Groups 1–3), each group containing two rabbits. Groups 1 and 2 received GaM twice a day every 12 hours. Group 1 rabbits began treatment at the first clinical evidence of infection by DFM analysis of lesion aspirates (day 4 post-inoculation), while Group 2 rabbits received GaM when lesions had become clearly indurated (day 14 post-inoculation). Administration at day 4 post-inoculation, was performed to evaluate GaM ability to prevent lesion progression, while application at day 14 post-infection aimed at evaluating GaM ability to reduce treponemal burden faster than in control animals and accelerate lesion healing. Group 3 rabbits received an equal amount of carrier only from day 4 post-inoculation. Treatment consisted in applying 250 μL of either GaM or carrier on top of each lesion, followed by gentle manual spreading to ensure uniform coverage of the whole lesion. Following application, animals were monitored for a few minutes to ensure that they would not remove the ointment, and were then taken back to their cages.
Development of skin lesions at injection sites was monitored in all experimental subjects by measuring diameter of indurated lesions each day, after shaving the animals and prior to the first application of GaM or carrier. Appearance of lesions at distant sites, due to pathogen hematogenous dissemination from the primary injections sites, was also monitored. Treponemal burden within primary lesions was first assessed by performing DFM analysis of lesion needle aspirates of all lesions at day 13 post-inoculation. At day 23 post-inoculation, a second set of aspirates was obtained from Group 2 (GaM-treated since day 14 post-inoculation) and Group 3 (carrier-treated control) rabbits from all lesions (except those that were previously biopsied). Approximately 100 fields per slide were examined and the treponeme number was recorded. Two lesion biopsies were obtained at day 13 post-inoculation using a 4 mm biopsy punch from each of Group 1 (GaM-treated since day 4 post-inoculation) and Group 3 (control) animals for evaluation of treponemal burden by real-time amplification of Tp mRNA (see below). Two additional biopsies were obtained at day 25 and day 33 post-inoculation from each animal in Group 2 (treated from day 14) and Group 3 (control). Inoculation sites to be biopsied were selected randomly. A flowchart describing the experimental design is provided in Fig 1. In all cases, biopsy samples were minced with a sterile scalpel immediately after collection and further homogenized in 400 μl of phenol-based TRIzol buffer (Life Technologies, Santa Clara, CA) using a disposable plastic pestle. Samples were stored at -80°C until use. Total RNA from biopsies was obtained according to the TRIzol extraction protocol. Extracted RNA was treated with DNaseI to obtain DNA-free RNA as previously described [57]. Reverse transcription into cDNA was performed with the Superscript III First-Strand Synthesis System (Life Technologies) using random primers according to the manufacturer's instructions. Message quantification was performed using an established relative quantification method that targets the mRNA for the treponemal 47 kDa lipoprotein (encoded by the tp0574 gene) and that normalizes the tp0574 signal to the message for the rabbit housekeeping gene Hypoxanthine-Guanine Phosphoribosyl Transferase (HPRT) [57]. Primer sequences and real-time amplification conditions for both targets, as well as details on plasmid standard preparation, were previously published [57,58]. Data from message quantification and treponemal counts from aspirates were analyzed with Student’s unpaired two-tailed t-test and significance set at p≤0.05. All animals were bled weekly for serology (FTA-ABS and VDRL), to assess seroconversion and confirm establishment of infection. Animals were euthanized 45 days post-inoculation after serological evidence demonstrated that all had become infected. Serological assays were performed on the same day for all sera collected at a given time point, to minimize test-to-test variability. The technologist performing the assays was blinded to the treatment status of the animals from which the samples were collected.
Immunohistochemistry (IHC). At day 13 post-inoculation a 4-mm lesion biopsy was taken randomly from each animal in Group 1 and Group 3. Two additional biopsies from disseminated skin lesions that appeared in one of the Group 1 rabbits were also taken. All biopsies were fixed in 10 ml of 10% neutral buffered formalin (NBF) at room temperature for approximately 72 hours and then transferred to 70% ethanol and stored at 4°C until paraffin embedding and sectioning. For embedding, biopsies were transferred back into 4% NBF (PanReac Applichem, Barcelona, Spain) for 3 hrs (2 x 1.5 hr passages). Subsequent sample processing was performed in a Leica ASP300 instrument (Leica Biosystems, Wetzlar, Germany). Samples were incubated in water for 10 min, and then transferred in 80% ethanol for 1 hr, in 96% ethanol for a total of 2 hrs (2 x 1 hr passages), and in absolute ethanol for 2 hrs (2 x 1 hr passages). Following dehydration, samples were transferred into paraffin solvent (Histo-Clear, National Diagnostics, Atlanta, GA) for 2 hrs (2 x 1 hr passages), followed by three passages of 1 hr each in liquid paraffin at 58°C (Paraplast X-tra, Millipore-Sigma, St. Louis, MO). From these samples, 3-μm sections were cut, placed on a heath block at 65°C for a total of 20 min to allow tissue adherence to the slide, and then stored at room temperature. For IHC procedures, silane-treated slides were used to further improve tissue adherence, and tissue sections were stored at 37°C. For hematoxylin and eosin (HE) staining, deparaffinized and rehydrated sections were placed in hematoxylin solution for 8 min and then rinsed for 3 min with tap water. Eosin staining was carried on for 1 min prior to washing under tap water for 5 minutes. Dehydration was obtained by passage in 96% ethanol for 4 min, followed by 2 passages in 100% ethanol for 2 and 3 min, respectively, and 2 passages in Histo-Clear for a total of 3 min. Sections were mounted using an acrylic resin (Eukitt, Orsatech, Gmbh), taking care to leave no bubbles during the process. Slides were left to air-dry overnight before being analyzed. For specific immunostaining, slides were first heated at 65°C for 1 hr, and then deparaffinized in EX-Prep solution (Roche Diagnostics, Indianapolis, IN) at 72°C. Cell conditioning was performed by applying ULTRA CC1 solution (Roche diagnostics) for a variable time (20–36 min, depending on the primary antibody) to correct epitope alteration due to fixation of the tissue sections. Polyclonal anti-CD4, -CD8, and -CD20 primary antibodies (Roche Diagnostics) were used at 1:100 dilution, and slides were incubated for 16 min (anti-CD4 and -CD8) or for 12 min (anti-CD20 antibodies). To avoid evaporation, tissue sections were covered with ULTRACS liquid coverslip (Roche Diagnostics) following application of the primary antibody. Primary antibody was removed by washes with a Tris-based buffer solution (Reaction buffer, pH 7.6; Roche Diagnostics). Reagents provided in the Ultra View Universal DAB (3,3’ diamino-benzidine) Detection kit (Roche Diagnostics) were used according to the manufacturer`s instruction for detection of primary antibody binding. Slides were then rinsed with water and counterstained with hematoxylin for 12 minutes. Tissue sections were dehydrated using two 5-min rinses with 96% ethanol followed by two 5-min rinses in absolute ethanol, and two 10-min washes with Histo-Clean. Prior to reading, coverslips were mounted on slides using the Eukitt acrylic resin and air-dried prior to being read.
Measurements of lesion diameter as a function of time (Fig 2) showed that in animals treated with GaM since day 4 post-inoculation, lesion development was significantly attenuated compared to controls (Group 3) or to animals treated since day 14 post-inoculation. Most lesions form Group 1 animals failed to develop into indurated papules and enlarge, but rather remained flat, although generally erythematous. Compared to controls, only one of the Group 2 rabbits that initiated treatment at day 14 post-inoculation showed a significant decrease in lesion diameter. Assessment of treponemal burden by DFM on lesion aspirates obtained at day 13 (Fig 3A) and day 23 (Fig 3B) post-inoculation showed that compared to controls and untreated animals, treponemal burden in lesions from GaM-treated animals was significantly reduced (p<0.05), while carrier-treated animals and untreated did not show significant differences in number of treponemes counted. Analysis of treponemal burden performed at day 23 post-inoculation showed that overall significantly fewer treponemes could be found in lesions from rabbits that started treatment at day 14 post-infection (Group 2) compared to controls. Table 1 summarizes the DFM results for each animal in each group, together with the results of FTA-ABS and VDRL tests following experimental infection. Notably, treated animals seroconverted approximately a week later than control animals, which is consistent with the reduced treponemal burden due to GaM.
Treponemal burden in primary lesions was further assessed using Tp mRNA quantification normalized to the rabbit housekeeping gene HPRT. Message quantification at day 13 post-inoculation showed that no treponemal mRNA was detected from Group 1 rabbit lesions compared to controls (Fig 4A). At day 23 post-infection, significantly less (p<0.05) Tp mRNA was detected in lesions from Group 2 rabbits compared to controls, while no difference was seen between these rabbits and the control ones at day 33 post-inoculation. By the end of the experiment (day 45 post-inoculation) all animals had developed erythematous disseminated skin lesions. Analysis of needle aspirates from a small subset of disseminated lesions revealed the presence of treponemes by DFM (not shown).
Both biopsies obtained from Group 1 rabbits showed the presence of very modest inflammatory infiltrates and absence of damage to follicles (Fig 4A and 4B), nearly like normal skin. Histological analysis of a disseminated lesion biopsy from one of the Group 1 rabbits (Fig 5C and 5D) showed a rich infiltrate of inflammatory cells, particularly eosinophils, and comparable amounts of CD4 and CD8 T-lymphocytes, plus B-lymphocytes (CD20 cells) and plasma cells, as well as follicular inflammation and intra-follicular abscesses (Fig 5C and 5D). Analysis of a second disseminated lesion from the same animal also showed extensive follicular inflammation and an elevated number of eosinophils, although the lymphocyte component could not be evaluated due to a scarcity of cells (not shown). Also, biopsies from carrier-treated animals showed a significant inflammatory infiltrate composed of eosinophils, lymphocytes (CD4, CD8, and CD20), and plasma cells (CD138+ cells) (Fig 5E–5G). Biopsies from control animals also showed elevated numbers of histiocytic cells and blood vessels with a thickened endothelium (Fig 4F and 4G).
In a rabbit model of yaws, GaM demonstrated treponemicidal activity against Tp pertenue is consistent with the previously claimed activity of gallium against T. pallidum [39]. When applied 4 days after inoculation, GaM significantly attenuated lesion development and decreased treponemal burden within lesions developed at the inoculation sites (assessed by both DFM and quantification of Tp-specific mRNA), indicating that GaM is bactericidal when applied locally. Secondary lesions distal to the treated inoculation site still appeared, with detectable treponemes by DFM and a histopathological picture consistent with inflammation, indicating that the predominant effect of topical GaM is localized to the area where applied. Treponemal systemic dissemination from the site of exposure is known to be rapid [59], and topical antimicrobial application was not expected to lead to disease eradication. When applied 14 days after inoculation, GaM promoted lesion healing in one rabbit but not the other. Both of those rabbits, however, showed a significant decrease in treponemal burden by DFM and qPCR. These limited data suggest that yaws lesions that have progressed from the initial erythematous stage may still benefit from GaM treatment. This small pilot study provides justification for conducting larger studies to further investigate GaM as a possible therapeutic agent for yaws. As part of its 2012 plan to overcome the global impact of neglected tropical diseases, the WHO set 2020 as the target year for yaws eradication [60]. The success of the previous anti-yaws campaigns conducted in the 1950`s [27] was due in large part to the efficacy of penicillin against the yaws pathogen, and the apparent inability of this pathogen to develop penicillin resistance. The reasons behind this inability to develop resistance are being elucidated only now [61]. Although penicillin is still the treatment of choice for yaws, in 2012 the yaws eradication strategy was revised to make single-dose oral azithromycin the treatment of choice for mass drug administration. This decision was driven by the difficulties associated with penicillin delivery, which requires a cold chain and trained personnel in loco. Similar to what has happened with syphilis treatment, the introduction of azithromycin has induced the emergence of macrolide-resistant yaws strains, with five epidemiologically related cases in Lihir island, Papua New Guinea [30]. The spread of these azithromycin-resistant strains could therefore undermine the current eradication goal, and make macrolides a decreasingly effective treatment option for yaws, similar to the situation with syphilis. GaM appears to be an attractive alternative treatment for yaws due to its novel mechanism of action (acting as a non-functional mimic of Fe(III), inhibiting bacterial DNA synthesis, for which the development of resistance appears unlikely), and its ability to be administered both locally (topically) and systemically (orally). Because azithromycin clears treponemes from dermal lesions slower than does penicillin [31], we hypothesize that in azithromycin-treated patients, concurrent topical GaM treatment of the yaws lesion during dressing changes may accelerate pathogen clearance from the lesion and consequently accelerate wound healing. This would allow a yaws patient to return to a normal life more quickly as well as decrease the infectivity of the lesion. If shown effective, co-administration of GaM with azithromycin as a “protective drug” could help delay appearance and spreading of genetic resistance to macrolides, greatly extending the time window in which azithromycin could be effectively used for yaws eradication. Systemically administered gallium was claimed in the past to have anti-treponemal activity in a rabbit model of syphilis [39]. We are planning to investigate the efficacy of orally administered GaM in our rabbit model of yaws. GaM has already completed several Phase 1 clinical trials in humans, where high safety and tolerability was demonstrated for this compound [38,62]. In those trials, GaM was administered orally to cancer patients at doses of as much as 3500 mg/day for repeated 28-day cycles; no dose-limiting or other significant toxicity was reported [63]. The topical GaM dose that would be applied to yaws lesions would be about a thousandth of the highest apparently safe oral dose, so the safety is expected to be high.
The major limitation of this study was the small number of laboratory animals used, which did not allow us to make clear conclusions on the efficacy of GaM in accelerating lesion healing when administered to indurated lesions (Group 2 animals). The promising results of this pilot study, however, suggest that the experiments described here should be repeated with groups of at least 8 rabbits each, according to power/sample size calculations. Furthermore, although topical GaM application was shown to have treponemicidal activity, as expected it did not prevent pathogen dissemination and establishment of the infection, even in animals that started treatment as soon as day 4 post-challenge. Our study did not address the efficacy of systemic GaM in eradicating experimental yaws. Additional studies on oral administration of GaM alone or in combination with conventional antimicrobials will need to be performed to fill this knowledge gap. Lastly, a large inoculum (106 cells/injection site) was used to induce lesion development within an acceptable experimental time-frame and to obtain samples in which the treponemal burden could be quantified. Very likely, during natural human infection, significantly fewer treponemes pass the epithelial barrier to cause disease. The use of large inocula has previously been used to evaluate the effectiveness of azithromycin against Tp pallidum, and was not shown to be a confounding factor, and we have no reason to believe it could be in our studies either. Gallium has previously been shown to be effective against many microorganisms in vitro and in animal models [41,42]. This study is the first to extend these observations to Tp pertenue and to report the use of GaM as a topical treatment for yaws. Our results suggest that this compound could be useful as a topical anti-treponemal agent, and justifies further research into the use of GaM as both a topical and an oral agent, alone and/or in combination with other antimicrobials to assess its full potential as a novel anti-yaws compound.
|
10.1371/journal.pntd.0003529 | Epidemiology and Molecular Characterization of Cryptosporidium spp. in Humans, Wild Primates, and Domesticated Animals in the Greater Gombe Ecosystem, Tanzania | Cryptosporidium is an important zoonotic parasite globally. Few studies have examined the ecology and epidemiology of this pathogen in rural tropical systems characterized by high rates of overlap among humans, domesticated animals, and wildlife. We investigated risk factors for Cryptosporidium infection and assessed cross-species transmission potential among people, non-human primates, and domestic animals in the Gombe Ecosystem, Kigoma District, Tanzania. A cross-sectional survey was designed to determine the occurrence and risk factors for Cryptosporidium infection in humans, domestic animals and wildlife living in and around Gombe National Park. Diagnostic PCR revealed Cryptosporidium infection rates of 4.3% in humans, 16.0% in non-human primates, and 9.6% in livestock. Local streams sampled were negative. DNA sequencing uncovered a complex epidemiology for Cryptosporidium in this system, with humans, baboons and a subset of chimpanzees infected with C. hominis subtype IfA12G2; another subset of chimpanzees infected with C. suis; and all positive goats and sheep infected with C. xiaoi. For humans, residence location was associated with increased risk of infection in Mwamgongo village compared to one camp (Kasekela), and there was an increased odds for infection when living in a household with another positive person. Fecal consistency and other gastrointestinal signs did not predict Cryptosporidium infection. Despite a high degree of habitat overlap between village people and livestock, our results suggest that there are distinct Cryptosporidium transmission dynamics for humans and livestock in this system. The dominance of C. hominis subtype IfA12G2 among humans and non-human primates suggest cross-species transmission. Interestingly, a subset of chimpanzees was infected with C. suis. We hypothesize that there is cross-species transmission from bush pigs (Potaochoerus larvatus) to chimpanzees in Gombe forest, since domesticated pigs are regionally absent. Our findings demonstrate a complex nature of Cryptosporidium in sympatric primates, including humans, and stress the need for further studies.
| Cryptosporidium is a common zoonotic gastrointestinal parasite. In a cross-sectional survey of humans, non-human primates (chimpanzees and baboons) and livestock in the Greater Gombe Ecosystem, Tanzania, Cryptosporidium infection rate was 4.3%, 16.0% and 9.6% respectively. Infection was not associated with clinical disease in people; however, living in a household with an infected person increased one’s risk of infection. Phylogenetic analyses identified clusters of Cryptosporidium with a mixed host background. Surprisingly, the Mitumba chimpanzee community, which shares a natural boundary with a human community, had a lower occurrence of C. hominis compared to the Kasakela chimpanzee community, which resides in the forest interior (less human exposure). However, Kasakela chimpanzees were also infected with C. suis, suggesting a transmission cycle linked to sympatric bush pigs. Our findings highlight the complex nature of zoonotic parasite transmission and stress the need for further studies in similar systems.
| Cryptosporidium is one of the most important parasitic diarrheal agents in humans in the world, is among the top four causes of moderate-to-severe diarrheal disease in young children in developing nations, and is problematic as an opportunistic co-infection with HIV due to increased morbidity and mortality [1,2]. Cryptosporidium is well adapted to zoonotic, waterborne, and foodborne transmission, with a life cycle occurring in suitable hosts and transmission by the fecal-oral route [3]. Zoonoses represent the majority of diseases emerging globally with potential to expand to new host systems, [4] yet despite these health threats, few studies have examined the ecology and epidemiology of this pathogen in rural tropical forest systems characterized by high rates of overlap among humans, domesticated animals, and wildlife [5,6].
In Tanzania, agriculture represents over a quarter of the national income and 80 percent of its labor force [7], but natural resources are declining, affected by desertification and soil degradation from recent droughts. This process has resulted in a high rate of loss of forest and woodland habitat [8]. The resulting fragmented landscape increases human-wildlife contact in these areas, elevating the risk for disease transmission. The Greater Gombe Ecosystem (GGE), Tanzania, in particular, is vulnerable to habitat disturbance, and this has both ecological and financial implications since it is home to diverse wildlife, including endangered chimpanzees (Pan troglodytes schweinfurthii), that are important contributors to the national economy through tourism [7].
Gombe National Park, established in 1968, is a small 35-km2 forest reserve located 16-km north of Kigoma in Western Tanzania (4°40′S 29°38′E). The park is 1500-m above sea-level with hills sloping westward from a rift escarpment to Lake Tanganyika [9]. It is home to a number of non-human primate species, including baboons (Pabio anubis), and a well-known wild chimpanzee population studied continuously for over 50 years [10,11]. There are three chimpanzee communities (Kasekela, Mitumba and Kalande); two of which, (Kasekela, and Mitumba) are habituated [12]. The habitat ranges of these two communities overlap slightly permitting opportunity for member contact. Their habitats have differing degrees of human encroachment [13]. Kasekela, the larger community (∼ 65 individuals), is situated at the center of the park in less disturbed forest, whereas Mitumba, the smaller Northern community (∼ 25 individuals), is in close proximity to Mwamgongo (4°40′S, 29°34’60′ E), a village home to ∼5000 inhabitants and their livestock. Another village borders the park to the South, but not along the Eastern ridge, due to high elevation and historic soil depletion. Human presence in the park is limited to researchers, tourists, park management staff, local field assistants and members of their families. The park border is not fenced and therefore villagers and their untethered animals (goats, sheep and dogs) are able to enter the park [14]). Mitumba chimpanzees are frequently reported raiding agricultural fields to the east in the Northern village, Mwamgongo, especially during the dry season (I. Lipende, personal communication). There is little evidence that the chimpanzee population has emigrated outside its established habitat for over 20 years, and immigration events are rare.
Death from infectious diseases is the leading cause of mortality for Gombe chimpanzees [15,16]. The chimpanzees have experienced SIVcpz-associated mortality and morbidity, with SIVcpz prevalence ranging between 9–18% and a 10–16-fold higher age-corrected death hazard for infected individuals [17]. Cryptosporidium is of special concern in this chimpanzee population, as SIVcpz illness may be complicated by Cryptosporidium co-infection, and mirror clinical features observed in human HIV/Cryptosporidium co-infections [2], that report Cryptosporidium infection rates from 8–30% [18,19]. To improve our understanding of this relationship, and highlight potential management options, we investigated risk factors for Cryptosporidium infection and assessed cross-species transmission potential among people, non-human primates, and domestic animals in the GGE, Kigoma District, Tanzania.
This project was reviewed and approved by the Emory University Institutional Review Board (approval #: IRB00018856) under the Expedited review process per 45 CFR 46.110(3), Title 45 CFR Subpart D section 46.404, one parent consent, and 21 CFR 56.110 and the Tanzanian National Institute for Medical Research Institute, Dar Es Salaam, Tanzania, which approved oral consent due to low literacy rates. All adult subjects provided informed consent, and a parent or guardian of any child participant provided informed consent on their behalf. Oral informed consent was obtained by trained local field assistants and documented by witnessed notation on IRB-approved enrollment forms. All animal use followed the guidelines of the Weatherall Report and the NIH Guide for the Care and Use of Laboratory Animals on the use of non-human primates in research, and was approved by the Tanzania Wildlife Research Institute and Tanzania Commission for Science and Technology (permit number 2009-279-NA-2009-184), and the Emory University Animal Care and Use Committee (protocol ID 087-2009). Approval was also obtained from Tanzania National Parks (Permit number TNP/HQ/C10/13) to collect samples from wild chimpanzees. The researchers did not have any interactions with the chimpanzees in the park. All domesticated animals were sampled from households in Mwamgongo village. The owners of the domesticated animals provide verbal consent for the collection of fecal specimens for this study, and the verbal consent was documented. We have included the GPS coordinates for Mwamgongo village at the first mention of the village in the Introduction.
The study period occurred between March 2010 and February 2011. Paired fecal samples from humans and domestic animals were collected during the dry (July 1-August 15) and wet (November 1-December 15) seasons. Human subjects were either residents of Mwamgongo village (estimated population size (n) ∼5000) or Gombe National Park (n ∼100). A baseline demographic survey was performed in June 2010 to identify households within Mwamgongo village with at least one domestic animal species: dog (Canis lupus) n ∼ 8, goat (Capra hircus), n ∼ 150 or sheep (Ovis aries), n ∼ 10. Twenty-five village households with domestic animals were randomly selected for study enrollment. Baboons (n ∼ 198) were opportunistically sampled in Mitumba and Kasekela during these two collection periods. Chimpanzees (n ∼ 90) were sampled in both Mitumba and Kasekela at quarterly intervals during the course of routine observational health monitoring [16].
Specimen cups were provided to enrolled village and park residents. Livestock specimens were aseptically collected by a village veterinary officer. Chimpanzee and baboon specimens were non-invasively collected from identified individuals as part of observational health monitoring. All fecal specimens were freshly voided and aseptically transferred to a screw cap plastic vial containing a 2.5% potassium dichromate solution (Fisher Scientific, Pittsburgh, PA). For baboon and non-human primate samples, care was taken to avoid the collection of soil, foliage or water contaminants, by transferring the interior and top most portion of stool to a collection cup using a sterile wooden spatula or swab and avoiding the collection of fecal material in contact with the ground. Each vial was labeled with a unique identification number, and date of collection. Wildlife samples were additionally labeled with the name of the observer, location and animal name. Samples were sealed with Parafilm (Pechiney Plastic Packaging, Chicago, IL) and stored at 4°C, and shipped in ice to Atlanta, GA United States.
Approximately 1-liter of water was collected in 55-oz sterile Whirl-pak bags (Nasco, Fort Atkinson, WI) and filtered for protozoa using a 0.45-μm Millipore MF-Millipore cellulose ester filter mounted on the Millipore (Billerica, MA) filtration system (diameter 47-mm). When possible, filtration was done on one filter, but in extreme cases where turbidity was high, sequential filtration was performed using two filters. Using sterilized forceps, filters were aseptically transferred to 2-ml cryovials containing a 2.5% potassium dichromate solution. Due to the logistics of field sampling, opportunistic water samples were collected from low, middle and high points of 6 continual streams (dry and wet) and two seasonal streams (wet only). GPS coordinates were obtained using a GPSmap 60CSx from Garmin (Garmin International Inc. Olathe, KS) for each collection point to assist in identifying locations for repeat sampling and if necessary, to assign sampled streams to watersheds associated with specific human or chimpanzee groups.
Nucleic acid was extracted from all fecal specimens and water filters using the FastDNA® SPIN Kit for Soil (MP Biomedicals, LLC, Solon, OH) following the methods described [20]. DNA extracts were subsequently tested using a polymerase chain reaction and restriction fragment length polymorphism (PCR-RFLP) approach where a segment (∼833 bp) of the Cryptosporidium SSU rRNA gene is amplified by nested PCR and then species and genotype diagnosis is made by restriction digestion of the secondary PCR product with SspI (New England BioLabs, Beverly, MA), and either VspI (Promega, Madison, WI) or MboII (New England BioLabs) [21,22]. Each sample was run in duplicate by PCR-RFLP analyses with appropriate controls. Specimens that were positive for Cryptosporidium by the SSU rRNA PCR were confirmed by DNA sequencing of the 18S PCR products (C. suis, C. xiaoi and C. hominis). A subset of specimens positive for C. hominis were also subtyped by sequencing the 60-kilodalton glycoprotein (GP60; ∼900 bp) in both directions on an ABI 3130 Genetic Analyzer (Foster City, CA) [23]. All sequences obtained were aligned with reference sequences using MEGA 6.0 or ClustalX software (http://www.clustal.org/) to identify Cryptosporidium species and C. hominis genotypes.
A survey was administered to each human subject focusing on demography, gastrointestinal symptoms (presence or absence within the previous 4 weeks), medication usage, and water usage. To minimize response bias, surveys were administered by trained local field assistants in the national language (Swahili). Data were manually recorded on paper forms, entered into spreadsheets in the computer program Microsoft Excel, and subsequently reviewed for accuracy.
Results were tabulated and compared in Microsoft Excel (Redmond, WA). To control for sample bias we calculated infection rate as the proportion of individuals in each group positive for Cryptosporidium divided by the total number of individuals in each group examined [24]. If a single individual sample was positive for Cryptosporidium, the subject was considered positive for the collection period (the season for most analyses). Statistical analyses were performed in SPSS version 20.0 (SPSS Inc., Chicago IL). Associations between human survey responses and infection status were compared using logistic regression for categorical binary data. Work history (agricultural fields or forest) were combined as a single factor in the final analysis. Odds ratios (OR) with 95% CI were calculated with significance set at 0.05 for all comparisons. Associations between chimpanzee demographic and observational health data and infection status were evaluated using a generalized estimating equation (GEE) method with exchangeable working correlation structure to account for repeat sampling of individuals. The Huber–White sandwich variance estimation technique was used to calculate confidence intervals (CI). In instances where cells contained less than 5 values, Fisher’s exact tests were used to calculate p-values.
Six hundred and eighty-four fecal specimens were screened for Cryptosporidium including 254 human, 99 domestic animal (n = 76 goat, n = 14 sheep, n = 9 dog) and 331 wildlife (n = 251 chimpanzee, n = 80 baboon) specimens. Cryptosporidium spp, were detected by PCR from 40 (5.8%) fecal samples but was not detected in any water samples (n = 42). The infection rate of Cryptosporidium was highest among 21/131 (16.0%) nonhuman primates tested, compared to 7/73 (9.6%) livestock and 8/185 (4.3%) humans. No significant differences in frequency were observed between chimpanzees and baboons (Table 1, Fisher’s exact test p = 0.457) or between the two chimpanzee communities (Table 1, Fisher’s exact test p = 0.7655). Of the 8 cases of Cryptosporidium detected in humans, 7 (87.5%) resided in Mwamgongo village and one (12.5%) in Mitumba camp. No human cases were detected in the Kasekela camp. Sheep had the highest occurrence of Cryptosporidium (22%) compared to goat (9%) and dogs (0%) but small sample size prevented evaluation of significance.
We identified three species of Cryptosporidium (C. hominis, C. suis and C. xiaoi) in this population (Table 1) based on RFLP and sequence analyses of the SSU rRNA gene. C. hominis was detected in all human cases and all 7 cases from sheep and goats were C. xiaoi. Six of the 12 positive chimpanzees from Kasekela were genotyped as C. suis (a Cryptosporidium species predominantly associated with pigs but has been found in a few human cases). This species was not detected in the Mitumba chimpanzee community (Fisher’s two-tailed exact test; p-value = 0.0537). The porcine species was also not found in the specimens from baboons, humans or domestic animals in the village. The remaining Kasekela chimpanzees (n = 6) and the 4 Mitumba chimpanzees had C. hominis. All baboons (n = 5 individuals) also had C. hominis. GP60 subtyping of a subset (n = 16) of the C. hominis positive samples identified a common subtype IfA12G2 in humans and nonhuman primates. The subtype sequence was identical to two sequences in GenBank; a human C. hominis IfA12G2 sequence from South Africa (GenBank accession number JN867334) and a sequence recovered from an olive baboon in Kenya (GenBank accession number JF681172).
We used data from the survey to identify potential risk factors for Cryptosporidium infection (Table 2). Among 95 respondents (100%), villagers in Mwamgongo were at greater risk for infection when living with a person who was positive for Cryptosporidium (OR = 9.722; 95% CI 1.741–54.279; p = 0.011). Persons living with Cryptosporidium-positive livestock tended to have a greater odds of infection (OR = 4.750; 95% CI 0.944–23.908; p = 0.059). Other factors related to behaviors, including location, occupation (either agricultural or forestry), and not boiling water for consumption were not statistically significant. Although presence of clinical signs was not statistically significant, when reviewing survey data for the Cryptosporidium positive patients, 4/8 reported having diarrhea; 2 sought treatment at the village clinic (Flagyl and Paracetemol). Four of 8 households (50%) reported at least one additional member of the household experiencing gastrointestinal symptoms, including diarrhea and cramping. Interestingly, affected individuals did not report consuming water from an open source (100%), which appeared protective (OR = 0.156; 95% CI 0.085–2.875) but not statistically significant (p = 0.162). 4% of study respondents reported boiling their water before use. No infected individuals reported boiling their drinking water. An association between season and Cryptosporidium infection was not observed in humans or non-human primates (Tables 2 and 3). Chimpanzee demographic factors such as age and sex were not risk factors for Cryptosporidium infection and evidence of diarrhea was not a reliable predictor of Cryptosporidium illness. Kasekela chimpanzees tended to have a higher likelihood (OR = 7.062; 95% CI 0.398–125.251, p = 0.07) of infection with C. suis as compared to the Mitumba community (Table 3).
Of 90 humans residing in camps within Gombe National Park, only one Cryptosporidium infection was observed (1%). In contrast, Cryptosporidium infection in residents of Mwamgongo village reached 10% during the drier months. These frequencies are comparable to those reported in some studies from children and adults without HIV, ranging from 0–18% [25–27], though frequencies as high as 32% have been reported elsewhere [28]. Site and HIV prevalence would seem to be important factors in human occurrence rates. Although HIV testing was beyond the scope of this study, a recent country report [9] indicates a low HIV prevalence (< 1%) from the Kigoma region where the study occurred. Infection was not statistically associated with gastrointestinal illness or stool consistency for humans or wildlife, findings consistent with earlier studies [28–32]. Surprisingly, unsafe drinking water (i.e. untreated/unboiled water, open water source) was not found to increase risk of Cryptosporidium infection, which could be the result of degradation or improper tapping of the water line, exposing water to environmental contamination [33]. This warrants further study as consumption of contaminated ground water has been repeatedly associated with Cryptosporidium infection [34,35].
Cryptosporidium was detected in baboons (11%) and the two chimpanzee communities (15–21%) at moderate frequencies. Similar frequencies of Cryptosporidium have been detected previously in other African primates. Habituated mountain gorillas in Bwindi Impenetrable National Park (BINP), Uganda had Cryptosporidium in (11%) of specimens sampled; 73% of positive specimens were detected from human-habituated gorillas [36]. Genetic characterization of this population determined that the gorillas and the local human community both carry C. parvum [37]. Eleven percent of red colobus and black-and-white colobus in Kibale National Park (KNP), Uganda were infected with Cryptosporidium; genetic sequences from some humans and colobus from KNP were identical, while two strains from red colobus in the forest interior were infected with a divergent subclade, suggesting the possibility of separate zoonotic and sylvatic cycles [28]. These findings support the zoonotic spillover potential of Cryptosporidium among humans, wildlife and livestock.
Similar patterns have been observed with other directly or environmentally transmitted enteric pathogens. At KNP, Giardia was detected in red colobus in forest fragments (5.7%), but not detected in undisturbed forest (0%) [38]. Genetic analyses of recovered strains determined that red colobus were infected with Giardia duodenalis assemblages associated with humans and livestock, suggesting complex cross-species transmission [39]. Escherichia coli strains from gorillas in BINP and chimpanzees in KNP that overlapped with humans were more similar to strains collected from resident humans and livestock compared to strains collected from gorillas and chimpanzees living in undisturbed forest [40,41]. Additionally, in KNP, genetic similarity between human/livestock and primate bacteria increased three-fold with a moderate to high increase in anthropogenic disturbance of forest fragment. Bacteria harbored by humans and livestock were more similar to those of monkeys that entered human settlements to raid crops, than to bacteria of other primate species [42]. These findings reinforce the notion that habitat overlap and anthropogenic disturbance increase the risk of interspecies transmission between wildlife, humans and livestock and that transmission can occur both via direct physical contact with or ingestion of contaminated feces and by indirect exposure via a shared (potentially contaminated) watershed. Although water samples screened in this study were negative for the parasite, waterborne outbreaks of Cryptosporidium as a result of human and animal fecal contamination are common [43–45]. Watershed sampling for Cryptosporidium in this study was opportunistic using smaller volumes of water than advocated by standard screening protocols [46] and provided only limited inter-seasonal sampling [47–49]. Therefore, our negative results do not assure that waterborne transmission is not important in this system. Future studies using more comprehensive watershed sampling would help to resolve this aspect of Cryptosporidium transmission.
The results of our RFLP and sequence analyses of the SSU rRNA gene suggest multiple potential zoonotic pathways for Cryptosporidium transmission in this study system. The village data reinforces our understanding that species of Cryptosporidium vary in their zoonotic potential [50]. Our data suggest that there is less likelihood that the C. xiaoi affecting the livestock is capable of causing illness in humans, considering the close proximity of livestock to humans in this community [i.e., animals often residing in homes], where animal-human contact is quite high though it has been documented in other studies [51]. Cryptosporidium hominis was also not detected among the domesticated animals in this study, but C. hominis has been found in other studies to be a zoonotic species, affecting both humans and domesticated animals [52,53]. Homes with positive livestock had a tendency for increased risk of human infection suggesting contribution of environmental factors or behaviors that may place the household at increased risk.
We anticipated that the Kasekela chimpanzee community would have a lower occurrence of Cryptosporidium compared to Mitumba, which shares a natural border with Mwamgongo village. Although not statistically significant, there was a higher frequency of Cryptosporidium recovered in Kasekela. The occurrence of C. hominis was not statistically higher for Mitumba compared to Kasekela. However, the results demonstrated that 50% (6/12) of the Cryptosporidium recovered from the Kasekela community are C. suis, a species associated with pigs. Domestic pigs are not found in the park or village due to religious preference (predominantly Muslim communities). However, bush pigs (Potaochoerus larvatus), native to the Gombe forest habitat, are common in Kasekela but have not been observed frequently in Mitumba forest. Thus bush pigs may serve as the reservoir host for C. suis in this system. Chimpanzees may be infected through contact with this animal (i.e., chimpanzees occasionally consume bush pigs) or by indirect contact with infective feces on the forest floor or the contamination of shared water sources. This putative pathway of transmission is supported by the fact that C. suis was not recovered in the Mitumba chimpanzee community, the baboons, domesticated livestock or village inhabitants. C. suis has zoonotic potential having been previously identified in an HIV+ patient in Lima, Peru, [54] and from patients in Henan, China and England [55,56]. Similarly, Salyer et al [28] found that while Cryptosporidium may be transmitted frequently among domesticated animals, humans and wildlife in areas of overlap, there may be host-parasite specific dynamics that occur in the absence of these regular interactions creating separate transmission cycles.
The appearance of the C. hominis species and its subtype IfA12G2C among the humans, baboons and chimpanzee communities demonstrates the zoonotic transmission potential of this parasite species among these closely related host species and points to a dominance of anthropozonotic transmission in this system. This subtype has been previously reported among captive olive baboons in Kenya [57], and is prevalent in humans in Africa [58]. Additionally, common primate behaviors may increase the likelihood for animal to human transmission. For example, baboons raid camp food reserves and homes, potentially transmitting etiologic agents via infected feces to humans. The Mitumba chimpanzees are reported to raid agricultural fields just outside the park boundaries, which can transmit diseases from the potentially contaminated feces of livestock or exposed human sewage.
Our results are based on small sample sizes, that if increased could alter the frequency and predictions of Cryptosporidium subtypes, infection and illness. The finding of C. suis in the Kasekela chimpanzee community is novel. We presume these chimpanzees experienced cross-species transmission from bush pigs in Gombe forest, because domesticated pigs are absent from the area. Unfortunately, we do not have access to fecal specimens from the local bush pig population to compare to specimens recovered from Kasekela chimpanzees. Despite the high overlap observed between people and livestock in villages in this region, our results suggest that the transmission dynamics of Cryptosporidium for humans and livestock are distinct but the dominance of C. hominis in humans and non-human primates suggest the potential for cross-species transmission. Our findings highlight the complex nature of zoonotic parasite transmission and stress the need for further studies in similar systems.
|
10.1371/journal.pgen.1007792 | Bacterial group II introns generate genetic diversity by circularization and trans-splicing from a population of intron-invaded mRNAs | Group II introns are ancient retroelements that significantly shaped the origin and evolution of contemporary eukaryotic genomes. These self-splicing ribozymes share a common ancestor with the telomerase enzyme, the spliceosome machinery as well as the highly abundant spliceosomal introns and non-LTR retroelements. More than half of the human genome thus consists of various elements that evolved from ancient group II introns, which altogether significantly contribute to key functions and genetic diversity in eukaryotes. Similarly, group II intron-related elements in bacteria such as abortive phage infection (Abi) retroelements, diversity generating retroelements (DGRs) and some CRISPR-Cas systems have evolved to confer important functions to their hosts. In sharp contrast, since bacterial group II introns are scarce, irregularly distributed and frequently spread by lateral transfer, they have mainly been considered as selfish retromobile elements with no beneficial function to their host. Here we unveil a new group II intron function that generates genetic diversity at the RNA level in bacterial cells. We demonstrate that Ll.LtrB, the model group II intron from Lactococcus lactis, recognizes specific sequence motifs within cellular mRNAs by base pairing, and invades them by reverse splicing. Subsequent splicing of ectopically inserted Ll.LtrB, through circularization, induces a novel trans-splicing pathway that generates exon 1-mRNA and mRNA-mRNA intergenic chimeras. Our data also show that recognition of upstream alternative circularization sites on intron-interrupted mRNAs release Ll.LtrB circles harboring mRNA fragments of various lengths at their splice junction. Intergenic trans-splicing and alternative circularization both produce novel group II intron splicing products with potential new functions. Overall, this work describes new splicing pathways in bacteria that generate, similarly to the spliceosome in eukaryotes, genetic diversity at the RNA level while providing additional functional and evolutionary links between group II introns, spliceosomal introns and the spliceosome.
| A hallmark of eukaryotic cells is the ability of their numerous introns, sequences that interrupt genes, to generate genetic diversity through the expression of several different protein variants from a single gene. These eukaryotic introns share a common ancestor with bacterial group II introns, which are mobile, scarce and maintained at low copy-levels. Because of this stark contrast, bacterial group II introns are considered as selfish mobile elements with no beneficial function to their hosts. As a challenge to this longstanding view, we describe here a new group II intron function that expands the genetic diversity and overall complexity of its bacterial host transcriptome. This ability of combining disparate mRNA fragments expands the repertoire of functional products at the disposal of the bacterial host. Our work thus provides a potential explanation of why group II introns were maintained in bacteria over the course of evolution, despite strong selective pressure to streamline their genomes. Moreover, generating new combinations of RNA molecules as a by-product of their otherwise selfish mobility pathway provides a new evolutionary link into how group II introns later evolved to fully support this function in the nuclei of eukaryotes.
| Bacterial group II introns are large RNA enzymes that mostly behave as retromobile elements [1–5]. Following their autocatalytic excision from interrupted RNA transcripts, they can reinsert within identical or similar DNA target sequences by retrohoming or retrotransposition, respectively [6–8]. These retromobile genetic elements are present in archaea, bacteria, and bacterial-derived organelles such as plant and fungal mitochondria, and plant chloroplasts [9]. While group II introns are somewhat infrequent in archaea, roughly one quarter of all sequenced bacterial genomes harbor one to a few copies displaying a broad phylogenetic distribution in the bacterial kingdom [10]. In sharp contrast, no functional group II introns were yet described in the nuclear genome of eukaryotes where they seem to be functionally excluded [11]. Although mitochondrial and chloroplastic group II introns mainly interrupt housekeeping genes, bacterial group II introns are generally found in non-coding sequences and associated with other mobile genetic elements [5]. Organellar group II introns thus primarily function as classic intervening sequences while bacterial group II introns behave like mobile elements. Bacterial group II introns were also shown to propagate by conjugation within and between species, invading the chromosome or resident plasmids of their new hosts using either the retrohoming or retrotransposition pathways [12–14].
Group II introns require the assistance of RNA binding proteins called maturases to adopt their active three-dimensional conformation and self-splice in vivo [15]. Specific sequence motifs within group IIA introns mediate the accurate recognition of the 5’ and 3’ splice sites. Exon binding sequence 1 (EBS1) and 2 (EBS2) identify the 5’ splice site by base pairing with complementary intron binding sequence 1 (IBS1) and 2 (IBS2) situated at the 3’ extremity of the upstream exon. The 3’ splice site is recognized by the ∂-∂’ base paring interaction at the 5’ extremity of the downstream exon. Group II introns self-splice from interrupted RNA transcripts through three different splicing pathways (Fig 1) [15]. The branching (Fig 1A), hydrolysis (Fig 1B) and circularization (Fig 1C) pathways release the intron as either branched structures called lariats, in linear forms or as closed circles, respectively. Each of these three splicing pathways involve two consecutive transesterification reactions (Fig 1, steps 1 and 2). Branching, however, is the only splicing pathway that is completely reversible where intron lariats can recognize single- and double-stranded nucleic acid substrates (RNA/DNA) through base pairing and reinsert themselves by reverse splicing (Fig 1A, double arrows) [15, 16]. Since reverse splicing is the initial step of both group II intron mobility pathways, retrohoming and retrotransposition, only released intron lariats are active mobile elements [16].
We recently unveiled and characterized at the molecular level the circularization pathway of Ll.LtrB, the model group II intron, from the gram-positive bacterium Lactococcus lactis [17, 18]. Our work showed that the intron excises simultaneously through the branching and circularization pathways in vivo leading to the accumulation of both intron lariats and circles respectively. While the majority of the excised intron circles were found to have their 5’ and 3’ ends perfectly joined, we identified Ll.LtrB RNA circles harboring additional nucleotides at their splice junction. Here we describe novel group II intron splicing pathways in which the release of intron circles, harboring or not mRNA fragments of various lengths at their splice junctions, occurs concurrently with the generation of intergenic E1-mRNA and mRNA-mRNA chimeras in vivo. Overall, this study unveils that, similarly to spliceosomal introns in eukaryotes, bacterial group II introns generate genetic diversity at the RNA level, producing novel splicing products with potential new functions.
To study the splicing pathway leading to the incorporation of additional nucleotides at the splice junction of group II intron circles [17] we performed an RT-PCR reaction across the Ll.LtrB-ΔLtrA+LtrA lariat and circle splice junctions (Fig 2) [17, 18]. We cloned and sequenced the amplicons located in the faint smear above the RT-PCR band that corresponds to perfect lariat and circle splice junctions (Fig 2C). They revealed excised intron circles harboring additional nucleotides (nts) between the first and the last nts of the intron (Fig 3A). The stretch of additional nts greatly varied in size (20–576 nts), originated from the L. lactis chromosome or the two plasmids used to express the intron (Fig 2A) [17, 18] and mapped to the transcribed strand of annotated genes. Some sequences were identified more than once while others corresponded to different portions of the same gene.
Additional nts within the same size range (26–593 nts) and with identical characteristics (S1 Fig) were identified at the circle splice junction of Ll.LtrB-WT (Fig 2C). Taken together, these data show that mRNA fragments are incorporated at the splice junction of Ll.LtrB RNA circles during circularization, regardless if LtrA, the intron-encoded protein, is expressed in trans (Fig 3A) or in cis (S1 Fig).
The flanking sequences on both sides of the mRNA fragments incorporated at the Ll.LtrB-ΔLtrA+LtrA (Fig 3A) and Ll.LtrB-WT (S1 Fig) circle splice junctions were retrieved, compiled and analyzed. Directly upstream from the 5’ and 3’ junctions we identified IBS1/2-like sequences partly complementary to the EBS1/2 sequences for both introns (Figs 3A and S1). Consensus sequences of 30 nts spanning the 5’ and 3’ junctions of the mRNA fragments confirmed the presence of IBS1/2-like sequence motifs. The IBS1-like motifs are better defined than the IBS2-like motifs, whereas the upstream IBS1/2-like motifs are stronger for both Ll.LtrB-ΔLtrA+LtrA and Ll.LtrB-WT (Fig 4A–4C).
Comparable mRNA fragments of various lengths (43–452 nts)(Fig 3B) were also found at the circle splice junction of Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA (Fig 2C), for which the EBS1 sequence was modified from 5’-GUUGUG-3’ to 5’-CAACAC-3’ (Fig 4D). Accordingly, the IBS1-like consensus sequence motifs upstream from both mRNA junctions were found to be different from Ll.LtrB-ΔLtrA+LtrA and Ll.LtrB-WT and complementary to the mutated EBS1 sequence (Fig 4E). In addition, similarly to Ll.LtrB-ΔLtrA+LtrA and Ll.LtrB-WT, the IBS1-like sequence motifs are both better defined than the IBS2-like motifs and the upstream IBS1/2-like motif much stronger.
The base pairing potential of Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA is more stringent than Ll.LtrB-ΔLtrA+LtrA and Ll.LtrB-WT because its EBS1 sequence (5’-CAACAC-3’) can perfectly recognize only 1 sequence (5’-GUGUUG-3’). In contrast, both introns harboring the wild-type EBS1 sequence (5’-GUUGUG-3’) can base pair perfectly with 64 different sequence combinations using G = U wobble base pairings. Consequently, the more stringent EBS1 sequence led to a fainter RT-PCR smear (Fig 2C), the identification of fewer mRNA fragments at the intron circle splice junction (Fig 3B), and to much stronger flanking consensus motifs when compared to Ll.LtrB-ΔLtrA+LtrA and Ll.LtrB-WT (Fig 4). These data confirm that both junctions of the incorporated mRNA fragments at intron circle splice junctions are recognized by the EBS1/2 motifs of Ll.LtrB through base pairing interactions during circularization.
Consensus sequences are slightly but consistently stronger when flexibility is allowed at both junctions of the mRNA fragments for all three constructs suggesting that Ll.LtrB does not always process mRNAs precisely downstream from the recognized IBS1/2-like motifs (S2–S5 Figs). We also identified mRNA fragments, at intron circle splice junctions, that either contained untranslated sequences or spanned two genes including the short intergenic regions of polycistronic mRNAs (Figs 3 and S1). This further supports our conclusion that Ll.LtrB can capture L. lactis transcripts at intron circle splice junctions during circularization.
Our findings indicate that cellular mRNAs can somehow be incorporated at the Ll.LtrB circle splice junction during the circularization pathway. Two models can explain how mRNA fragments could be incorporated at the splice junction of group II intron circles (Fig 5).
The external nucleophilic attack pathway (Fig 5A) was previously proposed to explain how short stretches of additional nts could be incorporated at the circle splice junction during intron circularization. However, the pathway of integration and the origin of the additional nts were never demonstrated [17, 19]. Taking into consideration the data presented here, Ll.LtrB would recognize, through base pairing interactions, an IBS1/2-like sequence on an L. lactis mRNA and guide its hydrolysis downstream of the recognized sequence (step 1). Next, the 3’-OH of the processed mRNA would induce a transesterification reaction at the exon 1-intron splice junction resulting in its ligation to the 5’ end of the intron and the release of exon 1 (step 2). The 3’-OH of exon 1 would then initiate the next transesterification reaction at the intron-exon 2 splice junction, releasing ligated exons and a linear intron harboring an mRNA fragment at its 5’ end (step 3). The final transesterifictaion reaction would be induced at the intron 5’ end (step 4a) or within the mRNA (step 4b) by the 2’-OH of the last nt of the linear intron, just downstream from IBS1/2-like sequences, resulting in the release of either a head-to-tail circular intron or an intron circle harboring an mRNA fragment at its splice junction respectively.
An alternative pathway (Fig 5B) would rather be initiated by the reverse splicing of an intron lariat within an L. lactis mRNA downstream of an IBS1/2-like sequence (step 1). The ectopically inserted group II intron would then excise from the mRNA through circularization (steps 2–4). The 3’-OH of free exon 1 would first attack the phosphodiester bond at the 3’ splice site between the last nt of the intron and the 3’ segment of the mRNA (step 2). This would generate a chimeric mRNA consisting of the ltrB-exon 1 (E1) linked to the 3’ segment of the mRNA (E1-mRNA) and a circularization intermediate where the linear intron is still attached to the 5’ segment of the mRNA. The final transesterifictaion reaction would then be induced at the intron 5’ end (step 3a) or within the mRNA fragment (step 3b) by the 2’-OH of the last nt of the intron, just downstream from IBS1/2-like sequences, resulting in the release of either a head-to-tail circular intron or an intron circle harboring an mRNA fragment at its splice junction respectively.
To investigate the proposed models we looked for unique intermediates of the reverse splicing pathway: the 3’ junction of Ll.LtrB reverse-spliced within mRNAs and chimeric E1-mRNAs (Fig 5B, asterisks). We first detected by RNA-Seq intron-interrupted mRNAs for Ll.LtrB-ΔLtrA+LtrA and Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA but not for the Ll.LtrB-ΔA-ΔLtrA+LtrA control which lacks the essential branch point A residue required for branching and reverse splicing [17, 18, 20] (Fig 6A). The reverse splice sites of Ll.LtrB-ΔLtrA+LtrA (Fig 6B) and Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA (Fig 6C) were shown to be immediately preceded by consensus IBS1/2-like sequence motifs complementary to their respective EBS1/2 sequences. On the other hand, similarly to the junctions between intron circles and mRNA fragments (Figs 4 and S5), we did not detect a ∂’-like sequence on the 3’ side of the intron insertion sites (Fig 6). This shows that Ll.LtrB can recognize IBS1/2-like sequences on various mRNAs by base pairing with its EBS1/2 sequences and invade them by reverse splicing, generating a population of intron-interrupted mRNAs in L. lactis. As expected, the more stringent EBS1 sequence of Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA led to the identification of fewer intron-interrupted mRNAs and a stronger IBS1/2-like consensus sequence upstream of the intron insertion sites compared to Ll.LtrB-ΔLtrA+LtrA.
We next studied in further details the reverse splicing of Ll.LtrB-ΔLtrA+LtrA within the Enolase (enoA) and Alanyl-tRNA synthetase (alaS) mRNAs. The enoA (167 nts) and alaS (304 nts) mRNA fragments, previously identified at the Ll.LtrB-ΔLtrA+LtrA circle splice junction, are both flanked by a strong (10/11 nts) and a weak (7/11 and 8/11 nts) IBS1/2-like sequence motif (Fig 3A). We amplified by RT-PCR the 5’ (Fig 7A and 7F) and 3’ (Fig 7B and 7E) junctions between the intron and the two mRNAs. Sequences of the four amplicons confirmed reverse splicing of the intron precisely downstream of the strong IBS1/2-like sequence within the enoA (Fig 7C, large open arrowhead) and alaS (Fig 7G, large open arrowhead) mRNAs. Importantly, no amplifications were detected for the reverse splicing deficient control, Ll.LtrB-ΔA-ΔLtrA+LtrA. Next, the faint smears above (Fig 7A and 7E) and below (Fig 7B and 7F) the main amplicons were cloned and shown to correspond to several independent 5’ and 3’ junctions of the intron inserted downstream of different weak IBS1/2-like sequences (7-9/11 nts)(Fig 7C and 7G, black and grey arrowheads). The weak IBS1/2-like sequences flanking the mRNA fragments previously identified within intron circles (Fig 3A), were also found invaded by the intron for both enoA (Fig 7C, small open arrowhead) and alaS (Fig 7G, small open arrowhead). Similarly, the Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA variant was shown to reverse splice at specific strong and weak IBS1/2-like sequences within the S12/S7 transcript (Fig 8A–8C). The identified reverse splice sites also include the strong and weak IBS1/2-like sequences flanking the S12/S7 mRNA fragment (161 nts) previously identified at the intron circle splice junction (Fig 3B).
Collectively, these results show that IBS1/2-like sequences are widespread within L. lactis mRNAs, providing abundant targets for Ll.LtrB reverse splicing. They also support the proposed alternative circularization model by which introns, reverse-spliced at ectopic sites within mRNAs, can circularize alternatively by recognizing upstream IBS1/2-like sequences leading to the capture of mRNA fragments at their splice junction (Fig 5B, step 3b). Accordingly, when additional nts are found at intron circle splice junctions, the upstream IBS1/2-like consensus sequences are consistently stronger (Figs 4 and S5) suggesting that when the intron reverse splices at a weak IBS1/2-like sequence, it is more likely to release intron circles harboring mRNA fragments by recognizing a stronger upstream alternative IBS1/2-like sequence.
The second distinguishing splicing intermediate between the two proposed models is a chimeric mRNA consisting of ltrB-exon 1 (E1) trans-spliced to an L. lactis mRNA fragment (E1-mRNA) (Fig 5B, asterisk). We specifically screened for E1-enoA and E1-alaS mRNA chimeras by RT-PCR. In both cases we detected, exclusively for the reverse splicing-competent intron, E1-mRNA chimeras ligated precisely downstream from the strong IBS1/2-like sequences (Fig 7D and 7H), the exact sites previously identified at one of the extremities of the mRNA fragments identified at intron circle splice junctions (Fig 3A) and invaded by reverse splicing (Fig 7A, 7B, 7E and 7F). The intron-catalyzed EBS1/2-specific generation of E1-mRNA chimeras was corroborated with the Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA variant again at the previously identified strong IBS1/2-like sequence of the S12/S7 transcript (Fig 8D). These results show that Ll.LtrB, reverse-spliced at IBS1/2-like sequences of various mRNAs, can recruit free E1 through EBS-IBS base pairing interactions, and catalyze the formation of E1-mRNA chimeras.
Ll.LtrB splicing via circularization, from a population of intron-interrupted mRNAs, generates processed mRNA fragments harboring IBS1/2-like sequences at their 3’ end (Fig 5B, step 3a and 3b). We next examined if these splicing products could be recruited by Ll.LtrB, similarly to free E1 through EBS-IBS base pairing, and used to generate intergenic mRNA-mRNA chimeras (Fig 9). We detected by RT-PCR both alaS-enoA (Fig 7I) and enoA-alaS (Fig 7J) mRNA-mRNA intergenic chimeras joined at specific IBS1/2-like sequences for Ll.LtrB-ΔLtrA+LtrA but not for the Ll.LtrB-ΔA-ΔLtrA+LtrA control. These data show that Ll.LtrB, reverse-spliced within various mRNAs, can recruit through base pairing processed mRNA fragments, harboring IBS1/2-like sequences at their 3’ end, to initiate the circularization splicing pathway (Fig 9, step 2). Ll.LtrB can thus catalyze the shuffling of coding sequences within a population of intron-interrupted mRNAs by a new intergenic trans-splicing pathway (Fig 9).
One quarter of currently sequenced bacterial genomes harbor one to a few group II introns [10]. This paucity, coupled with their irregular distribution and frequent lateral transfer [4], has led to the suggestion that they are selfish retromobile elements with no beneficial function to their host [5]. In contrast, many group II intron derivatives provide important functions in both eukaryotes and prokaryotes [1–3]. For example, the abundant spliceosomal introns, descendants of group II introns, generate significant genetic diversity and transcriptomic complexity via alternative splicing [21], intergenic trans-splicing [22], RNA circle formation [23] and by creating new genes through exon shuffling [24].
Even though the Ll.LtrB group II intron is present at only one copy in the L. lactis genome, the new splicing pathways described here (Figs 5B and 9) expand the genetic diversity and complexity of its host transcriptome. This stems from the ability of Ll.LtrB, following its release as RNP particles, to generate a population of intron-interrupted mRNAs through reverse splicing, which we were able to detect by RNA-Seq (Fig 6) and gene-specific RT-PCR (Figs 7 and 8). Ll.LtrB was recently shown to interact with its cognate ligated exons at the IBS1/2 site in vivo, leading to either complete reverse splicing or negative regulation of targeted mRNA through hydrolysis and degradation [25]. However, when we contrasted the counts per million (CPM) of Ll.LtrB-WT, Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA and Ll.LtrB-ΔA-ΔLtrA+LtrA constructs for alaS, the most abundant target for reverse splicing that we identified by RNA-Seq, we obtained differential expression ratios that showed very little change in the abundance of the alaS transcript: 0.97 between Ll.LtrB-WT and EBS1/Mut-ΔLtrA+LtrA and 0.96 between Ll.LtrB-WT and Ll.LtrB-ΔA-ΔLtrA+LtrA. This suggests that the IBS1/2-like sites we identified within host mRNAs are not efficient targets for hydrolysis, but rather seem to be used for reverse splicing. Interestingly, several of the reverse-splicing sites found by RNA-Seq were also identified independently at the extremity of mRNA fragments captured at intron circle splice junctions (Figs 3A and S1), yet there was only a small overlap of IBS1/2-like motifs between these two sets of data. Moreover, when we analysed the enoA and alaS genes in greater detail, we found a multitude of additional IBS1/2-like motifs that were used as targets for Ll.LtrB reverse splicing and whose base paring interactions with the intron varied from strong (11/11 nts) to weak (7/11 nts) (Fig 7C and 7G). Overall, our data thus suggest that the reverse-splicing of group II introns into ectopic sites within host mRNAs is a widespread, dynamic and transient process whose exact scope is hard to determine.
We demonstrated that circularization of Ll.LtrB from interrupted mRNAs, using free E1 or mRNA fragments harboring IBS1/2-like sequences at their 3’ end, generates two types of trans-spliced transcripts: E1-mRNA (Fig 5B) and mRNA-mRNA (Fig 9) chimeras respectively. Ll.LtrB was recently found to generate free E1 in vivo through hydrolysis of ligated cognate exons at the IBS1/2 site [25]. This Spliced Exon Reopening (SER) reaction (Fig 1) could thus produce the initial source of E1 required to initiate Ll.LtrB circularization from both its cognate exons and ectopic insertion sites. In addition, we found that alternative circularization of Ll.LtrB from interrupted mRNAs releases intron circles harboring mRNA fragments at their splice junction (Fig 5B, step 3b). These novel bacterial splicing products, generated by alternative circularization and intergenic trans-splicing, may have and/or lead to novel biological functions for their host cell. For instance, chimeric RNAs, intron circles and different circular RNAs that accumulate in vivo have been recently associated to a variety of interesting new functions such as RNA sponges, protein sponges and transcriptional regulators in various biological systems [23, 26, 27]. Moreover, the trans-spliced E1-mRNA and mRNA-mRNA chimeras could be reclaimed by the host and potentially lead to the creation of new genes. Group II introns may thus serve a beneficial function for their hosts by increasing the complexity and genetic diversity of their transcriptomes (Fig 10) which could explain why they were retained in bacteria.
Our work also unveils two additional functional and evolutionary links between group II introns, spliceosomal introns and the spliceosome. First, the trans-splicing of E1 at the 5’ end of various mRNA fragments is analogous to the second step of the spliced leader (SL) trans-splicing pathway, which has a patchy evolutionary distribution amongst eukaryotes and whose origin has remained enigmatic [28, 29]. Second, we showed that group II introns, similarly to the spliceosome [22], can catalyze the trans-splicing of intergenic mRNA-mRNA chimeras in bacteria. Since group II introns are considered as the progenitors of both spliceosomal introns and the snRNAs of the spliceosome [1–3], our findings suggest that the spliceosome-dependent formation of SL trans-spliced transcripts and intergenic mRNA-mRNA chimeras in eukaryotes both consist of ancient group II intron splicing functions still shared with their contemporary bacterial relatives.
Overall, we described here new group II intron splicing pathways that generate and expand the genetic diversity and complexity of its host transcriptome which represents a new function for these bacterial retroelements. Our work also unveils new functional and evolutionary links with their nuclear relatives in eukaryotes, and provide a potential explanation of why group II introns were maintained in bacteria.
Lactococcus lactis strain NZ9800ΔltrB (TetR) [8] was grown in M17 media supplemented with 0.5% glucose (GM17) at 30°C without shaking. The Escherichia coli strain DH10β, used for cloning purposes, was grown in LB at 37°C with shaking. Antibiotics were used at the following concentrations: chloramphenicol (CamR), 10 μg/ml; spectinomycin (SpcR), 300 μg/ml. Previously constructed plasmids (pDL-P232-Ll.LtrB-ΔLtrA [30], pDL-P232-Ll.LtrB-WT [30], pLE-P232-LtrA [31]) were used to study Ll.LtrB splicing. Additional variants were constructed by site-directed mutagenesis (New England Biolabs Q5 Site-Directed-Mutagenesis Kit): pDL-P232-Ll.LtrB-ΔA-ΔLtrA, pDL-P232-Ll.LtrB-EBS1/Mut-ΔLtrA. The alanyl tRNA synthetase (alaS) and enolase (enoA) genes were cloned in pLE-P232-LtrA (BssHII), downstream of the second P23 promoter, and expressed with the intron in a two-plasmid system. Primers used for mutagenesis and cloning are in S2 Table.
Total RNA was isolated from NZ9800ΔltrB harboring various plasmid constructs as previously described [31]. RT-PCR reactions [17, 18] were performed on total RNA preparations of NZ9800ΔltrB harboring various intron constructs (primers in S2 Table).
RNA-Seq was performed on rRNA-depleted total RNA from L. lactis (NZ9800ΔltrB) expressing Ll.LtrB-ΔLtrA+LtrA, Ll.LtrB-ΔA-ΔLtrA+LtrA, or Ll.LtrB-EBS1/Mut-ΔLtrA+LtrA using the Illumina HiSeq 2500 paired-end sequencing system [32].
Aligned and adjusted consensuses were prepared using the WebLogo software [33]. Adjusted consensuses were determined using a code that calculated contiguous nucleotides with the highest capacity of base pairing to EBS1 and EBS2 (total of 11 nucleotides), separated from each other by 0–2 nucleotides, in a region that spanned -14, +4 nts around the junction with the intron.
|
10.1371/journal.ppat.1004599 | Uropathogenic Escherichia coli Superinfection Enhances the Severity of Mouse Bladder Infection | Urinary tract infections (UTIs) afflict over 9 million women in America every year, often necessitating long-term prophylactic antibiotics. One risk factor for UTI is frequent sexual intercourse, which dramatically increases the risk of UTI. The mechanism behind this increased risk is unknown; however, bacteriuria increases immediately after sexual intercourse episodes, suggesting that physical manipulation introduces periurethral flora into the urinary tract. In this paper, we investigated whether superinfection (repeat introduction of bacteria) resulted in increased risk of severe UTI, manifesting as persistent bacteriuria, high titer bladder bacterial burdens and chronic inflammation, an outcome referred to as chronic cystitis. Chronic cystitis represents unchecked luminal bacterial replication and is defined histologically by urothelial hyperplasia and submucosal lymphoid aggregates, a histological pattern similar to that seen in humans suffering chronic UTI. C57BL/6J mice are resistant to chronic cystitis after a single infection; however, they developed persistent bacteriuria and chronic cystitis when superinfected 24 hours apart. Elevated levels of interleukin-6 (IL-6), keratinocyte cytokine (KC/CXCL1), and granulocyte colony-stimulating factor (G-CSF) in the serum of C57BL/6J mice prior to the second infection predicted the development of chronic cystitis. These same cytokines have been found to precede chronic cystitis in singly infected C3H/HeN mice. Furthermore, inoculating C3H/HeN mice twice within a six-hour period doubled the proportion of mice that developed chronic cystitis. Intracellular bacterial replication, regulated hemolysin (HlyA) expression, and caspase 1/11 activation were essential for this increase. Microarrays conducted at four weeks post inoculation in both mouse strains revealed upregulation of IL-1 and antimicrobial peptides during chronic cystitis. These data suggest a mechanism by which caspase-1/11 activation and IL-1 secretion could predispose certain women to recurrent UTI after frequent intercourse, a predisposition predictable by several serum biomarkers in two murine models.
| Urinary tract infections (UTIs) affect millions of women each year resulting in substantial morbidity and lost wages. Approximately 1.5 million women are referred to urology clinics suffering from chronic recurrent UTI on a yearly basis necessitating the use of prophylactic antibiotics. Frequent and recent sexual intercourse correlates with the development of UTI, a phenomenon referred to clinically as “honeymoon cystitis.” Here, using superinfection mouse models, we identified bacterial and host factors that influence the likelihood of developing chronic UTI. We discovered that superinfection leads to a higher rate of chronic UTI, which depended on bacterial replication within bladder cells combined with an immune response including inflammasome activation and cytokine release. These data suggest that bacterial inoculation into an acutely inflamed urinary tract is more likely to lead to severe UTI than bacterial presence in the absence of inflammation. Modification of these risk factors could lead to new therapeutics that prevent the development of recurrent UTI.
| Nearly nine million people present each year to primary care physicians with a urinary tract infection (UTI), costing nearly $2 billion yearly [1], [2]. Women suffer the majority of these infections, with the lifetime risk approaching 50% [3]. Furthermore, 25–40% of these women will suffer recurrent UTI (rUTI), with 1.5 million women referred to urology clinics and often requiring prophylactic antibiotics to prevent recurrence [4]–[6]. Uropathogenic E. coli (UPEC) are responsible for>80% of community acquired UTI and 50% of nosocomial UTI [7], [8]. In the absence of antibiotic therapy, up to 60% of women experience symptoms and/or bacteriuria lasting months after initial infection [9]–[12], implying that cystitis is not always self-limiting. Furthermore, if the infection persists without adequate treatment, the organisms have the capacity to ascend the ureters, causing pyelonephritis and sepsis [13]. Antibiotic resistant organisms further complicate infection and threaten to increase the likelihood of chronic UTI, pyelonephritis and potentially bacteremia [14], [15]. UTIs are increasingly being treated with fluoroquinolones, which in turn has led to a rise in resistance and the spread of multi-drug resistant microorganisms globally, which is a looming worldwide crisis [16], [17]. It is therefore imperative to understand the molecular mechanisms that underlie this problematic disease in order to develop novel therapies.
Sexual intercourse is one of the most significant risk factors predisposing otherwise healthy women to UTI. Early studies demonstrated that sexual intercourse led to a 10-fold increase in bacteria/ml of urine and a subsequently increased predisposition to developing a UTI within 24 hours thereafter [5], [18]–[21]. More recent studies have shown that the frequency with which a woman has sexual intercourse dramatically impacts the likelihood of developing both acute and rUTI [4], [22], [23]. Scholes et. al found a direct association between the number of episodes of sexual intercourse in a given month and the risk of developing rUTI. However the significance of the timing between these episodes of sexual intercourse is unknown. Are evenly spaced episodes associated with an equal risk or, instead, does an episode prime the bladder for rUTI if another insult follows within a sensitive period? To address this question, we developed a model of sequential infection in mice to explore the hypothesis that a sensitive period exists after an initial bacterial insult to the bladder in which the likelihood of developing severe, chronic infection is dramatically increased.
Murine models of UTI have been used to decipher complexities of this disease in naïve individuals. UPEC are capable of colonizing multiple body habitats and niches, including both intracellular and extracellular locations within the bladder, as well as in the gastrointestinal (GI) tract and the kidneys. Selective pressure and bacterial population bottlenecks during colonization impact the ultimate fate of disease [24]–[27]. Adhesive pili assembled by the chaperone/usher pathway (CUP), such as type 1 pili, contain adhesins at their tips that function in adherence and invasion of host tissues and in biofilm formation on medical devices. Upon introduction of UPEC into the bladder, bacteria bind to either mannosylated uroplakin plaques or β1-α3 integrin receptors on the epithelial surface of the bladder via the type 1 pilus FimH adhesin [28]–[30]. Upon internalization, UPEC can be exocytosed as part of a TLR4 dependent innate defense process [31]. In addition to expulsion of individual bacteria, the host can exfoliate superficial facet cells to shed attached and invaded bacteria into the urine for clearance [29]. A small fraction of invaded bacteria escape into the host cell cytoplasm, where they are able to subvert expulsion and innate defenses by replicating into biofilm-like intracellular bacterial communities (IBCs) [24], [32]. UPEC eventually flux out of these communities with a substantial proportion existing as neutrophil resistant filaments [33], [34]. Importantly, evidence of IBCs and bacterial filaments have been observed in women suffering acute UTI, one to two days post self-reported sexual intercourse, but not in healthy controls or infections caused by Gram-positive organisms, which do not form IBCs [21]. IBCs have also been observed in urine from children with an acute UTI [35]. Additionally, IBC formation and the innate immune response of cytokine secretion and exfoliation have been observed in all tested mouse strains, but the long-term outcome of infection differs [36]–[38].
There are two main, mutually exclusive, outcomes to acute infection in C3H/HeN mice: either chronic bacterial cystitis (chronic cystitis), which is characterized by persistent high titer bacteriuria (>104 CFU/ml) and high titer bacterial bladder burdens (>104 CFU) two or more weeks after inoculation, accompanied by chronic inflammation [37], [39], or resolution of bacteriuria [37]. Mice that resolve infection may harbor small populations of dormant UPEC called Quiescent Intracellular Reservoirs (QIRs) [40]. Other mouse strains exhibit varied proportions of these two outcomes. C57BL/6J mice resolve bacteriuria within days and thus are resistant to chronic cystitis, but are susceptible to QIR formation [40], [41]. In contrast, other TLR4-responsive C3H background sub-strains and closely related CBA/J and DBA/2J mice experience persistent high-titer bacteriuria and bladder colonization by UPEC in the presence of chronic inflammation lasting at least four weeks post-infection (wpi). During chronic cystitis, persistent lymphoid aggregates and urothelial hyperplasia with lack of superficial facet cell terminal differentiation accompany luminal bacterial replication [37]. These same histological findings of submucosal lymphoid aggregates and urothelial hyperplasia have been observed in humans suffering persistent bacteriuria and chronic cystitis [42]. Since murine chronic cystitis predisposes to recurrent chronic UTI after antibiotic-mediated bacterial clearance, this is also a relevant model to interrogate the mechanism of recurrent cystitis [37]. In mouse models of UTI, mice initially experience urinary frequency and dysuria as determined by reaction to noxious stimuli and nerve responses during acute infection [43], [44]; however, during chronic cystitis bacterial replication may exist in an asymptomatic carrier state as studies have not been conducted to determine whether dysuria persists. Interestingly, higher serum levels of interleukins (IL) 5 and 6, keratinocyte cytokine (KC/CXCL1), and granulocyte colony-stimulating factor (G-CSF) in C3H/HeN mice at 24 hours post infection (hpi) predicted the development of persistent bacteriuria and chronic cystitis thereafter, suggestive of a host-pathogen checkpoint during acute infection that predicts long term outcome [26], [37]. In women with an acute UTI, increased amounts of serum CXCL1, M-CSF, and IL-8 correlated with subsequent rUTI, suggesting a similar checkpoint [45].
In this manuscript, we developed a superinfection model to mimic the clinical scenario of frequent sexual intercourse whereby sequential inocula are introduced within a brief period of time. C57BL/6J mice are resistant to chronic cystitis when singly infected; however, 30% of C57BL/6J mice developed chronic cystitis when superinfected 24 hours after the initial infection. Serum elevations of IL-6, KC, and G-CSF prior to superinfection predicted the development of persistent bacteriuria in C57BL/6J mice similar to singly infected C3H/HeN mice. Superinfecting C3H/HeN mice 1–6 hours after the initial inoculation increased the proportion of mice experiencing chronic cystitis. In order for this elevation to occur, we found that the initial UPEC inoculum (the “priming” inoculation) must be alive, invasive, capable of intracellular replication, and able to regulate hemolysin expression. Inhibition of the caspase 1/11 inflammasome prior to priming reduced bacterial CFU at four wpi relative to DMSO-treated mice. Microarray analysis of mouse bladders four wpi revealed that both C57BL/6J and C3H/HeN mice secreted antimicrobial peptides and IL-1 during chronic infection. In contrast to C3H/HeN mice, immunoglobulin expression was upregulated in C57BL/6J mice experiencing chronic cystitis. This immunoglobulin expression was absent in C57BL/6J mice that resolved infection and in C3H/HeN mice. Our data suggest mechanisms whereby certain women may be susceptible to rUTI after frequent sexual intercourse dependent on intracellular bacterial replication and the host immune response.
Studies suggest that a host-pathogen checkpoint within the first 24 hpi determines UTI outcome in C3H/HeN mice [26], [37]. In addition, the chronic inflammation observed in mice experiencing chronic cystitis was found to predispose to rUTI after re-infection [37]. Thus, we hypothesized that superinfecting mice during this period of acute inflammation would increase the proportion of mice experiencing chronic cystitis. We transurethrally infected 7–8 week old female C3H/HeN mice with 107 CFU UTI89 or PBS as the priming inoculation and superinfected them 1–2, 6, or 24 hours thereafter. Enumeration of bacterial CFU at one wpi as an initial screen revealed a dramatic increase in the proportion of mice experiencing chronic cystitis in mice superinfected 1–6 hours after priming compared to singly infected or PBS treated mice (Fig. 1A). We used a cutoff of 106 CFU to demarcate mice experiencing high-titer bacterial infection at one week. Importantly, we did not observe a significant increase in CFU when a single inoculum was doubled (2×107 CFU). Superinfection at 24 hpi had no effect on bacterial titers at one week, suggesting that the factors predisposing to increased susceptibility to chronic cystitis upon superinfection wane over time [26]. However, inoculation with PBS followed by UTI89 24 hpi did lead to high titers in 60% of mice. While this result is perplexing, it possibly reflects that sacrifice six days post infection was not sufficient to delineate the typical bimodal distribution of outcomes [37]. The process of catheterization also induces inflammation, which may not have resolved by 6 dpi [46]. We conducted all subsequent C3H/HeN superinfections one hour after priming.
Since early severe inflammatory responses predispose to chronic cystitis [37], we hypothesized that the initial inoculum primed the bladder by initiating an innate immune response to intracellular bacteria that predisposed to a higher proportion of mice experiencing chronic cystitis upon superinfection. We utilized a panel of UTI89 mutants in fimH, ompA, and kps that have been shown to differ in their ability to: i) invade and form IBCs and ii) persist during chronic cystitis in co-infection experiments [47], [48]. Mature IBCs caused by WT bacteria are clonally derived from a single invasive event [24]. The mannose-binding pocket of FimH is invariant among sequenced UPEC [47], and the binding pocket mutant, FimH::Q133K, is defective in mannose-binding and can neither invade the bladder epithelium nor form IBCs. FimH undergoes compact and elongated conformational changes wherein the receptor binding domain bends approximately 37° with respect to the pilin domain. The mannose-binding pocket is deformed in the compact conformation whereas the elongated conformation is mannose binding proficient [49], [50]. Several residues outside the mannose-binding pocket (positions 27, 62, 66 and 163) are under positive selection in clinical UPEC isolates compared to fecal strains [47] and have been shown to function in modulating the conformational changes between the elongated and compact states [48]. FimH::A27V/V163A predominantly adopts a high-mannose binding, elongated conformation. Its expression results in: i) a 10-fold reduction in intracellular CFU one hpi and ii) a defect in the ability to form IBCs at six hpi. FimH::A62S shifts the equilibrium towards the compact conformation. Expression of this allele results in: i) a 10-fold reduction in intracellular CFU one hpi and ii) a 10-fold reduction in IBC formation compared to WT UTI89 [47], [48]. UTI89ΔompA forms half the number of IBCs as UTI89 [51], and UTI89Δkps is defective in IBC formation. UTI89Δkps can replicate intracellularly and the IBC defect can be rescued by co-inoculation with WT UTI89, which results in mixed strain, non-clonal, IBCs [52].
We primed mice with these strains and superinfected one hpi with WT UTI89 and assessed bacteriuria at days 1, 7, 14, and 21 and enumerated bladder titers at 28 dpi. Mice were designated as having chronic cystitis if they had urine bacterial titers greater than 104 CFU/ml at each time point and bladder titers greater than 104 CFU at sacrifice [37]. We found that the FimH::A27V/V163A allele was incapable of priming the bladder for the development of chronic cystitis (p<0.05 relative to WT superinfection). In contrast, FimH::A62S did not significantly differ from PBS or WT superinfection; therefore, it may be capable of priming, though to a lesser degree. UTI89ΔompA and UTI89Δkps were both able to prime the bladder for enhanced chronic cystitis relative to PBS when superinfected one hpi with WT UTI89 (p<0.05 and p<0.01 respectively; Fig. 1C). We also primed with heat-killed UTI89 and found that live, but not heat killed, UTI89 were capable of priming the bladder indicating that bacterial products such as LPS were insufficient (Fig. 1B). These data indicate that live and invasive UTI89 capable of at least some degree of intracellular replication are required for the priming to enhance the incidence of chronic cystitis upon superinfection of UTI89. Taken together these data suggest that priming begins during invasion and early IBC formation.
One of the most potent host defenses to eliminate adherent and invaded UPEC is superficial facet cell exfoliation [29]. The process of exfoliation is activated in part by the bacterial expression of hemolysin (HlyA) [53](Nagamatsu et al. in review). UTI89ΔcpxR overexpresses HlyA, leading to exfoliation and attenuation in our murine model of cystitis (Nagamatsu et al. in review). The UTI89ΔcpxRΔhlyA double mutant was not attenuated, suggesting that the in vivo defect was due to increased hemolysin expression (Nagamatsu et al. in review). The ability of UPEC to rapidly build up in numbers in the form of IBCs and then disperse to neighboring cells may be part of a mechanism to subvert an exfoliation response. Thus, fine-tuning the expression of HlyA during acute bladder infection may serve to maximize UPEC persistence and give UPEC a fitness edge against the host innate inflammatory response. Interestingly, in C3H/HeN mice, UTI89 ΔhlyA is not attenuated throughout infection and causes chronic cystitis comparable to UTI89; however, other reports suggest deletion of HlyA in UPEC CFT073 decreases virulence [54]. We investigated the role of hemolysin in priming the bladder for chronic cystitis upon superinfection by utilizing UTI89ΔhlyA or UTI89ΔcpxR as the initial inoculation followed by WT UTI89 one hpi. Both of these strains were statistically significantly different when compared to WT UTI89 as the priming inoculum. Therefore, we conclude that neither was capable of priming the bladder for enhanced chronic cystitis (Fig. 1D). Thus, too high or low expression of hemolysin abolished the ability of UTI89 to prime for enhanced chronic cystitis implying that an optimal level of hemolysin expression is critical for priming the bladder for enhanced chronic cystitis.
HlyA-mediated exfoliation is in part due to its ability to trigger degradation of paxillin, a scaffold protein that modulates the dynamics of cytoskeletal rearrangements [55]. HlyA can also trigger cell death in human bladder epithelial cells and release of IL-1α via caspase-4 (the murine ortholog is caspase-11) activation and caspase-1-dependent IL-1β secretion via activation of the NLRP3 inflammasome pathway, which orchestrates additional cell death (Nagamatsu et al. in review). We hypothesized that inflammasome and caspase 1/11 activation were essential for superinfection. Thus, mice were treated intravesically with a dose of caspase 1/11 inhibitor or DMSO one hour prior to priming and a second dose with the priming inoculum to test this hypothesis (Fig. 2A). Providing two doses of the inhibitor was previously shown to be effective in dampening in vivo inflammatory responses. In vitro, the inhibitor dramatically reduced downstream elements of inflammasome activation, IL-1α and IL-1β secretion, when bladder cells were infected with UTI89 (Nagamatsu et al. in review). Caspase 1/11 inhibition significantly reduced median bladder titers at four weeks after superinfection relative to the DMSO control group (Fig. 2B). We also saw a trend of caspase 1/11 inhibition in reducing the proportion of WT superinfected mice experiencing chronic cystitis to single infection levels (Fig. 2B). DMSO also reduced the proportion of mice experiencing persistent bacteriuria and chronic cystitis, but to a lesser degree than caspase 1/11 inhibition (Fig. 2B vs. Fig. 1B–D), suggesting an anti-inflammatory role of DMSO alone. Intriguingly, DMSO was recently found to inhibit the NLRP3 inflammasome [56]. Taken together, these data implicate hemolysin and the NLRP3 inflammasome in the priming response to enhanced chronic cystitis.
We further investigated whether chemical exfoliation could enhance the proportion of mice experiencing chronic cystitis prior to a single infection. We utilized the cationic protein, protamine sulfate, which has previously been used to exfoliate the superficial facet cell layer of the urothelium [40], [57]. A 10 mg/mL dose delivered intravesically in 50 µL PBS was shown to exfoliate 65% of the facet cell layer 12 hours after treatment while an additional booster dose of 50 mg/mL led to 95% exfoliation [40]. We utilized these concentrations to initiate, but likely not complete, the process of exfoliation one hour prior to infection with UTI89. We did not observe a significant increase in the proportion of mice experiencing chronic cystitis over PBS pretreatment (Fig. 2C). Thus, these data suggest that at least partial IBC formation in conjunction with caspase 1/11 activation primes the bladder for enhanced chronic cystitis, but chemical initiation of exfoliation is not sufficient. Taken together, these data suggest that exfoliation per se might not play a significant role in impacting the likelihood of enhanced chronic cystitis but instead may reflect a downstream marker of the priming event.
C57BL/6J mice typically rapidly resolve bacteriuria and are resistant to chronic cystitis upon single inoculation with UPEC [37], [38]. Five to ten percent of the time after inoculation with UTI89, C57BL/6J mice experience persistent bacteriuria, but this is generally due to kidney infection without concomitant high titer bladder infection [37], [41]. This degree of kidney infection is not infectious dose dependent and therefore likely due to ureteric reflux of the bacteria during experimental inoculation [37]. We investigated whether superinfecting C57BL/6J mice during acute infection would stimulate an immune response leading to chronic cystitis. We inoculated bladders with PBS or 107 CFU of UTI89 followed by superinfection with UTI89 1, 6, 24, 48 hours or one week after initial infection and collected urine at days 1, 7, 14, and 21 dpi followed by enumeration of bladder and kidney titers at 28 dpi (Fig. 3). A 24 hpi superinfection resulted in 35% of mice sustaining persistent bacteriuria with bladder titers >104 CFU at four weeks compared to 0% in the singly infected group (Fig. 3A). Kidney titers were also increased in the mice with persistent bacteriuria, but we did not observe a significant increase in the proportion of mice with kidney infection greater than 104 CFU (Fig. 3B). These data suggest that at 24 hours after infection the bladders of C57BL/6J mice were primed to develop chronic cystitis upon superinfection. We investigated whether an ascending kidney infection plays a role in predisposing these mice to chronic cystitis by inoculating PBS into the bladder, either 24 hours before or after infection with UTI89, to stimulate a bladder and ureter stretch response or potentially increase reflux of bacteria into the kidneys, respectively. We determined the percentage of mice with persistent bacteriuria and those with bladder and kidney titers greater than 104 CFU at sacrifice (Table 1). We found in all conditions that persistent bacteriuria was a 100% predictor of kidney titers>104 CFU at four wpi. Persistent bacteriuria also predicted bladder titers greater than 104 CFU at four wpi in C57BL/6J mice superinfected 24 hpi with UTI89. For the group of mice inoculated with PBS before the initial UTI89 infection, persistent bacteriuria did not correlate with high bladder titers suggesting these bacteria were only replicating in the kidneys. Serially infecting with two inocula of UTI89 trended towards increased persistent bacteriuria and chronic cystitis compared to the group inoculated with UTI89 followed by PBS at 24 hpi (P = 0.066; Table 1 and Fig. 4A). Kidney titers of UTI89 superinfected mice were significantly higher than when PBS was used to prime or superinfect perhaps suggesting that repeat infection may also increase susceptibility to pyelonephritis (Fig. 4B). Thus, a 24 hpi superinfection of WT UTI89 led to increased rates of persistent bacteriuria and chronic cystitis; however, bladder/ureter stretch or kidney ascension at 24 hpi may contribute to this increase.
C3H/HeN mice that progress to chronic cystitis upon single inoculation can be predicted by elevated serum levels of IL-5, IL-6, KC, and G-CSF at 24 hpi [37]. We hypothesized that similar elevations would predict sensitization to chronic cystitis in C57BL6/J mice if they were subsequently superinfected. Thus, we determined levels of 23 serum cytokines from C57BL/6J mice 24 hrs after initial inoculation with PBS or UTI89 prior to superinfection. We then superinfected a subset of the mice initially infected with UTI89 (superinfection in Fig. 5) leaving the other mice untouched (UTI89 group). All mice were evaluated with urine titers over 28 d and sacrificed to enumerate bladder titers. We stratified the superinfected mice based on outcome four weeks later as determined by persistent bacteriuria and chronic cystitis. We found elevations of serum KC (Fig. 5A), IL-6 (Fig. 5B), and G-CSF (Fig. 5C) in mice that progressed to chronic cystitis relative to those that resolved infection or were mock-infected with PBS. Therefore, higher levels of these cytokines correlate with chronic cystitis that develops later if mice are superinfected. At the time we obtained serum, the single infection and superinfection groups were identical, and no statistical differences existed among them. These data demonstrate that a subset of C57BL/6J mice respond to an initial infection in a way that results in higher specific serum cytokine levels and primes them to develop chronic cystitis if an additional insult is delivered 24 hpi.
During chronic cystitis of singly-infected C3H/HeN mice, the bladder epithelium is hyperplastic and normal terminal differentiation of the superficial facet cell layer, including the expression of surface uroplakins, does not occur [37]. In this environment, the bacteria are able to persist extracellularly by an unknown mechanism. To assess this, we conducted scanning electron microscopy analysis on bladder tissue harvested at four wpi and found that bacteria replicate in the presence of ongoing epithelial exfoliation and neutrophil influx in chronic cystitis of both C3H/HeN and C57BL/6J mice (S1A–D Fig.). This analysis supports previous experiments that have shown that during chronic cystitis the majority of bacteria are extracellular, replicating in the urine or adherent to underlying transitional epithelial cells [24], [37]. The mechanism by which bacteria adhere in the absence of uroplakins has not been demonstrated in vivo, but in vitro studies have shown that FimH binds integrins and other host proteins such as TLR4 [30], [58], [59]. Alternatively additional adhesive factors such as other CUP pili may play a role. Interestingly, during chronic cystitis, neutrophils, which we observed to be actively engulfing bacteria, are insufficient for clearing infection; however, the reason for this is unclear. Mature superficial facet cells could not be discerned at this time point, but were present in mock-infected mice (S1E Fig.). Patients with persistent bacteriuria or rUTI have been reported to have similar histopathology [42]. In order to identify the bladder micro-environment in which UPEC replicate during chronic cystitis, we conducted microarray analysis on RNA extracted from bladders four wpi. C3H/HeN mice were singly-infected and C57BL/6J mice were superinfected to develop chronic cystitis. Mice from each strain inoculated with PBS were used as controls. Depicted in Fig. 6 are the expression profiles relative to the global average with green indicating increased expression and red denoting decreased. C3H/HeN mice experiencing chronic cystitis had a dramatically different expression profile from resolved and mock-infected mice (Fig. 6A). Uroplakins were among the most downregulated genes during chronic cystitis in both mouse models, consistent with the lack of terminally differentiated superficial facet cells (S1 Fig.). Eleven of the 20 (55%) most upregulated genes during chronic cystitis were the same in both mouse strains (S1 Table). The functional categorization revealed that most of the up-regulated genes function in inflammatory response, cytokine release, and ion binding [60]–[62]. Of interest among these genes in both of these mouse models is the inflammasome-related cytokines IL-1. We have shown that UPEC activate the caspase 4 murine homologue, caspase 11, during acute infection in a hemolysin-dependent fashion (Nagamatsu et. al. in review). Despite these similarities, interesting differences existed in the ongoing inflammatory response in mice experiencing chronic cystitis (S1 Table). In C57BL/6J mice, the inflammatory response is immunoglobulin- and cytokine-mediated whereas in C3H/HeN mice, we noted a remarkable absence of upregulated immunoglobulin genes. The increased expression of antimicrobial peptides such as RegIIIγ and the calgranulins (s100a8 and s100a9) is interesting because this increased expression is not sufficient to eliminate bacterial replication during chronic cystitis. Interestingly, C3H/HeN mice that were mock infected exhibited a very similar profile to mice that resolve infection (Fig. 6A). Contrary to C3H/HeN mice, C57BL/6J mice that resolved infection differed significantly from either chronic cystitis or mock infected mice, suggesting an element of altered physiology and immunological memory of the infection (Fig. 6B). This information supports research that serially infecting mice that resolve infection makes them less susceptible to recurrent infection [37], [63]. What is interesting here is that the mechanisms by which this occurs may differ between mouse strains, and possibly by extension, women.
We have developed models of bacterial superinfection of the urinary tract, which may provide insight into the connection between recent and frequent sexual intercourse and the susceptibility to the development of chronic UTI [5], [22]. Our results demonstrate that superinfection resulted in increased susceptibility to chronic cystitis in both susceptible and resistant mouse genetic backgrounds, but the time window for priming differed between strains. We have previously shown that chronic cystitis predisposes to severe rUTI upon a subsequent infection weeks to months after clearance of the first infection with antibiotics [37]. Clinically, millions of women take post-coital and prophylactic antibiotics so as not to develop rUTI [64]. Therefore, if clinically applicable, our results detailed here may partially explain why frequent sexual intercourse is such a strong risk factor for UTI. The necessity of prophylactic antibiotics could be obviated if the risk factors and bacterial traits identified here can be altered in the clinical population of women suffering chronic rUTIs.
Frequent sexual intercourse is among the most important risk factors for rUTI in young women [22]. Peri-urethral carriage of the causal strain and sexual intercourse immediately precede the development of a rUTI [5]. Sexual intercourse likely introduces mixed populations of bacteria into the urinary tract, with E. coli being the most common [18]. In this environment, UPEC invade bladder tissue and replicate, forming IBCs and bacterial filaments, which have been observed in human urine in 40% of patients suffering acute UTI, 24–48 hours after reported sexual intercourse [21]. These data may provide mechanistic insight as to the frequent clinical observation that recent and frequent sexual intercourse over a brief period of time leads to increased rates of rUTI [23]. Furthermore, elevated levels of serum CSF1, CXCL-1, and CXCL-8 in women with acute UTI were associated with a higher rate of rUTI [45]. Using C3H/HeN and C57BL/6J mice, we have shown that superinfection during the period of acute infection dramatically increases the proportion of mice that experience chronic cystitis with inoculations of 107 UPEC (Fig. 1A and 3A). The bacterial characteristics responsible for frequent recurrences are beginning to be assessed [65]. Hemolysin is expressed by 50% of UPEC isolates, but is more likely to be associated with symptomatic UTI [66]. It is possible that hemolysin-mediated exfoliation and caspase 1/11 activation leads to UTI-associated symptoms. In our studies, we found that an increase in priming for chronic cystitis correlated with the bacterial ability to invade and replicate within the bladder tissue (Fig. 1B–C), and through hemolysin to activate caspase 1/11 leading to IL-1 secretion and bacterial replication (Fig. 1D and 2B). Activation of caspase 1/11 has been shown to contribute to epithelial cell death in vitro and exfoliation in vivo in C3H/HeN mice, suggesting that caspase-mediated exfoliation may expose the underlying epithelium upon which UPEC replicates during chronic cystitis (Nagamatsu et. al. in review). Inhibition of caspase 1/11 protected superinfected mice from chronic cystitis (Fig. 2), suggesting a role for cytokines downstream of caspase activation including IL-1α and IL-1β, identified in our microarray of four-week bladders (Fig. 6; S1 Table). A microarray analysis revealed that in C3H/HeN and C57BL/6J mice, 11/20 of the most upregulated genes during chronic cystitis were the same. Differences between the responses to infection in these mouse strains may result from the dramatic increase in kidney infection or QIR presence in C57BL/6J relative to C3H/HeN mice [37], [40]. Further, this data supports the hypothesis that a muted inflammatory response to UPEC infection is more likely to lead to resolution [26]. Also, our studies suggest that serum biomarkers such as IL-6, KC, and G-CSF may predict a predisposition to rUTI (Fig. 5) [37]. Recently, it was demonstrated that cytokines involved in immune cell chemotaxis and maturation (the human homolog of KC included) during acute UTI enhanced the likelihood of developing rUTI [45].
We have created mouse models that have identified both bacterial and host immune factors that may predispose women to rUTI. Inhibiting caspase-mediated inflammation or downstream effectors may serve to prevent a UTI from becoming a chronic or recurrent UTI. Further work to identify bacterial and host factors that influence the balance between resolution and chronic infection is required to lead to better treatments clinically. The ability of UPEC to invade bladder tissue allows it to transcend stringent bottlenecks during infection [24], [25], [27]. The ability to replicate intracellularly also impacts the ability of a second invading strain to proliferate in the bladder environment (Fig. 1B–C). The molecular basis of bacterial colonization of the bladder during chronic cystitis is an area of active investigation. Previously, it has been shown that mannosides are effective in treating chronic cystitis arguing that FimH-mediated binding plays an important role [67]. It has recently been demonstrated that FimH variation outside of the binding pocket affects protein conformation and pathogenicity of UPEC [48]. This variation may impact bacterial adherence and replication during chronic cystitis. Furthermore, because invasion and intracellular replication appear to influence the likelihood to develop chronic cystitis, treatments with soluble compounds such as mannosides that block the ability of UPEC to invade the tissue or compounds that might alter FimH conformation hold promise as effective means to prevent or treat rUTI [67]–[70]. These analyses may allow us to identify high-risk patients for more aggressive therapy and/or anti-virulence compounds to limit this troubling disease.
All WT bacterial strains utilized were derivatives of UTI89, including tagged, isogenic UTI89 isolates, kanamycin resistant UTI89 attHK022::COM-GFP, kanamycin resistant UTI89 with re-integrated UTI89 FimH, spectinomycin resistant UTI89 attλ::PSSH10-1, and chloramphenicol resistant UTI89 [24], [47], [71]. FimH mutant strains, ΔompA, Δkps, ΔhlyA, ΔcpxR were all previously published [47], [51], [52](Nagamatsu et al. in review).
Bacteria for infection were prepared as previously described [72]. Six to seven week old female C3H/HeN (Harlan) or C57BL/6J (Jackson) were transurethrally infected with a 50 µL suspension containing 5×106–2×107 CFU of UTI89 or relevant mutant in PBS under 3% isofluorane. Protamine Sulfate (Sigma) was dissolved in PBS and caspase 1/11 inhibitor Ac-YVAD-CMK (BACHEM) was dissolved in DMSO and transurethrally inoculated into the bladder. At indicated timepoints after infection, mice were anesthetized and infected again. Venous blood was obtained at 24 hpi, just prior to re-infection, by submandibular puncture and centrifuged at max speed at 4°C in Microtainer serum separation tubes (BD) and stored at −20°C until use. Cytokine expression was measured using the Bio-Plex multiplex cytokine Group I bead kit array (Bio-Rad), which measures 23 cytokines. Urine was obtained by gentle suprapubic pressure and serially diluted and plated on appropriate antibiotic plates. Mice were sacrificed by cervical dislocation under isofluorane anesthesia, and their organs were aseptically removed. Chronic cystitis was determined if animals had urine titers>104 CFU/mL at 1, 7, 14, 21 dpi and bladder titers>104 CFU at sacrifice at 28 dpi [37]. Animals that resolved infection and had a recurrence or had resolved the infection with reservoir titers>104 CFU were marked in red and considered to have resolved the chronic infection. Organ titers shown are the total bacterial burden.
The Washington University Animal Studies Committee approved all mouse infections and procedures as part of protocol number 20120216, which was approved 01/11/2013 and expires 01/11/2016. Overall care of the animals was consistent with The guide for the Care and Use of Laboratory Animals from the National Research Council and the USDA Animal Care Resource Guide. Euthanasia procedures are consistent with the “AVMA guidelines for the Euthanasia of Animals 2013 edition.”
C3H/HeN or C57BL/6J mice were infected as discussed above. After 28 days, animals that had developed chronic cystitis, resolved the infection, or aged matched PBS controls were sacrificed for RNA isolation. Upon sacrifice, 5 bladders from each condition were immediately pooled and homogenized in Trizol for RNA isolation according to the manufacture's suggested protocol. DNase treatment was performed to remove any contaminating DNA before submission to the Genome Technology Access Center for sample processing and hybridization on Affymetrix Mouse Gene 1.0 chips in triplicate. Data was analyzed using the Partek Genomics Suite. Gene lists were compiled using fdr-ANOVA analysis with a significance cut off of p<0.001. Experiments were repeated twice with a representative analysis shown. Microarray data are available in the ArrayExpress database (www.ebi.ac.uk/arrayexpress) under accession number E-MTAB-2930.
Mice were infected as described above. Bladders were aseptically harvested, bisected, and splayed. Bladders were fixed in 2.0% glutaraldehyde in 0.1M sodium phosphate buffer overnight. Bladders were then washed three times with 0.1M sodium phosphate buffer and de-ionized water before being fixed in 1.0% osmium tetroxide. Bladders were washed and then critical point drying was performed with absolute ethanol and liquid carbon dioxide. Sputter coating was performed with gold-palladium using a Tousimis Samsputter-2a. Images were obtained on a Hitachi S-2600H operated at 20 kV accelerating voltage.
Datapoints below the limit of detection (LOD) were set to the LOD for graphical representation and statistical analysis. For cytokine data, values out of the range of the instrument were not included for analysis. Fisher's exact test was utilized to determine differences between groups for rates of chronic cystitis. One-way ANOVA was utilized to determine whether any cytokine differences were apparent and pairwise assessment of median values was determined by Mann-Whitney test. Unless otherwise indicated, p<0.05 was considered significant. Analyses were performed in Graphpad Prism 5.0.
|
10.1371/journal.ppat.1003152 | Cyclosporine A Impairs Nucleotide Binding Oligomerization Domain (Nod1)-Mediated Innate Antibacterial Renal Defenses in Mice and Human Transplant Recipients | Acute pyelonephritis (APN), which is mainly caused by uropathogenic Escherichia coli (UPEC), is the most common bacterial complication in renal transplant recipients receiving immunosuppressive treatment. However, it remains unclear how immunosuppressive drugs, such as the calcineurin inhibitor cyclosporine A (CsA), decrease renal resistance to UPEC. Here, we investigated the effects of CsA in host defense against UPEC in an experimental model of APN. We show that CsA-treated mice exhibit impaired production of the chemoattractant chemokines CXCL2 and CXCL1, decreased intrarenal recruitment of neutrophils, and greater susceptibility to UPEC than vehicle-treated mice. Strikingly, renal expression of Toll-like receptor 4 (Tlr4) and nucleotide-binding oligomerization domain 1 (Nod1), neutrophil migration capacity, and phagocytic killing of E. coli were significantly reduced in CsA-treated mice. CsA inhibited lipopolysaccharide (LPS)-induced, Tlr4-mediated production of CXCL2 by epithelial collecting duct cells. In addition, CsA markedly inhibited Nod1 expression in neutrophils, macrophages, and renal dendritic cells. CsA, acting through inhibition of the nuclear factor of activated T-cells (NFATs), also markedly downregulated Nod1 in neutrophils and macrophages. Silencing the NFATc1 isoform mRNA, similar to CsA, downregulated Nod1 expression in macrophages, and administration of the 11R-VIVIT peptide inhibitor of NFATs to mice also reduced neutrophil bacterial phagocytosis and renal resistance to UPEC. Conversely, synthetic Nod1 stimulating agonists given to CsA-treated mice significantly increased renal resistance to UPEC. Renal transplant recipients receiving CsA exhibited similar decrease in NOD1 expression and neutrophil phagocytosis of E. coli. The findings suggest that such mechanism of NFATc1-dependent inhibition of Nod1-mediated innate immune response together with the decrease in Tlr4-mediated production of chemoattractant chemokines caused by CsA may contribute to sensitizing kidney grafts to APN.
| Patients who have received a kidney graft are treated with immunosuppressive drugs, such as the cyclosporine A (CsA). Transplanted patients under CsA are prone to bacterial infections. In this study, we used an experimental mouse model of kidney infection with Escherichia coli (E. coli) bacteria to study the effect of CsA. We show that CsA treatment of mice reduced their renal defense against E. coli. We found that CsA, in addition to its inhibitory action on the TLR4-mediated production of chemoattractant chemokines, also inhibited the expression of nucleotide-binding oligomerization domain 1 (Nod1), an intracellular receptor involved in the innate immune response against bacteria, in phagocytic cells. CsA acts by inhibiting the functions of the transcription factor NFAT. We show that NFAT is required for the proper expression of Nod1. Since Nod1 has already been reported to be involved in the phagocytic functions of polymorphonuclear neutrophils, we looked for and found a severe defect in neutrophil bacterial killing associated with reduced expression of Nod1 in both mice and patients treated by CsA. Interestingly, when mice treated with CsA are given synthetic molecules known to bind Nod1, this permitted the restoration of the Nod1 expression and renal defenses. This paper describes a novel mechanism which may explain, at least in part, why transplant patients receiving CsA have increased susceptibility to bacterial infection, and also provides a potential therapeutic strategy to restore renal antibacterial defenses.
| Urinary tract infection (UTI) often complicated by acute pyelonephritis (APN), which is mainly caused by uropathogenic Escherichia coli (UPEC), is the most frequent infectious complication following renal transplantation [1], [2]. Despite improvement of the surgical procedures and the use of post-operative antibiotic prophylaxis, the rates of post-graft APN still remain higher than in the normal population [2], and late UTI occurring after more than 6 months after transplantation are associated with increased risk of death, and post-graft APN may also compromise long-term graft outcome [3], [4]. Although many factors including age, sex, and co-morbidity conditions play a role in the susceptibility to infection, long-term immunosuppressive therapy used to prevent episodes of acute graft rejection obviously increases the risk of bacterial, viral or fungal infections in the context of transplantation [5], [6].
Calcineurin inhibitors, such as Cyclosporine A (CsA), are almost incontrovertible drugs widely used to prevent renal graft rejection. Their main function is to inhibit the phosphatase activity of calcineurin, which regulates the nuclear translocation of the nuclear factor of activated T-cells (NFATs) transcription factor [7]. Impaired activation of NFATs then prevents the transcription of cytokine genes, including IL-2, in activated T cells [8]. However, the mechanism(s) by which CsA could alter the innate immune system, and thereby decrease renal host defenses against ascending UPEC remain largely unknown.
Early recognition of bacterial motifs by a number of pattern recognition receptors, including Toll-like receptors (TLRs) and (Nod)-like receptors (NLRs), is essential for the removal of bacterial pathogens [9]. UPEC colonizing the urinary tract are recognized by several TLRs, including TLR2, 4, 5, and 11 [10]. Studies using experimental murine models of ascending UTI have demonstrated that Tlr4, which senses lipopolysaccharide (LPS) from Gram-negative bacteria [10], and also Tlr11, that is expressed in murine bladder epithelial cells and RTECs [11], regulate susceptibility to UTIs in mice. TLRs play key roles in activating the transcription factor NF-κB and the mitogen-activated protein kinases (MAPKs) signaling pathways leading to the production of chemoattractant cytokines and subsequent recruitment of neutrophils and monocytes/macrophages for efficient clearance of the bacteria. Nod1 and Nod2 also promote the activation of NF-κB and MAPKs through the recruitment of the kinase RIP-2 (also known as RIP2K or RICK), which is a member of the caspase activation and recruitment domain (CARD) protein family [12], [13]. Nod1 recognizes muramyl tripeptide (M-TriDAP), a degradation product of peptidoglycan (PGN) containing DAP which is present in most Gram-negative bacteria and some Gram-positive bacteria [14], [15], while Nod2 recognizes muramyl dipeptide (MDP), a motif common to PGNs from all classes of bacteria [15]. Nod2 is mainly expressed in monocytes and macrophages, and mutations of NOD2 are associated with Crohn's disease, an inflammatory bowel disease mainly driven by T cells [16], [17]. The functions of Nod1, which is more ubiquitously expressed, differ somewhat from those of Nod2. Recent studies have demonstrated that Nod1 plays a key role in the migration of neutrophils into the intestine and liver [18], and in activating phagocytic mechanisms of bacterial killing [19], [20]. The fact that altered leukocyte functions and decreased capacity for bacterial phagocytosis are the most common abnormalities in the immune status of renal transplant recipients [21], [22], led us to investigate the possibility that CsA alters the Nod1-mediated neutrophil functions and bacterial phagocytic killing of UPEC.
In the present study, we used an experimental mouse model of ascending UTI and show that the administration of CsA to wild-type (WT) mice decreases renal resistance to UPEC infection. CsA impaired Tlr4-mediated activation and subsequent production of chemoattractant chemokines in the epithelial collecting duct cells, to which UPEC bind preferentially during their retrograde ascent along the urinary tract system [23]. In addition, CsA, through its inhibitory action on NFATs, also markedly inhibited the functional expression of Nod1 in phagocytic cells, including neutrophil migration capacity and phagocytic killing of UPEC. Similar to CsA-treated WT mice, Nod1−/− mice exhibited greater susceptibility to UPEC than their WT counterparts. Using 11R-VIVIT, a synthetic peptide inhibitor of NFATs [24], we also demonstrate in vitro and in vivo the relevance of the regulatory role of the NFATc1 isoform in controlling the Nod1-mediated renal susceptibility to UPEC. We also report the functional downexpression of human NOD1 and decreased phagocytic capacity of E. coli in leukocytes from renal transplant recipients treated with CsA. The combined inhibitory effects of CsA on Tlr4-mediated chemokine production and Nod1-mediated migration of neutrophils and bacterial phagocytic capacities, which contribute to decrease renal antibacterial defenses in mice, may explain, at least in part, the susceptibility of CsA-treated renal transplant recipients to bacterial infections.
WT mice treated with CsA (15 mg/kg) or its vehicle for 5 days were then infected by transurethral inoculation with the UPEC HT7 strain isolated from the urine of a woman with acute pyelonephritis [4], [25], to test whether CsA affected renal antibacterial defenses. 24 h after the inoculation of live UPEC, the bacterial burden and E. coli positive immunostaining were greater in CsA-treated mice than vehicle-treated mice (Figure 1A and B). As a control, we checked that CsA did not modify the growth rate of UPEC (not shown), thus excluding any direct effect of the calcineurin inhibitor on the bacteria. CsA also increased renal bacterial loads in kidneys from Rag2−/− mice, which lack mature lymphocytes, to almost the same extent as in WT kidneys (Figure 1A), suggesting that CsA promotes UPEC infection independently of its inhibitory effect on the adaptive immune system. The amount of secreted chemokines MIP-2/CXCL2 and KC/CXCL1, which play key roles in the chemoattraction of neutrophils during experimental UTI [23], [26]–[28], was also significantly lower in kidneys from CsA-treated mice than vehicle-treated mice (Figure 1C). Fewer Ly-6G+ neutrophils were detected in the kidney medulla and in the urinary space in the CsA-treated mice than in untreated mice (Figure 1D). Quantification of neutrophils assessed by measuring myeloperoxidase (MPO) activity also revealed significant lower MPO activity in the 24 h post-infected kidney homogenates from CsA-treated mice than vehicle-treated mice (Figure 1E).
Flow cytometry (FACS) analysis revealed that the CD45+ leukocyte population detected in the 24 h post-infected kidneys was essentially composed of F4/80+ CD11bLO Gr1−/LO MHC-II+ CD11c+ renal dendritic cells (DCs), which have been shown to form a contiguous network in the renal tubulointerstitium [29], F4/80− CD11b+ Gr1HI MHC-II− CD11c− neutrophils, and to a lesser extent F4/80+ CD11b+ Gr1INT MHC-II− CD11c− inflammatory monocytes/macrophages. FACS analysis revealed a significant decrease in the proportion of neutrophils over the total CD45+ renal cell population detected in the 24 h post-infected kidneys from CsA-treated mice compared to vehicle-treated mice kidneys (Figure S1 and Figure 2A and B). In contrast, CsA only slightly, and non-significantly reduced the number of monocytes/macrophages or DCs present in the 24 h post-infected kidneys (Figure S1). This suggests that CsA preferentially impairs the migration capacity of neutrophils in the UPEC-infected kidneys. In vitro experiments using the Boyden chamber method also revealed that the migration capacity of neutrophils isolated from the blood of the 5-day CsA-treated mice and stimulated by the neutrophil activating agent N-formyl-methionyl leucyl-phenylalanine (fMLP) or by CXCL2 was significantly decreased compared to that of neutrophils from vehicle-treated WT mice (Figure S2A and B). CsA also altered the bacterial phagocytic killing capacity by neutrophils. In contrast to neutrophil-enriched peritoneal cells (NPCs) isolated from vehicle-treated mice, NPCs collected from CsA-treated mice exhibited significantly lower ex vivo capacity to internalize Texas red-coupled E. coli and kill serum-opsonized E. coli (Figure 2C and D).
Nod1−/− neutrophils were shown to exhibit deficient capacity of bacterial phagocytic killing and lower migration capacity than WT neutrophils [20], [30]. Given that neutrophil migration and their phagocytic killing capacities were markedly reduced in CsA-treated mice, we tested whether CsA directly alters intrarenal expression of Nod1. Quantitative real-time PCR revealed that the levels of Nod1 mRNA expression and, to a lesser extent those of Tlr4, but not of Nod2, were markedly decreased in the 24 h post-infected CsA-treated mice kidneys compared to those of the infected vehicle-treated mice (Figure 2E). In accordance with these findings, the amount of the immunodetected Nod2 protein remained equivalent in the infected kidneys from CsA-and vehicle-treated mice, whereas the amounts of Nod1 and Tlr4 proteins were ∼50% and ∼30%, respectively, lower in the infected kidneys from CsA-treated mice than in those from their vehicle-treated counterparts (Figure 2F). Despite the decrease in Nod1 expression, the level of phosphorylated over total RIP-2, which is involved in the control of Nod-mediated NF-κB activation [12], was similar in the 24 h post-infected kidneys from CsA-treated mice and vehicle-treated mice (not shown). Collectively, these findings suggest that, in addition to an inhibitory effect on Tlr4 mRNA and protein expression, CsA impairs the recruitment and functions of neutrophils in the inflamed kidneys by a mechanism possibly linked to downregulation of Nod1 expression.
We next investigated the consequence of Nod1 deficiency in host renal bacterial defenses. The renal bacterial burden and the number of immunodetected UPEC were significantly greater in the kidneys of Nod1−/− mice than in those of WT, 24 h after the inoculation of UPEC (Figure 3A and B). Less Ly-6G+ neutrophils were detected in the urinary space from post-infected Nod1−/− mice than from post-infected WT or Nod2−/− mice (Figure 3C), and, like in UPEC-infected CsA-treated mice, the MPO activity also remained significantly lower in the post-infected Nod1−/− than in post-infected WT or Nod2−/− kidneys (Figure 3D). FACS also showed that the proportion of polymorphonuclear neutrophils infiltrating the 24 h post-infected Nod1−/− mice kidneys was lower than in the WT kidneys (Figure 3E and F). In accordance with the findings of Dharancy et al. [30], in vitro experiments using a Boyden chamber revealed that the migration capacity of neutrophils isolated from naive Nod1−/− mice and activated by fMLP (or with CXCL2, not shown) was significantly lower than that of WT neutrophils (Figure S2C and D). These findings also suggested that Nod1 is implicated in the migration capacity of neutrophils. Consistently with a role for Nod1 in the bacterial phagocytic killing capacity of neutrophils [20] NPCs collected from Nod1−/− mice, but not those isolated from WT or Nod2−/− mice, were unable to internalize Texas red-coupled E. coli (Figure S3A and B). A significant reduction in the killing of serum-opsonized UPEC was also observed in Nod1−/− neutrophils compared to WT or Nod2−/− neutrophils (Figure S3C). We then analyzed the effects of CsA on the renal susceptibility of Nod1−/− mice to UPEC. Administration of CsA slightly, but not significantly, increased the renal bacterial loads in kidneys from Nod1−/− mice compared to those from untreated Nod1−/− mice (Figure 3G), further suggesting that CsA impairs Nod1-mediated antibacterial defenses to UPEC.
Previous studies had demonstrated that bladder epithelial cells and renal epithelial tubule cells are actively involved, together with bone marrow-derived cells, in the chemoattraction of neutrophils to the site of inflammation in experimental models of ascending UTI [31], [32]. We also showed that UPEC preferentially binds to the apical side of epithelial cells constituting the terminal collecting duct (Figure 4A) [33]. Activation of TLR4 signaling in the urinary tract system infected by UPEC plays a key role in this process [23], [34]. Because renal tubule cells also express the Nod1 and Nod2 receptors, that can be activated in inflamed kidneys [35], [36], experiments were carried out to analyze the effects of CsA on both Tlr4, and Nod1 or Nod2 signaling in renal tubule cells and bone marrow-derived cells activated by LPS, Nod ligands or UPEC. We checked that the renal medullary collecting duct (MCD) cells did express Tlr4, Nod1, and Nod2 mRNAs (Figure 4B). LPS, and to a much lesser extent the Nod1 agonist FK156 and the Nod2 agonist MDP, stimulate the production of CXCL2 in cultured WT MCD cells (Figure 4C). CsA inhibited in a dose-dependent manner the Tlr4 mRNA and protein expressions without altering Nod1 and Nod2 expressions (Figure 4D and E). We then tested whether 100 nM (corresponding to 120 ng/ml) CsA alters the TLR4- and/or Nod-mediated cellular response in MCD cells. The CXCL2 production caused by UPEC significantly decreased in untreated Tlr4−/− MCD cells compared to WT, Nod1−/−, or Nod2−/− MCD cells, and also decreased to almost the same extent in CsA-treated WT, Nod1−/−, or Nod2−/− MCD cells than in CsA-treated Tlr4−/− MCD cells (Figure 4F, upper panel). As controls, similar profiles of CXCL2 production were obtained by incubating WT, Nod1−/−, and Nod2−/− MCD cells with LPS, and no significant production of CXCL2 was detected in Tlr4−/− MCD cells (Figure 4F, lower panel). These findings thus suggest that CsA mainly affects the predominant TLR4-mediated production of CXCL2 and has only a minor effect on epithelial Nod1- and Nod2-mediated renal tubule cell activation caused by UPEC. CsA also significantly reduced the ability of LPS-activated confluent WT MCD cells, which developed high electrical transepithelial resistance (∼4500 Ω. cm2), to stimulate the in vitro migration capacity of neutrophils as compared to untreated WT MCD cells incubated with LPS (Figure S4A to C). As a control, Tlr4−/− MCD cells challenged with LPS did not stimulate the migration of neutrophils (Figure S4B and C).
We next analyzed the effects of CsA on Tlr4 and Nod mRNAs expression in neutrophils, macrophages, and renal DCs. Incubating primary bone marrow neutrophils with CsA for 8 h or bone marrow macrophages (BMMs) with CsA for longer times (48 h) significantly decreased the relative levels of Nod1 mRNA and protein expressions, and to a lesser extent reduced the expression of Nod2 mRNA, but not that of Tlr4 mRNA (Figure 5A, B, D and E). Renal DCs expressing Nod1 and Nod2 [37], were shown to produce substantial amounts of CXCL2, which is involved in the recruitment of neutrophils in the kidneys following UPEC challenge [26], [27], [38]. Incubating highly-enriched CD11c+ cells isolated from WT kidneys by gradient centrifugation and magnetic beads separation [39] with 100 nM CsA for 48 h significantly reduced Nod1 mRNA expression without affecting the expression of Tlr4 or Nod2 (Figure 5G). However, the small number of purified renal CD11c+ cells obtained did not permit reliable Western blot analysis of the Nod1 protein. Consistent with an inhibitory action of CsA on Nod1, the production of CCL5, which has been shown to be highly sensitive to Nod agonist stimulation [40], when stimulated by the Nod1 agonist FK156 was significantly lower in CsA-treated than in untreated neutrophils and macrophages (Figure 5C and F). CsA only slightly, and non-significantly, reduced the FK156-stimulated production of CCL5 in renal DCs (Figure 5H). Given that CXCL2 plays an essential role in the recruitment of neutrophils, we went on to test the effects of CsA on CXCL2 production in neutrophils, macrophages, and renal DCs. Unlike renal MCD cells, CsA only slightly reduced the CXCL2 production stimulated by LPS in neutrophils, macrophages, and renal DCs (Figure 5C, F and H). Taken as a whole, these findings indicate that CsA globally impairs the functional expression of Nod1 in neutrophils, macrophages, and renal DCs.
Since Nod1 senses a number of invasive Gram-negative bacteria, we tested whether UPEC can directly activate Nod1 mRNA expression in macrophages and whether CsA impairs the UPEC-induced activation of Nod1. Incubating WT BMMs with UPEC for 3 h had almost no stimulatory effect on Tlr4 mRNA expression, but in contrast induced a significant increase in Nod1 and Nod2 mRNAs expression (Figure S5). Pre-incubating BMMs with CsA impairs the increase in Nod1 mRNA expression, and to a much lesser extent that of Nod2, caused by the subsequent incubation with UPEC for additional 3 h (Figure S5). These findings further suggest that CsA preferentially alters the activation of Nod1 induced by UPEC in phagocytic cells.
CsA inhibits the nuclear translocation of NFATs, which in turn inhibit the transcription of T cell effector cytokines [8]. Because CsA preferentially alters Nod1 expression in phagocytic cells, experiments were carried out to test whether the downregulation of the NFATc1 isoform, which is highly expressed in both murine and human neutrophils and macrophages [41], [42], impairs mRNA expression of Nods. Because neutrophils have a limited life-span, experiments were carried out on mouse BMMs. Knockdown of NFATc1 mRNA expression using a multiple set of NFATc1 siRNAs (referred to as NFATc1a–d siRNA) in WT BMMs resulted in the almost complete inhibition of the expression of NFATc1 mRNA when compared to non-transfected WT BMMs or cells transfected with a control siRNA (Figure 6A). Silencing NFATc1 by the set of NFATc1 siRNAs markedly inhibited the relative level of Nod1 mRNA expression, but had almost no effect on Nod2 mRNA expression (Figure 6B). To further assess the inhibitory action of CsA on Nod mRNAs expression, experiments were performed on WT BMMs incubated with 11R-VIVIT, a cell-permeable peptide that specifically inhibits the calcineurin-NFATs interaction without affecting calcineurin phosphatase activity [43]. Incubating WT BMMs with 1 µM 11R-VIVIT for 48 h also markedly inhibited the relative levels of Nod1 mRNA expression, and, to a lesser extent, that of Nod2 mRNA (Figure 6C). In contrast, knock-down of NFATc1 mRNA expression or incubation of BMMs with the 11R-VIVIT had no effect on Tlr4 mRNA expression (Figure S6A and B). Given that inhibition of NFATc1 can affect the transcriptional expression of many proteins, the possibility that a contrario in vitro activation of NFATs could specifically stimulate the expression of Nod1 was investigated. NFATs are activated by increased intracellular calcium concentration during T-cell activation [7]. Calcium mobilization induces the dephosphorylation of cytosolic NFATs which translocate into the nucleus [44]. Ionomycin (2 µM, 60 min) induced the translocation of NFATc1 from the cytosol into nuclei from WT BMMs, whereas the pre-incubation of WT BMMs with CsA or 11R-VIVIT totally or almost totally impaired the nuclear translocation of NFATc1 caused by subsequent addition of ionomycin (Figure 6D). Ionomycin also significantly stimulated Nod1, but failed to stimulate Nod2 and Tlr4 mRNAs expression in WT BMMs compared to untreated cells, or to cells pre-treated with CsA or 11R-VIVIT (Figure 6E and Figure S6C). Collectively, these data strongly suggest a role for NFATc1 as a transcriptional activator of Nod1.
Given that the 11R-VIVIT inhibited Nod1 mRNA expression, 11R-VIVIT should also impair the Nod1-mediated bacterial phagocytic function and decrease the renal defense against UPEC. NPCs isolated from WT mice given daily intraperitoneal injections of 10 µg/kg 11R-VIVIT for 48 h exhibited a significant lower ex vivo capacity to internalize E. coli and lower phagocytic killing of serum-opsonized UPEC than untreated NPCs (Figure 7A and B). 24 h after the inoculation of UPEC, 11R-VIVIT-treated mice exhibited reduced intrarenal MPO activity and reduced amount of immunodetected Nod1 protein, and significant greater renal bacterial burden than in non-treated WT mice (Figure 7C to E).
The fact that the stimulation of Nod1 can enhance systemic innate immunity [20] and that the administration of Nod1 peptide agonists to mice confer resistance against several pathogens [45], led us to test whether the stimulation of Nod1 by synthetic Nod1-stimulating agonists could reinforce renal defense against UPEC. The cell-permeable Nod1 activating agonist C12-iEDAP (50 µg/ml for 24 h) induced a significant increase in Nod1 mRNA expression, which overcome the inhibition of Nod1 mRNA expression caused by CsA alone (not shown). We then tested whether the in vivo administration of synthetic Nod1 agonists can reactivate Nod1-mediated phagocytic function and reinforce renal resistance of CsA-treated mice to UPEC. The capacity of neutrophils to internalize Texas red-coupled E. coli was greater in neutrophils from CsA-treated mice that had been treated with C12-iEDAP than in those collected from CsA-treated mice which had not received C12-iEDAP (Figure 8A). Furthermore, intra-peritoneal injection of C12-iEDAP or FK156 (80 µg/mouse) one day before the transurethral inoculation of UPEC to CsA-treated WT mice, induced significant reduction in the renal bacterial burden when compared to CsA-treated mice which had not received C12-iEDAP or FK156, or CsA-treated mice which had received the Nod2 agonist MDP (Figure 8B). The observed decrease in renal bacterial burden was associated with greater MPO activity in the infected kidneys from CsA-treated mice pre-treated with the Nod1 agonists (Figure 8C). Figure 8D illustrates the greater amount of Nod1 protein in the infected kidneys from CsA-treated mice which had been pre-treated with FK156. The amounts of CXCL2 and CXCL1 secreted were not significantly different in the 24-h, post-infected kidneys from CsA-treated mice and the vehicle-treated mice (Figure 8E), suggesting that the administration of Nod1 agonists does not induce any major renal inflammatory response. Collectively, these findings are consistent with the restimulation of Nod1-mediated host protective functions in CsA-treated mice.
Investigations were performed on blood samples from random renal transplant recipients treated with CsA (n = 25) to test whether human renal transplant recipients exhibit similar decrease in NOD1 expression and defective NOD1-mediated bacterial phagocytosis capacity. The demographic characteristics of renal transplant recipients are summarized in Table S1. Transplant recipients all received CsA and additional immunosuppressive drugs, including prednisolone and mycophenolate mofetil, which is a selective inhibitor of the de novo synthesis of guanosine nucleotides in T and B lymphocytes [46]. For comparison, investigations were also performed on healthy volunteers (n = 10) used as controls. The production of IL-8 was first measured in whole blood samples incubated with various TLR and NOD agonists. The levels of IL-8 triggered by all TLR agonists tested did not significantly differ in blood samples from transplant recipients and normal healthy controls (Figure 9A). In contrast, the levels of IL-8 production stimulated by the human NOD1 synthetic agonist M-TriDAP was significantly less in blood samples from the transplant recipients treated with CsA than in normal controls (Figure 9A and B). The levels of NOD1 mRNA, but not those of NOD2, TLR2, or TLR4 mRNAs extracted from whole blood samples, were also significantly lower in transplant recipients than in normal controls (Figure 9C). Flow cytometry analysis of phagocytosis of Texas red-coupled E. coli and quantification of NOD1 in human neutrophils also revealed that the low levels of NOD1 in the neutrophils from CsA-treated renal transplant recipients were closely correlated with their capacity to phagocyte E. coli, which was significantly lower than in the neutrophils from normal healthy controls (Figure 9D). Moreover, three of the transplant recipients analyzed who exhibited low NOD1 expression and low capacity of E. coli phagocytosis by neutrophils had a previous history of UTI/APN. These findings suggest that the NFATs-dependent inhibitory mechanism of Nod1-mediated innate immune response identified in the mouse also occurs in human transplant recipients treated with CsA.
In the present work, we show that in mice, CsA reduces renal resistance to the retrograde inoculation of uropathogenic E. coli. CsA induces a significant decrease in the production of the chemoattractant chemokines and impairs the recruitment of neutrophils in kidneys from mice infected by UPEC. The primary source (i.e. the epithelial tubule cells or circulating immune cells) of pro-inflammatory mediators produced in experimental models of UTI remains discussed. In accordance with a number of previous studies, medullary collecting duct epithelial cells, which are the first renal tubule cells to come into contact with UPEC during their retrograde ascent, produce substantial amounts of TLR4-mediated CXCL2, which play a key role in the recruitment of neutrophils in the infected kidneys [23], [27], [47]. The fact that CsA impairs LPS-induced production of CXCL2 in renal MCD cells and LPS-induced recruitment of neutrophils, further suggests that the Tlr4-mediated activation of renal epithelial cells contributes to the recruitment of neutrophils in the infected kidneys, at least during the initial phase of infection. Nor can we exclude the possibility that the decrease in Tlr4 mRNA expression detected in the infected kidneys and in the Tlr4-mediated cell activation detected in murine MCD cells, are due, at least in part, to the cytotoxic action of calcineurin inhibitors [48]. In contrast, CsA had no in vitro inhibitory effect on Tlr4 expression in neutrophils, macrophages, or renal DCs. After being bound to renal collecting duct cells, UPEC induces the rapid recruitment of neutrophils (during the first 6 h), followed by the recruitment of monocytes/macrophages over the next 12–24 h. Although LPS stimulates the in vitro production of CXCL2 by neutrophils and inflammatory monocytes/macrophages, these cells do not seem to play major roles in the renal production of chemoattractant chemokines during UPEC infection [38]. Recently, resident renal DCs have been shown to be major source of CXCL2 production, compared to neutrophils and monocytes/macrophages, 20 h following the retrograde inoculation of UPEC [38]. Furthermore, the migration capacity of neutrophils has been reported to be significantly lower in UPEC-infected CD11c deficient mice than in their WT counterparts [38], indicating that CXCL2 production by renal DCs certainly plays some role in the chemoattraction of neutrophils. Here we show that CsA altered Nod1 mRNA expression in renal DCs, without impairing Tlr4 mRNA expression and the number of renal DCs in the infected kidneys. Although in vitro incubation of renal DCs with CsA did not have much effect on the in vitro LPS-induced CXCL2 production, we cannot exclude any in vivo participation of renal DCs in the defective migration capacity of neutrophils within infected kidneys from CsA-treated mice.
The present study demonstrated an unexpected effect of CsA on Nod1-mediated neutrophils migration capacity and bacterial phagocytosis. We show both in vivo and in vitro that CsA impairs Nod1 expression in neutrophils and macrophages. The stimulation of Nod1 by Nod1 stimulating agonist or bacteria was shown to play a role in the recruitment of neutrophils in the intestine [18], [49], and that the number of infiltrating neutrophils was shown to be significantly reduced in injured livers from Nod1−/− mice challenged with carbon tetrachloride [30]. We also detected defective recruitment of neutrophils in kidneys from Nod1−/− mice infected by UPEC, and in infected kidneys of WT mice treated with CsA. Given that CsA, which affects TLR4-mediated CXCL2 production in MCD cells and alters the expression of Nod1 in neutrophils, macrophages, and also renal DCs, these findings suggest that, in addition to impairing the epithelial TLR4-mediated production of chemoattractant chemokines, CsA may also alter the Nod1-mediated capacity of neutrophils to migrate in kidneys infected with UPEC.
Recent studies have highlighted the role of NFAT/calcineurin signaling pathways in controlling innate immunity and in regulating homeostasis of immune cells. Calcineurin/NFATs signaling was shown to negatively regulate myeloid lineage development [50]. The susceptibility to fungal infection caused by CsA was also shown to be the consequence of NFAT-dependent inhibition of an immune innate pathway regulating antifungal resistance in neutrophils. Indeed, Greenblatt et al. [42] reported that the neutrophils of both calcineurin-deficient mice and CsA-treated mice exhibited a defective ability to kill Candida albicans without any noticeable changes in the classical fungicidal activity of neutrophils. These authors showed that calcineurin regulates the ability of neutrophils to kill C. albicans via another anti-microbial pathway, which involves the C-type lectin-like receptor dectin-1 and IL-10 production. Given that Nod1 and Nod2 are not directly involved in the recognition of C. albicans [51], these findings suggest that CsA may affect NFATs-dependent cellular signaling activated by Gram-negative bacteria or fungi in different ways. The NFATc3/c4 isoforms were also shown to be required for TLR-induced innate inflammatory response in monocytes/macrophages [52]. Our results strongly suggest that NFATc1 controls Nod1 at the transcriptional level. In silico analysis (Genomatix Software GmbH) has identified putative NFAT binding sites in human and murine Nod1 and Nod2 promoter regions. Downexpression of NFATc1 inhibited Nod1 more efficiently than Nod2. However, the in silico analysis did not permit us to predict differences in the number of putative binding sites for NFATc1 on the promoter regions of Nod1 and/or Nod2. We have no direct explanation for the preferential inhibitory effect of silencing NFATc1 on Nod1 expression. The fact that much less Nod1 than Nod2 is present in immune cells may account, at least in part, for the greater reduction in Nod1 mRNA expression induced by NFATc1 silencing. Nevertheless our findings strongly suggest that CsA, through NFATc1 inhibition, alters Nod1-mediated phagocytic functions. In agreement with this, inhibition of the calcineurin phosphatase activity has been reported to decrease phagocytosis in macrophages [53], further suggesting that NFATc1 is essential for proper activation of the phagocytosis process. Collectively, these findings indicate that NFATs, which play key roles in adaptive T cell functions, are critical cellular mediators of the innate immune responses.
Although the present findings indicate that CsA directly alters Nod1 expression, it may also affect other immune receptors and downstream signaling pathways in various different ways. Calcineurin serine threonine phosphatase downregulates TLR-mediated signaling pathways in macrophages, whereas CsA and its newer counterpart Tacrolimus have been shown to activate NF-κB and induce cytokine expression in inactivated macrophages [54]. It has been also reported that the activation of DCs and macrophages by Tacrolimus can induce a state of reduced responsiveness to LPS [55], similar to the LPS-induced transient state of tolerance observed following a subsequent LPS challenge [56]. During bacterial infection, it seems likely that various TLRs and NLRs are activated in response to the invading pathogen or to microbial components, such as LPS or PGN, released into the bloodstream and in the infected tissues [57]. Moreover, the interplay between Nods and TLRs may be critical for the induction of protective immune responses [14], [58], [59]. Therefore, it is conceivable that altered epithelial TLR4-mediated chemokine production caused by CsA, may potentiate the deleterious inhibitory effects of CsA on Nod1-mediated neutrophil functions, leading to a more pronounced decrease in host resistance to bacterial infection.
Long-term use of immunosuppressive drugs, used to prevent graft rejection, increases the susceptibility of transplant recipients towards bacterial infection [2], [5], [6]. Until recently, the impact of immunosuppressive therapy was considered to be largely non-specific. However, several groups of researchers have reported changes in the numbers and/or functions of circulating leukocytes, including polymorphonuclear neutrophils from transplant patients acquiring infections [21], [22]. Moreover, it has already been suggested that abnormalities in neutrophil functions, including impaired migration capacity following fMLP stimulation, are indicators of sepsis in solid organ transplant recipients [60]. Neutrophils from renal transplant recipients have been reported to exhibit diminished phagocytic activity and reduce bactericidal activity against Klebsiella pneumoniae, compared to the activities seen with neutrophils from healthy subjects [61]. In vitro studies have also shown that CsA reduces both neutrophil phagocytosis capacity and ROS production [62], [63]. Analysis of a panel of blood leukocyte phenotypes and functions also revealed that transplant recipients (renal and renal/pancreas) most of whom were receiving CsA and subjected to infection exhibited a reduction in ROS production [21]. In the present study, the investigations performed on renal transplant recipients have revealed downregulated expression of NOD1 in circulating leukocytes, similar to that found in CsA-treated mice. The low capacity for E. coli phagocytosis appears to be closely correlated with a low expression of NOD1 in the neutrophils of renal transplant recipients. Since Nod1−/− mice are more susceptible than WT to early Streptococcus pneumoniae sepsis, and conversely, that PGN recognition by Nod1 enhances killing of S. pneumoniae and Staphylococcus aureus by neutrophils [19], it is conceivable that the observed downregulation of NOD1 caused by CsA may also impair the capacity of neutrophils from renal transplant recipients to kill Gram-positive bacteria. However, we cannot exclude the possibility that the results from investigations performed in human renal transplant recipients could have been flawed by several confounding factors, such as the concomitant administration of several immunosuppressive drugs and some degree of renal impairment. The consistency with which CsA downregulated NOD1 strongly suggests that the impairment of NOD1-mediated bacterial phagocytic capacity caused by CsA may therefore represent an additional risk factor for the occurrence of UTI/APN in human transplant recipients. Despite antibiotic prophylaxis, the frequency of post-graft UTI/APN still remains relatively high, and increasing the resistance of bacteria to antibiotics may also increase the risk of recurrent episodes of UTI/APN. This raises the question of whether alternative therapeutic strategies could help to reduce the frequency of post-graft UTI/APN. A number of studies have provided convincing evidences that pre-treatment of mice with Nod agonists enhances host protection against sepsis, bacterial infection, viruses, or even parasites [64]. Here we also show that administration of Nod1 agonists can restore efficient renal clearance of UPEC in CsA-treated mice. However, further studies will be required to find out whether the administration of synthetic Nod1 agonists alone or in combination with antibiotics could potentially help to reduce the occurrence of UTI/APN in renal grafts.
In summary, we have identified a hitherto-unknown mechanism of the NFATc1-dependent inhibitory action of the Nod1-mediated innate immune response, which may affect host renal antibacterial defenses against invasive uropathogens, and possibly favor the emergence of bacterial infection in renal transplant recipients receiving long-term CsA treatment.
All animal experiments were approved by and conducted in accordance with guidelines of the French Agricultural Office and in compliance with the French and European regulations on Animal Welfare (Service de la protection et Santé animale; Approval Number 75–687, revised 2008) and with Public Health Service recommendations. All the efforts were made to minimize suffering of mice.
Blood samples were obtained from transplant recipients and healthy volunteers after being informed and given oral consent, according to French law for non interventional studies using a leftover or a small additional blood sample (Public Health Code, article L1121-1, revised in May 2009). All samples were anonymized. Human and animal studies were approved by the Institutional Ethics Committee (Comité de Protection des Personnes (CCP #5) affiliated to the Tenon Hospital (AP-HP)-University Paris 6 (Approval CCP-0612/2011). All experiments were conducted in accordance with the principles expressed in the Declaration of Helsinski.
Adult female (8–10 week old) WT mice (supplied by the Centre d'Elevage Janvier, Le Genest-Saint-Isle, France), Rag2−/− mice, and Nod1−/− and Nod2−/− mice from the same C57BL/6 genetic background were used. Mice were infected with the uropathogenic E. coli strain HT7 (108 bacteria in 50 µl sterile PBS) introduced via the transurethral route into the bladder as described [4], [25]. 100 µl CsA (Neoral, Novartis International Pharmaceutical Ltd, 15 mg/kg), or its vehicle (castor oil) were administered sub-cutaneously to mice for 5 days before the inoculation of UPEC. Bacterial loads (CFU) in kidneys were determined 24 h after infection by plating. Kidney sections were stained using anti-E. coli antibody (Interchim), anti-Ly6-G antibody (BD Biosciences), or aquaporin-2 (AQP-2) as described [23].
Primary cultures of medullary collecting duct (MCD) isolated from WT, Nod1−/−, Nod2−/−, and Tlr4−/− mice kidneys were grown as described [23]. Experiments were carried out on confluent cells two weeks after seeding.
Bone marrow neutrophils and circulating blood neutrophils were isolated by gradient density centrifugation using Ficoll-Paque PREMIUM (GE Healthcare) as described elsewhere [20]. Bone marrow-derived macrophages (BMMs) were isolated and grown as described [40]. Indirect immunofluorescence studies were performed on WT BMMs using a mouse anti-NFATc1 monoclonal antibody (Thermo Scientific Pierce Antibodies) and Sytox green nucleic acid stain (Invitrogen). Renal dendritic CD11c+ cells were isolated as previously described with slight modifications [39], [65]. For each cell preparation both kidneys from 5 naïve WT mice were used. Briefly, the kidneys of each mouse were minced and then digested for 45 min at 37°C with 1 mg/ml collagenase (Roche Diagnostics, Meylan, France) and 10 µg/ml DNAse I in RPMI 1640-Glutamax medium (Life Technologies) supplemented with 10% heat-inactivated fetal calf serum, 10 mM HEPES, 100 U penicillin, and 0.1 mg/ml streptomycin. Kidney homogenates from each mouse were then filtered through 70 µm nylon mesh, washed with PBS, centrifuged (250 g, 5 min), resuspended in 3 ml of 0.01 M ethylenediaminetetracetic acid (EDTA) in FCS and layered on top of 3 ml Histopaque-1077 (Sigma). Density centrifugation (400 g, 30 min) was performed at room temperature. The interphase cells were then harvested, washed, and resuspended in 600 µl MACS buffer (Miltenyi Biotec.). CD11c+ cells were then isolated using microbead-labeled specific monoclonal antibody (clone N418, Miltenyi Biotec.), and separated using magnetic beads according to the manufacturer's instructions. The enriched- CD11c+ cell suspension obtained from 5 mice were then pooled and used for the cytokine assay.
The migration capacity of neutrophils isolated from untreated or CsA-treated WT mice or Nod1−/− mice was analyzed using the Boyden chamber technique as previously described [30]. After the lysis of red blood cells, blood samples from vehicle- and CsA-treated WT were laid on the top of a Ficoll-Paque PREMIUM (GE Healthcare, Uppsala, Sweden) density gradient, then centrifuged (400 g, 30 min at 4°C), and the bottom layer containing the neutrophil-enriched fraction was collected. 106 neutrophils were then resuspended in 200 µl Hank's buffered salt solution (HBSS) containing 0.5% bovine serum albumin and added to the upper compartment of a Transwell Clear membrane insert (3 µm pore size, Corning Inc., Lowel, MA). The lower compartment (600 µl) of the chamber contained either HBSS alone or supplemented with fMLP (10−7 M) or CXCL2 (200 ng/ml). Incubations were performed at 37°C for 40 min in a 5% CO2/95% air atmosphere. A neutrophil migration assay was also carried out using isolated WT and Tlr4−/− MCDs seeded and grown to confluence in defined DMEM/Hams'F12 culture medium [23] on the apical side of the filters. Confluent WT MCD cells were then incubated with or without 100 ng/ml CsA for 48 h, then with or without (10 ng/ml) LPS, which was added when required to the upper compartment of the chamber for 4 h in a 5% CO2/95% air atmosphere. The lower compartment contained 106 WT neutrophils resuspended in 600 µl defined culture medium. In all cases, the filters were rinsed, then fixed in methanol and stained using the RAL 555 Kit (Réactifs RAL, Martillac, France). The neutrophils (stained deep purple) detected in the filters were counted by microscopic observation. MCD cells (stained pale red) were also stained using the RAL kit containing eosin. In parallel, the transepithelial electrical resistance (RT) was measured using dual silver/silver chloride (Ag/AgCl) electrodes connected to a Millicel-ERS voltohmmeter (Millipore, Billerica, MA).
Enriched-neutrophil peritoneal cells collected by peritoneal lavages 3 h after a single intraperitoneal injection of 1.5 ml of thioglycollate (Bio-Rad Laboratories) were incubated with Texas red-coupled E. coli (104 bacteria/107 cells) for 30 min at 37°C, and then stained with CD11b-FITC or Wheat Germ Agglutinin (WGA)-Alexa Fluor 647 (Invitrogen) to delineate cell peripheries. The internalization of E. coli was determined by measuring the intracellular red fluorescence intensity using confocal microscopy analysis as described elsewhere [66]. For the ex vivo bacterial killing assay, E. coli were mixed without or with peritoneal neutrophils (103 bacteria/106 neutrophils) following the same procedure as described elsewhere [20].
Blood samples from 25 renal transplant recipients with a functioning graft during the first three years after surgery and exposed to CsA (Table S1) were randomly taken during the regular routine consultations at Tenon hospital (Assistance Publique-Hôpitaux de Paris, France). In all cases, the blood samples were taken at least 6 months after surgery. In addition, ten healthy volunteers served as normal controls.
Total RNA from mouse kidneys, neutrophils, or macrophages was purified with RNAble (Eurobio laboratories) and reverse transcribed using Moloney Murine Leukemia Virus reverse transcriptase (Invitrogen). cDNA was subjected to quantitative real-time PCR using a Chromo IV sequence detector (MJ Research). The mouse Tlr2, Tlr4, Nod1, Nod2 and ß-actin primers used and the corresponding Taqman probes are listed in Table S2. PCR data were reported as the relative increase in mRNA transcripts versus that found in uninfected kidneys or vehicle-treated neutrophils or macrophages cells and corrected using the respective levels of ß-actin mRNA. Quantitative real-time PCR was also performed on RNA extracted from blood samples of renal transplant recipients using human TLR2, TLR4, NOD1, NOD2, and ß-ACTIN primers and corresponding Taqman probes (listed in Table S2). PCR data were reported as the relative increase in mRNA transcripts versus that found in a pool of RNA of untreated leukocytes from healthy volunteers. For reverse transcription PCR, cDNA and non-reverse transcribed RNA (250 ng) from cultured mouse MCD cells or BMMs were amplified for 35 cycles in 35 µl of Platinum Blue PCR SuperMix (Invitrogen) containing 10 pmol of mouse NFATc1, Tlr4, Nod1, Nod2, or GAPDH primers (described in Table S2). Amplification products were run on a 2% agarose gel containing SYBR Safe DNA gel stain (Invitrogen) and photographed.
Experiments were performed using different predesigned HP GenomeWide (Qiagen, Courtaboeuf, France) siRNAs (referred to as NFATc1a, b, c, and d) for the murine NFATc1 gene target DNA sequence. NFATc1a DNA sequence: 5′-TCGGATCGAGGTGCAGCCCAA-3′; sense: 5′-GGAUCGAGGUGCAGCCCAATT-3′; antisense: 5′-UUGGGCUGCACCUCGAUCCGA-3′; NFATc1b DNA sequence: 5′-CACGGTTACTTGGAGAATGAA-3′; sense: 5′-CGGUUACUUGGAGAAUGAATT-3′; antisense: 5′-UUCAUUCUCCAAGUAACCGTG-3′; NFATc1c DNA sequence: 5′-CCCGTCCAAGTCAGTTTCTAT-3′; sense: 5′-CGUCCAAGUCAGUUUCUAUTT-3′; antisense: 5′-AUAGAAACUGACUUGGACGGG-3′; NFATc1d DNA sequence: 5′-CCGGGACCTGTGCAAGCCAAA-3′; sense: 5′-GGGACCUGUGCAAGCCAAATT-3′; antisense: 5′-UUUGGCUUGCACAGGUCCCGG-3′. A universal negative control siRNA (target DNA sequence: 5′-AATTCTCCGAACGTGTCACGT-3′; sense: 5′-UUCUCCGAACGUGUCACGUdTdT-3′; antisense: 5′ ACGUGACACGUUCGGAGAAdTdT-3′) was also used (Qiagen). Single strand sense and antisense RNA nucleotides were annealed to generate a RNA duplex according to the Manufacturer's instructions. WT BMMs were seeded in 6-well plates and incubated with 10 nM of each siRNA tested and 2 µl of Lipofectamine RNAiMAX Reagent (Invitrogen) for 48 h at 37°C before use. As a control, we checked that each of the NFATc1 (a to d) siRNAs inhibited NFATc1 mRNA expression in macrophages using reverse-transcription PCR (not shown).
Mouse kidney homogenates and BMMs were lysed and processed for Western blotting using mouse anti-TLR4 [25]), anti-Nod1 (Cell Signaling) or anti-Nod2 (eBiosciences) antibodies, and phospho-RIP-2 (Ser 176), and total RIP-2 (Ozyme) antibodies. Protein bands were revealed using horse raddish peroxidase-conjugated goat anti-rabbit IgG (Jackson Immunoresearch), and analyzed by Western Blotting.
Cytokine production was measured in mouse kidney homogenates, or cell supernatants using DuoSet mouse ELISA kits (R&D Systems, Minneapolis, MN). Neutrophils, macrophages, or cultured MCD cells were incubated either with LPS (Escherichia coli 0111:B4 LPS Ultra-Pure, InvivoGen, Toulouse, France), 50 µg/ml C12-iEDAP (InvivoGen), 1 µM FK156 (provided by Nami Kawano, Astellas Pharma Inc., Osaka, Japan), or 1 µM MDP (InvivoGen, Toulouse, France) for 8 to 18 h at 37°C. For FK156 and MDP stimulations, mouse macrophages were pre-treated with 1 µM cytochalasin D (Calbiochem, Darmstadt, Germany) for 30 min to allow efficient internalization of the synthetic Nod activating agonists as described elsewhere [40]. Human blood samples (10 µl) were incubated in 500 µl RPMI culture medium (Invitrogen) at 37°C alone or with 1 ng/ml Pam3CSK4 (InvivoGen) or LPS, 1 µg/ml flagellin (InvivoGen); 50 µg/ml unmethylated CpG-DNA (HyCult Biotechnology, Uden, The Netherlands), 50 nM MDP or various concentrations (0.05–2 µM) of M-TriDAP (InvivoGen) for 18 h. IL-8 production was measured using a DuoSet human ELISA kit (R&D Systems, Lille, France). All the reagents used were tested negative for endotoxin contamination using the Limulus amoebocyte assay according to the Manufacturer's recommendations (QCL-1000, Biowhittaker, Buckinghamshire, UK). MPO activity was measured using HyCult Biotechnology ELISA kit.
The cell populations infiltrating the infected mouse kidneys were analyzed by flow cytometry. 24 h after UPEC infection, kidneys were carefully rinsed with PBS to remove the remaining circulating blood cells. The kidneys were then minced and digested for 45 min at 37°C with 1 mg/ml collagenase (Roche Diagnostics) and 10 µg/ml DNAse I in the same RPMI 1640-Glutamax medium (Life Technologies) as described above for the isolation of renal DCs. After rinsing, kidney homogenates were then passed through a 70 µm pore sized nylon Cell Strainer (BD Biosciences) with 15 ml PBS. The resulting cell suspension was centrifuged (1600 rpm, 10 min) again and then resuspended (10×106 total cells/ml) in FACS buffer containing 2% BSA and 0.05% sodium azide. Non-specific binding of antibody to Fc receptors was blocked by incubating the cell suspension with the anti-mouse CD16/32 (2.4G2) antibody (10 µg/ml) and ChromePure rat IgG (100 µg/ml, Jackson Immunoresearch) for 30 min at 4°C. Cells were then incubated in pre-determined optimal concentrations of fluorochrome-conjugated antibodies to cell surface antigens or matching isotype control antibodies for 30 min at 4°C. APC anti-mouse Ly-6G/Ly-6C (Gr1; RB6-8C5), Pe anti-mouse F4/80 (BM8), PerCP/Cy5.5 anti-mouse CD11b (M1/70), Pacific Blue anti-mouse (MHC-II/IA/IE, M5/114.15.2) and PeCy7 anti-mouse CD45 (RA3-6B2), and matching fluorophores-conjugated antibodies isotypes were purchased from Biolegend. PeCy7 anti-mouse CD11c (HL3) and V500 anti mouse CD45.2 (104) were purchased from BD Pharmingen. Fluorescent measurements were conducted with identical settings on at least 100,000 CD45+ cells per kidney per experiment using a BD FACSCanto II cytometer operating BD FACSDiva software v6.1.3 (BD Biosciences, Erembodegem, Belgium), and FlowJo v7.6.5 (Tree Star. Inc., Ashland, OR). Analyses of NOD1 expression and bacterial phagocytosis capacity by human neutrophils were also analyzed by flow cytometry. Human whole blood samples (1 ml) were incubated with Texas red-coupled E. coli (107 bacteria) (Molecular Probes) with gentle stirring for 30 min at 37°C, while negative control samples were kept on ice before analysis. Red blood cells were then lyzed by adding 10 ml NH4Cl 0.8% wt/vol for 15 min. Trypan blue (0.05 mg/ml) was added to the samples to reduce the quenching of surface-bound fluorescence. Samples were centrifuged (400 g, 10 min) at 4°C to remove cell debris, and pelleted leukocytes were then rinsed in PBS. In parallel, aliquots of NH4Cl-treated blood samples were permeabilized with methanol, and then incubated with an anti-human NOD1 antibody (Imgenex Corp.). All fluorescence measurements were conducted with identical settings and forward and side-scatter parameters to identify the neutrophil population and to gate out other cells and debris [67].
Statistical analysis was performed using the GraphPad Prism program. The unpaired t test, (two-tailed p values) was used to compare two groups. The distribution of three or more groups was analyzed by One-Way ANOVA and the Kruskal-Wallis test. The Mann-Whitney test was used to compare the group with one another. A p value<0.05 was considered significant.
The mouse gene accession numbers (GenBank) are as follows: ß-actin, NM_007393.3; GAPDH, AK144690; NFATc1/NFAT2, NM_016791.4; Nod1, NM_172729; Nod2, NM_145857; RipK2, NM_138952.3; Tlr2, NM_011905.3; Tlr4, NM_021297, Tlr5, AF186107.1; Tlr9, AF314224.
The human gene accession numbers (GenBank) are as follows: GAPDH, NM_002046.3; NOD1, NM_006092.2; NOD2, NM_022162.1; TLR2, NM_003264.3; TLR4, NM_138554.3.
The mouse protein accession number (UniProtKB/Swiss-Prot) are as follows: ß-actin, P60710; CCL5, P30882; CD11b/integrin alpha-M/beta-2; P05555; CXCL1; P12850; CXCL2, P10889; NFATc1/NFAT2, 088942; Nod1, Q8BHB0; Nod2, Q8K3Z0; RipK2, P58801; Tlr2, Q9QUN7; Tlr4, Q9QUK6, Tlr5, Q9JLF7; Tlr9, Q9EQU3.
The human protein accession numbers (UniProtKB/Swiss-Prot) are as follows: IL-8, P10145; NOD1, Q9Y239.
|
10.1371/journal.ppat.1006440 | Streptococcus gallolyticus subsp. gallolyticus promotes colorectal tumor development | Streptococcus gallolyticus subsp. gallolyticus (Sg) has long been known to have a strong association with colorectal cancer (CRC). This knowledge has important clinical implications, and yet little is known about the role of Sg in the development of CRC. Here we demonstrate that Sg promotes human colon cancer cell proliferation in a manner that depends on cell context, bacterial growth phase and direct contact between bacteria and colon cancer cells. In addition, we observed increased level of β-catenin, c-Myc and PCNA in colon cancer cells following incubation with Sg. Knockdown or inhibition of β-catenin abolished the effect of Sg. Furthermore, mice administered with Sg had significantly more tumors, higher tumor burden and dysplasia grade, and increased cell proliferation and β-catenin staining in colonic crypts compared to mice receiving control bacteria. Finally, we showed that Sg is present in the majority of CRC patients and is preferentially associated with tumor compared to normal tissues obtained from CRC patients. These results taken together establish for the first time a tumor-promoting role of Sg that involves specific bacterial and host factors and have important clinical implications.
| Colorectal cancer (CRC) is a leading cause of cancer-related death. The recognition that microbial agents can contribute to the development of CRC raises hope for improving CRC diagnosis and treatment by incorporating both microbial and patient characteristics into clinical strategies. S. gallolyticus subsp. gallolyticus (Sg) has been implicated in CRC for decades. Patients with Sg infections display a much higher risk of having CRC compared to the general population. Despite this, the precise role of Sg in the development of CRC—i.e., whether this organism plays an active role in the development of tumor or its presence is merely a consequence of the tumor environment being favorable for its colonization of the colon—was unknown. Here using in vitro cell cultures and mouse models of CRC, we demonstrate that Sg actively promotes colon cancer cell proliferation and tumor growth, suggesting that it is not an innocent “passenger”. These results represent a major step forward in understanding the relationship between Sg and CRC. This combined with the prevalence of Sg in CRC patients highlight Sg as being both clinically relevant and functionally important for CRC.
| Colorectal cancer (CRC) is the second to third most common cancer and a leading cause of cancer-related death [1,2]. The human gut is normally exposed to ~ 1014 microorganisms, which are increasingly recognized as being intimately involved in the health and disease status of the gut [3]. The identification of specific microbes driving colon tumorigenesis [4–9] raises hope that we may exploit knowledge about specific tumor-promoting microbes to improve cancer diagnosis, prevention and treatment [10–15].
Streptococcus gallolyticus subsp. gallolyticus (Sg) belongs to the Streptococcus bovis/Streptococcus equinus complex (SBSEC) and was previously known as S. bovis biotype I [16,17]. This is a Gram-positive, opportunistic pathogen that causes bacteremia and endocarditis in humans. Patients with bacteremia/endocarditis due to Sg display a strong association with CRC as shown in numerous case reports and case series [3,12–15,18–29]. A meta-analysis of case reports and case series published between 1970 and 2010 found that patients with S. bovis infections had ~60% chance of having concomitant colon adenomas/carcinomas [30], much higher than that in the general population. The SBSEC group includes a number of closely related species such as Sg, S. pasteurianus, S. macedonicus and S. infantarius [16,31]. Among these different species, Sg infection has the strongest association with CRC (~ 7 fold higher risk compared to infections caused by the other species), suggesting the existence of a Sg-specific mechanism(s) that promotes the association between the pathogen and CRC. Other similar studies consistently revealed elevated risks of colorectal adenoma/carcinoma for patients with Sg infections [14,21–25,32]. These epidemiological studies clearly show a strong association between Sg bacteremia/endocarditis and CRC. Due to the overwhelming clinical association, it is recommended that patients with Sg infections undergo thorough colon examination. In addition, recent prospective studies followed patients with endocarditis due to Sg or other pathogens for several years and found that a significantly higher percentage of Sg endocarditis patients developed new colonic neoplasm during the follow-up period, compared to patients with endocarditis caused by other pathogens [33,34]. This suggests that Sg plays a role in the early stages of tumor development. On the other hand, CRC patients may be colonized with Sg in their colon with no signs of bacteremia or endocarditis [35,36], although the prevalence of Sg in CRC patients is not as well defined as the risk for CRC in patients with active Sg infections.
Despite a strong clinical association between Sg and CRC, the role of Sg in CRC development, i.e. whether it drives colon tumorigenesis or merely colonizes the colon tumor environment, was unknown. In this study, we investigated the effect of Sg on colon tumor development using in vitro cultured human colon cancer cells and mouse models of CRC. We demonstrated that Sg promoted colon cancer cell proliferation in a manner that depends on the specific cell context, bacterial growth phase, and direct contact between bacteria and colon cancer cells. In addition, we showed that Sg promoted cell proliferation by up-regulating β-catenin, a central signaling molecule in colon tumorigenesis. Exposure to Sg also resulted in larger tumors in a xenograft model. In an azoxymethane (AOM)-induced mouse model of CRC, mice given Sg had significantly higher tumor burden, higher average dysplasia grade, increased cell proliferation and β-catenin level in colon crypts compared to control mice. Lastly, we analyzed tumors and adjacent normal tissues from CRC patients. The results indicate that Sg is present in the majority of CRC patients and preferentially associates with tumor tissues. These results demonstrate a tumor-promoting role of Sg and have important implications with respect to microbial contributions to CRC as well as clinical practices to combat CRC.
The overall effect of Sg on cell growth and proliferation was examined using a variety of cell lines. Human colon cancer cell lines HCT116, HT29, LoVo, SW480, SW1116, normal human colon epithelial cell lines FHC and CCD 841 CoN, human kidney epithelial cell HEK293 and human lung cancer cell line A549 were co-cultured with Sg strains TX20005 and TX20030, and Lactococcus lactis MG1363 (used as a negative control bacterial strain). The number of viable cells was counted after 24 and 48 hours of incubation. We found that, in the presence of the Sg strains, HCT116, HT29 and LoVo had significantly more viable cells than colon cancer cells cultured in the presence of L. lactis or no bacteria (~ 50–60% more at 24 hours and ~ 20–30% more at 48 hours) (Fig 1A–1C). Interestingly, we did not observe any increase in cell numbers for the other cell lines tested including SW480, SW1116, A549, HEK293, CCD841 CoN, and FHC (Fig 1D–1I). These results suggest that Sg strains TX20005 and TX20030 promote colon cancer cell growth in a cell context-dependent manner. Therefore, we refer to HT29, HCT116 and LoVo hereafter as “responsive” colon cancer cells, and the others as unresponsive cells.
The increased number of viable cells after co-culture with Sg could be due to increased proliferation, reduced apoptosis, or both. We therefore examined the effect of Sg on cell proliferation and apoptosis. Cells co-cultured with Sg or L. lactis were labeled with bromodeoxyuridine (BrdU) and analyzed by flow cytometry. Co-culture with Sg TX20005 resulted in ~3 and ~ 1.7-fold increase in the percentage of S phase cells in HCT116 and HT29 cells, respectively, compared to cells incubated with L. lactis or cells only control (Fig 2A and 2B and S1 Fig). No significant changes in the percentage of S phase cells were observed in FHC cells following incubation with TX20005, as compared to no bacteria or L. lactis-incubated FHC cells (Fig 2C, S1 Fig). We further determined the level of proliferating cell nuclear antigen (PCNA), a marker for cell proliferation [37], in cells incubated with Sg, L. lactis or cells only. The results showed that HCT116 and HT29 cells incubated with TX20005 had significantly higher levels of PCNA compared to cells co-cultured with L. lactis or cells only control (Fig 2D, 2E, 2G and 2H). No difference was observed in the PCNA level in FHC cells between the different conditions, as expected (Fig 2F and 2I). These results indicate that Sg promotes cell proliferation in responsive cells.
We next examined the effect of Sg on cell apoptosis in HCT116, HT29, and FHC cells co-cultured with TX20005, L. lactis, or media only. The cells were stained with anti-Annexin V antibodies and propidium iodide followed by flow cytometry analysis. No significant difference was observed in the percentage of apoptotic cells between the different groups in any of the cell lines (Fig 2J–2L and S2 Fig). To further confirm this, we compared the level of cleaved caspase 3 and observed no difference between the different conditions in any of the cell lines tested (S3 Fig). Taken together, these results indicate that Sg does not affect cell apoptosis, but promotes colon cancer cell proliferation in a cell context-dependent manner.
We next examined the effect of an expanded panel of bacterial strains on HT29 and HCT116. The panel included Sg strains TX20005, TX20030 and TX20031, and strains of closely related species within the SBSEC—S. infantarius (TX20012), S. macedonicus (TX20026), and S. pasteurianus (TX20027). E. coli strain XL-1 Blue and L. lactis were included as negative control bacteria. All three Sg strains significantly increased HT29 and HCT116 cell numbers at 24 (Fig 3A and 3D) and 48 hour (S4A and S4B Fig) time points, whereas none of the other bacterial strains had any effect.
Bacterial growth under co-culture conditions were determined (S5 Fig). There was no significant difference in the growth curve between the different strains as analyzed by ANOVA. TX20026 showed virtually the same growth curve as TX20005. For TX20012, TX20027 and L. lactis, the bacterial counts were lower than TX20005 at 6 hours. However, there was no significant difference in the bacterial counts between these strains and TX20005 after 6 hours.
In the co-culture experiments described above, the bacteria added to the wells were from stationary phase cultures. We examined the effect of exponential phase cultures of TX20005, TX20030 and L. lactis on HT29 and HCT116 cells. In contrast to stationary phase bacteria, exponential phase TX20005 or TX20030 did not cause any significant increase in HT29 (Fig 3B) or HCT116 (Fig 3E) cell numbers at 24 hours compared to the no bacteria controls. Thus, the results suggest that the ability of Sg to promote cell proliferation is growth phase-dependent.
We next examined whether secreted bacterial factors or bacterial metabolites in the culture supernatant were sufficient to promote colon cancer cell growth. Supernatants from stationary phase cultures of TX20005, TX20030 and E. coli were collected and filtered to remove any residual bacteria. HT29 and HCT116 cells were cultured in media only or media supplemented with the culture supernatants. Bacterial culture supernatants were insufficient to promote cell proliferation in HT29 and HCT116 cells (S6A and S6B Fig). The inability of Sg culture supernatants to stimulate host cell proliferation could be due to the possibility that the proliferation-promoting effect of Sg required bacteria to directly contact host cells. It could also be due to that the factors/metabolites in the culture supernatants were unstable and required a continuous presence of live bacteria in the culture. To distinguish these two possibilities, we used a transwell system in which bacteria were cultured in inserts with permeable membranes of 0.4 μm pore size. This pore size allows the passage of secreted bacterial factors and metabolites but not bacteria. Culturing Sg bacteria in transwells resulted in a complete loss of the proliferation-promoting effect of TX20005 on both HT29 and HCT116 cells (Fig 3C and 3F). HT29 cells were also incubated with either heat-killed bacteria or bacterial lysates, prepared from stationary phase culture. We did not observe any increase in the cell number compared to the control group under either treatment conditions (S7 Fig), suggesting that live bacteria are necessary to produce the observed effect on proliferation shown in Fig 1.
Taken together, these results suggest that the proliferation-promoting effect of Sg is dependent on Sg-specific factors, bacterial growth phase and direct contact between live bacteria and responsive cells.
Since direct contact between bacteria and cancer cells is required to promote cell proliferation, we investigated the ability of Sg to adhere to responsive and unresponsive cell lines. The results showed that both TX20005 and TX20030 adhered to HCT116, HT29, and A549 cells at a similar level (~ 20% of the initial inoculum) and slightly higher to SW1116 and SW480 (~ 30% of the initial inoculum). Adherence to CCD 841 CoN colonic epithelial cells was significantly lower than to the colon cancer cell lines (~15% of the initial inoculum) (S8A Fig). The ability of Sg strains to invade these cells was also investigated using gentamicin protection assay. Hardly any intracellular bacteria were detected, suggesting that the two Sg strains are poorly or non-invasive (S8B Fig). We further investigated the adherence ability of stationary and exponential phase TX20005 and TX20030. Exponential phase TX20005 and TX20030 adhered to HCT116 cells significantly less than stationary phase bacteria (S8C Fig).
Together these results suggest that the ability of Sg to adhere to responsive cells is also growth phase-dependent, consistent with the growth phase-dependency observed in cell proliferation assays. On the other hand, Sg strains are able to adhere to unresponsive cells as well as to responsive cells, suggesting there may be multiple interactions between Sg and different cell lines.
The Wnt/β-catenin signaling pathway regulates cell fate and proliferation and is a critical pathway in colon tumorigenesis [38–40]. We investigated the effect of Sg on β-catenin in responsive and unresponsive cells. For HCT116 and HT29 cells, co-culture with TX20005 led to a significantly elevated level of total β-catenin compared to cells co-cultured with L. lactis or no bacteria after 12 hours of incubation (Fig 4A–4D). In contrast, no increase in β-catenin level was observed in unresponsive FHC, SW480 and SW1116 cells following Sg co-culture (Fig 4E, 4F, 4H and 4I). In addition, Sg had no effect on the level of β-catenin in A549 cells (S9 Fig). Upon activation, β-catenin is translocated into the nuclei and triggers the enhanced expression of downstream targets, such as c-Myc and cyclin D1 [41]. We examined the level of nuclear β-catenin. The results showed that HCT116 and HT29 cells co-cultured with TX20005 had significantly increased nuclear β-catenin compared to cells co-cultured with L. lactis or cells only (Fig 4A–4D). No change in nuclear β-catenin was observed in FHC cells under the same experimental conditions (Fig 4E and 4F). In accordance with this observation, the level of c-Myc (Fig 4A–4D) and cyclin D1 (Fig 4G) in HCT116 and HT29 was also significantly increased following co-culture with TX20005 compared to that in the control groups (Fig 4A–4D). No increase in the level of c-Myc (Fig 4E and 4F) or cyclin D1 (Fig 4G) was observed in FHC cells, as expected. In A549 cells, we also observed no changes in the level of c-Myc (S9 Fig). Taken together, these results suggest that incubation of responsive cells with Sg results in up-regulation of β-catenin and its oncogenic downstream targets.
To determine the role of β-catenin in Sg-mediated cell proliferation, β-catenin stable knockdown cells were generated using two specific shRNA. Knockdown was confirmed using western blot assays (S10A Fig). In co-culture experiments, β-catenin knockdown completely abolished the effect of Sg on cell proliferation, whereas cells transfected with control shRNA showed a similar increase in cell numbers as untransfected cells (Fig 5A and S10B Fig). To further confirm this, we used a β-catenin responsive transcription (CRT) inhibitor iCRT3, which disrupts β-catenin-TCF4 interaction [42]. In the presence of iCRT3, TX20005 co-culture with HT29 cells did not increase cell proliferation compared to the control groups (Fig 5B). We next examined the effect of TX20005 on the level of c-Myc and PCNA in the presence of iCRT3. Treatment of HT29 cells with iCRT3 significantly reduced the effect of TX20005 on c-Myc and PCNA expression (Fig 5C and 5D). To determine if β-catenin knockdown or iCRT3 treatment affected bacterial adherence to HT29 cells, we performed adherence assays. We did not observe any significant change in the adherence of TX20005 to either β-catenin knockdown cells or cells treated with iCRT3 compared to untransfected or untreated cells (S10C Fig). Taken together, these results indicate that promotion of cell proliferation by Sg is through up-regulation of β-catenin dependent signaling.
HCT116 cells cultured with TX20005 or L. lactis were injected into nude mice and tumor growth monitored (Fig 6A). Starting from day 13, TX20005-treated cells formed significantly larger tumors than L. lactis-treated cells. Expression of β-catenin, c-Myc and PCNA was analyzed in tumors obtained at day 21 (Fig 6B and 6C). A significant increase in the levels of β-catenin, c-Myc and PCNA were observed in tumors from TX20005-treated cells compared to those from L. lactis-treated cells. We also tested the effect of Sg using the non-responsive SW480 cells. The results showed that TX20005-treated SW480 cells did not form bigger tumors compared to cells treated with L. lactis (S11 Fig). These results indicate that TX20005 treatment promoted tumor growth in the xenograft model, and that this growth promotion requires responsive cells.
To further evaluate the role of Sg in the development of colon tumor, we used an AOM-induced mouse model of CRC. This model is commonly used to represent sporadic CRC. Mice were treated with 2 doses of AOM followed by antibiotic treatment for a week and then orally gavaged with TX20005, L. lactis or saline for 24 weeks. Colons were harvested for visual examination for macroscopic tumors. Overall, most of the tumors were found in the distal portion of the colon. We observed that Sg-treated mice had higher tumor burden compared to both the saline and L. lactis control groups (Fig 7A).
H&E stained colon sections were evaluated. Colons from Sg-treated mice showed a significantly higher average dysplasia grade compared to those from L. lactis-treated or saline control mice (Fig 7B). Adenocarcinomas were observed in Sg-treated mice but not in the control groups (Fig 7C). We next examined cell proliferation and apoptosis in mouse colonic crypt cells. Sg-treated mice had a significantly higher percentage of proliferating cells (BrdU+) in the colonic crypts compared to L. lactis- or saline-treated control groups (Fig 7D and 7E). In addition, Sg-treated mice had higher levels of β-catenin in the colon epithelium as compared to L. lactis-treated or saline controls (Fig 7F). In contrast, we did not observe any significant difference in the percentage of apoptotic cells between the different treatment groups as determined by TUNEL assays (Fig 7G, S12 Fig). These results are consistent with the observations from our in vitro cell culture assays.
Inflammation in the colon of mice was also scored. Both Sg- and L. lactis-treated groups displayed significantly higher average inflammation scores compared to the saline group. However, overall inflammation in these groups was mild and there was no apparent difference between the Sg- and L. lactis-treated groups (Fig 7H). We further measured the level of cytokines TNFα, IL-1β, IL-6, IL-17, IL23 and Cox-2 in tumor and normal tissues collected from the mouse colon using RT-qPCR. There was no significant difference between TX20005 and L. lactis treated groups in any of the cytokines tested (S13 Fig). These results suggest that Sg and L. lactis trigger similar inflammatory responses.
We further tested the effect of Sg on tumor development using a different shorter procedure, in which mice were treated with four doses of AOM and gavaged with bacteria for 12 weeks. Similar to the results from the first longer procedure, a significant increase in tumor numbers was observed in Sg-treated mice compared to the saline control (S14A Fig). When compared to the L. lactis group, mice gavaged with TX20005 also had more tumors; however, the difference was not statistically significant (p = 0.08). Tumor burden also displayed a similar trend as that observed in the longer procedure, in which Sg-treated mice had a higher average tumor burden than the other two groups (S14B Fig). In the shorter procedure, however, the difference was not statistically significant (p = 0.06 vs. L. lactis-treated mice) perhaps due to reduced duration, less bacterial gavage or more AOM injections in this second procedure. Overall, results from the two procedures show a consistent trend pointing towards Sg acting as a promotional agent for tumor development in the mouse colon.
To determine whether the abundance of TX20005 in the colon correlates with tumor number or burden in the mice, we collected fecal material from mice at the end of the 12-week gavage experiment. Relative abundance of Sg was determined by qPCR using Sg specific primers. We observed positive and statistically significant correlations between the relative abundance of TX20005, tumor number (Fig 8A, Pearson’s r = 0.77, p = 0.001), and tumor burden (Fig 8B, Pearson’s r = 0.60, p = 0.02), respectively. We further stained colon sections from mice treated with Sg or saline with anti-Sg antiserum. The antiserum was tested against strains of closely related species in SBSEC, as well as Enterococcus faecalis, E. coli, and L. lactis. The antiserum specifically recognized Sg but not the other strains (S15 Fig). In colon sections, we observed positive staining in Sg-treated mice but not in the saline control (Fig 8C and 8D), further indicating that the antiserum was specific. Sg bacteria were found within tumor tissues. The presence of Sg around normal-looking crypts was observed only occasionally, suggesting a preferential association of Sg with tumor tissues. Taken together, the results described above suggest that Sg promotes colon tumor development in the AOM-induced mouse model of CRC and this promotion involves up-regulation of β-catenin and increased colonic crypt cell proliferation.
While the association between patients with bacteremia/endocarditis caused by Sg and CRC is well documented, this population only represents a small proportion of CRC patients. Previous studies indicated that CRC patients might be “silently” infected with Sg in their colon with no symptoms of bacteremia or endocarditis [35,36]. However, the prevalence of Sg in CRC patients has not been extensively studied. We analyzed 148 tumors and 128 adjacent matched normal tissues from CRC patients by qPCR using Sg-specific primers. Overall, we found that ~74% of tumor tissues and ~47% of the normal tissues were positive for Sg (p < 0.0001, tumor vs. normal, Fisher’s exact test), suggesting that Sg is present in the majority of CRC patients and preferentially associates with tumor tissues. We further divided the positive samples into those with relatively high or low abundance of Sg. More tumor tissues were highly enriched with Sg (26%) than normal tissues (9%) (p = 0.0025, Fisher’s exact test), indicating a higher bacterial abundance in the tumor tissues (Table 1). Streptococcus pasterianus (Sp), previously S. bovis biotype II, is closely related to Sg however; patients with bacteremia or endocarditis due to Sp did not display a strong association to CRC as that reported for Sg in epidemiological studies [19,43,44]. The prevalence of Sp in CRC patients with no signs of bacteremia or endocarditis is unknown. We examined a small subset of randomly selected tumor and matched normal tissues using Sp-specific primers. The results showed that only 11% of tumor tissues (n = 27) and 7% of normal tissues (n = 27) were Sp positive, significantly lower than that of Sg (p < 0.0001, Fisher’s exact test) (Table 2). Thus the high prevalence in CRC patients is observed for Sg but not for a closely related organism. Taken together, these results suggest that Sg is present in the majority of CRC patients and preferentially associates with tumor tissues.
Tissue sections of tumors from CRC patients were stained using Sg-specific antibodies to visualize the bacteria. We examined 4 normal and 21 tumor samples from CRC patients. Sg was detected in one of the normal samples (25%), and in 10 of the tumor samples (48%). Sg were seen attached to tumor tissues (Fig 9). This result is consistent with our in vitro observation that proximity of Sg to colon epithelial cells is critical for its effect on cell proliferation.
CRC is the second to third most common cancer and a leading cause of cancer death in the world. Annually, over a million people are diagnosed with CRC and ~700,000 die due to CRC [45]. In recent years, the role of microbial agents in the development of CRC has gained increasing recognition, raising hope that we may be able to exploit the knowledge on how microbes modulate the development of CRC to improve CRC diagnosis, prevention and treatment. To achieve this goal, a clear understanding of how precisely microbes exert their influence on tumor development is important. Sg infections have long been known to display a strong association with CRC, and yet virtually nothing was known about the nature of this association—i.e., whether this organism plays an active role in tumor development or its presence is merely a consequence of the tumor environment being favorable for its colonization—was unknown. The results described in this study demonstrate a tumor-promoting role of Sg that is dependent on cell context, specific bacterial factors, direct contact with colon cancer cells, and β-catenin signaling. We further show that Sg is present in the majority of CRC patients and preferentially associates with tumor tissues. Taken together, these findings highlight the importance of Sg in the development of CRC that extends far beyond previous recognitions.
That Sg promotes cell proliferation in responsive cells is based on the following results. Responsive cells co-cutlured with Sg had significantly higher number of viable cells, higher percentage of cells in S phase, higher level of PCNA, β-catenin (both total and nuclear), c-Myc and cyclin D1, compared to cells cultured in media only or with negative control bacteria. Sg did not affect the level of β-catenin or c-Myc in unresponsive cells. In addition, Sg did not affect cell apoptosis, as shown by a lack of sigficannt difference in annexin V staining or the level of cleaved caspase 3 between cells co-cultured with Sg and those with media only or with negative control bacteria.
Our results also provide evidence indicating that β-catenin is functionally important for Sg stimulated cell proliferation. Knockdown of β-catenin in responsive cells by two independent shRNA abolished Sg’s effect on cell proliferation. Treatment of cells with a β-catenin inhibitor iCRT3, which blocks the interaction between β-catenin and its partner transcriptional factor, also abolished Sg’s effect on cell proliferation, c-Myc and PCNA. Thus, Sg promotes the proliferation of responsive cells in a β-catenin dependent manner. The Wnt/β-catenin signaling pathway regulates cell proliferation and cell fate. Dysregulation of this pathway plays a central role in the development of CRC [46–50]. It is highly pertinent, therefore, that Sg also targets this critical pathway. The upstream events leading to the activation of β-catenin signaling upon exposure to Sg are unknown and are the focus of on-going studies in our laboratory. Studies on other tumor-promoting bacteria indicate that diverse strategies are used to influence β-catenin signaling. For example, Fusobacterium nucleatum modulates β-catenin signaling by binding to E-cadherin through its FadA adhesin [5]. Bacteroides fragilis secretes a zinc-dependent metalloprotease toxin that cleaves E-cadherin, leading to nuclear translocation of β-catenin, increased c-Myc expression and cell proliferation [51]. Helicobacter pylori, which is an important cause for gastric cancer, activates β-catenin signaling in multiple ways including affecting the expression of Wnt ligands [52], activating Wnt receptors [53], suppressing GSK3β [54,55], interfering with β-catenin/TCF4 complex by downregulating the gastric tumor suppressor Runx3 [56–58], and interacting with E-cadherin to disrupt the E-cadherin/β-catenin complex [59]. In addition, there have been numerous studies in recent years linking microRNA (miRNA) dysregulation to CRC (recent reviews [60–64]). Evidence suggests that microbes (e.g., H. pylori, Citrobacter rodentium, and human papillomavirus (HPV)) can regulate β-catenin signaling and cell proliferation by affecting certain miRNAs [65–67]. However, Sg does not appear to encode any homologs to any bacterial factor known to modulate β-catenin or cell proliferation [68–71], suggesting the involvement of a potentially novel mechanism.
The observation that some colon cancer cell lines (SW480 and SW1116) as well as a lung cancer cell line A549 are not responsive to the effect of Sg suggests that Sg functions by engaging host factors specific to the responsive cells. All five of the colon cancer cell lines we tested contain mutations in the Wnt/β-catenin signaling pathway; HT29, LoVo, SW480 and SW1116 have mutations in APC whereas HCT116 contains a mutated version of β-catenin that results in increased protein stability [72,73]. Sg further increases β-catenin level in HT29, HCT116, and LoVo, but not in SW480 and SW1116 cells. It is possible that Sg up-regulates β-catenin at a more upstream level or by affecting factors outside the canonical Wnt/β-catenin signaling pathway. The fact that TX20005 adheres to unresponsive colon cancer cells as well as, or even better than, responsive cells suggests that the differential effects of Sg on responsive and unresponsive cells are not due to differences in the amount of bacteria adhering to these cells. Rather, whether or how the signal is transduced from the cell surface where Sg is attached is likely to be responsible for the difference. It is also possible that Sg adheres to different receptors on responsive and unresponsive cells. Overall, our results suggest that the effect of Sg depends on specific cell context. This implies that not everyone colonized by Sg may be equally affected; some individuals with certain genetic or epigenetic makeup may be more susceptible to the tumor-promoting effect of Sg than others. Identifying host factors that render cells responsive to Sg will be important.
The observation that bacterial species closely related to Sg failed to stimulate cell proliferation in vitro suggests the involvement of Sg-specific factors in this process. This finding is consistent with previous clinical observations that among the closely related species in the S. bovis group, Sg displays a particularly strong association with CRC. Our results also suggest that direct contact between Sg and responsive cells is required for Sg’s effect on cell proliferation. We observed that Sg culture supernatants or Sg cultured in transwell inserts were unable to stimulate cell proliferation. Furthermore, heat-killed Sg or Sg lysates had no effect on the proliferation of responsive cells. These results suggest that direct contact or close proximity of live Sg to host target cells is required for its stimulation of cell proliferation. We do need to point out that with respect to the requirement of live Sg, we cannot exclude the possibility that the relevant Sg factor(s) is inactivated during the heating or bacterial lysis process. The results also show that both adherence to responsive cells and promotion of cell proliferation by Sg depend on bacterial growth phase. Taken together, these results suggest several possibilities. Firstly, it is possible that adherence and promotion of cell proliferation is mediated by the same Sg surface factor(s) up-regulated in bacterial stationary phase. Secondly, it is possible that adherence brings Sg close to the surface of responsive cells. Subsequent actions by other Sg surface or secreted factors then promote cell proliferation. A third possibility is that adherence of Sg to responsive cells results in the up-regulation or activation of other Sg surface or secreted factors, which in turn stimulate cell proliferation. Studies are on-going to identify Sg factors involved in cell adherence and stimulation of cell proliferation. It was reported previously that an Sg mutant deficient of the Pil3 pilus was significantly impaired in the ability to adhere to mucus-producing cells and to colonize mouse distal colon compared to a Pil3 overexpressing Sg variant strain [74]. The Pil3 pilus has been shown to bind intestinal mucins and fibrinogen [75]. Sg also expresses a collagen binding protein Acb that mediates binding to collagen I, IV and V [70]. The genome of Sg encodes proteins with homology to fibrinogen/fibronectin binding proteins and major cell surface adhesion pac, as well as additional pilus operons [68,69]. These surface proteins can potentially play a role in Sg adherence to host target cells and/or colonization of the mouse colon.
The results from mouse models suggest that Sg promotes tumor development. Sg-treated responsive cells developed larger tumors in a mouse xenograft model than cells treated with control bacteria L. lactis. Higher levels of β-catenin, c-Myc and PCNA were also observed in Sg-treated xenografts compared to L. lactis-treated ones. In the AOM CRC model, mice treated with Sg had more tumors and higher tumor burden compared to L. lactis or saline-treated mice. This was confirmed using two different experimental procedures. In addition, Sg-treated mice had a higher percentage of proliferating cells and stronger β-catenin staining in colonic crypts compared to the control mice. Apoptosis in colon epithelial cells of Sg-treated mice was similar to that in L. lactis-treated mice. These findings are consistent with the results from cell culture assays. Furthermore, we observed a significant and positive correlation between Sg bacterial burden in the mouse colon and tumor number and burden, respectively, suggesting a dose effect. Finally, Sg bacteria were detected within tumor tissues from the AOM CRC model and were seen directly associated with tumor tissues from CRC patients. These results are consistent with the in vitro finding and suggest that Sg promotes tumor development by adhering to or in close proximity with tumor.
The observation that Sg and L. lactis induced similar levels of inflammatory responses suggests that Sg-induced immune responses may not play a major role in Sg-mediated tumor promotion. However, this does not exclude the possibility that Sg may induce specific types of immune reactions that favor tumor development or Sg’s promotion of tumor development requires a certain component of the immune system. In addition, the role of microbiota in Sg-mediated tumor promotion remains unclear. The results here suggest a direct effect of Sg on colon epithelial cells. However, whether Sg functions in concert with other microbial agents in the gut or elicits specific responses when mixed with certain other microbes is unknown. It is also unclear whether Sg plays a role in tumor initiation, since all mice were treated with AOM. Further studies are needed to clarify these issues. Overall, the results presented here support a model in which Sg actively promotes colon tumor growth. This promotion appears to involve up-regulation of β-catenin by Sg, resulting in increased cell proliferation.
Previous epidemiology studies have primarily focused on CRC risks for patients with bacteremia/endocarditis due to Sg. These patients only constitute a small proportion of CRC patients. Limited information was available on the prevalence of Sg in CRC patients with no signs of infections [35,36]. Abdulamir et al. studied 52 CRC patients without symptoms of bacteremia and found that approximately 33% of tumors and 23% of matched normal colon tissues were Sg-positive when a conventional PCR method was used for the detection of Sg [36]. A recent study investigated the presence of S. bovis group organisms in the colonic suction fluid from individuals who underwent colonoscopy [35]. S. bovis was isolated from each of the 17 patients diagnosed with malignant tumors. However, it was not clear if the S. bovis isolates were Sg or other species within the S. bovis group. Here we analyzed 148 tumor and 128 matched adjacent normal tissues from CRC patients and showed that ~ 74% of CRC patients were Sg positive and that Sg preferentially associates with tumor tissues. The difference in Sg prevalence between this study and others could be due to the different detection methods used, patient characteristics and/or the fact that samples were collected from different geographical regions (Malaysia [36], Israel [35], and United States (this study), respectively). Despite the difference, these studies appear to have a common theme–Sg is present in a much larger proportion of CRC patients than previously recognized and preferentially associates with tumor tissues. More extensive epidemiology studies are needed to further define the prevalence of Sg in CRC patients and its correlation with clinical, pathological and molecular characteristics of CRC. Overall, the results taken together indicate that a large proportion of CRC patients are “silently” infected with Sg. This information combined with our finding that Sg actively promotes tumor growth further highlights the clinical importance of this organism.
Previous studies have shown that S. bovis or Sg was isolated from fecal cultures in approximately 10% of healthy individuals [27,36]. The factors that affect Sg abundance in healthy individuals are unclear. It is plausible that host factors, gut microbiome composition and antibiotics usage can all have an impact. The prevalence of Sg in healthy individuals is lower than that in CRC patients. It was found that metabolites produced by colon cancer cells facilitate Sg outgrowth, resulting in a significant growth advantage of Sg over other bacteria in a simulated colon tumor microenvironment [76]. This could be a possible explanation for the higher prevalence of Sg in CRC patients. It is also possible that as tumors develop in the colon, changes in the inflammatory status and microbiome composition creates a colon environment that favors Sg survival and growth. Altered expression of surface proteins on tumor cells may also facilitate Sg adherence to the cells and colonization of the colon tissue.
In summary, this is the first report demonstrating a tumor-promoting role of Sg. The findings here have important clinical implications. Going forward, identifying the Sg factor(s) responsible for promoting cell proliferation and tumor development, and host factors targeted by Sg, will be critical for understanding how Sg functions as a tumor-promoting agent and for developing optimized strategies based on both bacterial and host characteristics to better diagnose and treat CRC.
S. bovis group strains (S. gallolyticus, S. pasteurianus, S. infantarius, S. macedonicus) [70], Lactococcus lactis MG1363 (provided by Timothy J. Foster, Trinity College Dublin, Ireland), and E. coli XL-1 Blue were grown at 37°C in brain-heart infusion (BHI) broth with shaking or on BHI agar (Difco Laboratories, Sparks, MD). Starter cultures were prepared by growing strains overnight in 3 ml BHI broth. Fresh BHI broth was then inoculated with the overnight culture at 1:100 ratio. Cells were harvested at 0.5 OD600nm for exponential phase bacteria and at 1.0 OD600nm for stationary phase bacteria.
Human colon cancer cell lines HCT116, HT29, LoVo, SW1116, SW480 and kidney epithelial cell line HEK293 (obtained from Dr. Scott Koptetz, Universoity of Texas M. D. Anderson Cancer Center, Houston, Texas and Cheryl L Walker, Baylor College of Medicine, Houston, Texas) were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM, GIBCO, USA) supplemented with 10% fetal bovine serum (FBS) (GIBCO, USA). Human normal colon epithelial cell lines CCD 841 CoN and FHC were cultured in DMEM/F12 (GIBCO, USA) supplemented with 15% FBS (GIBCO, USA). Human lung carcinoma cell line A549 was maintained in F12-K media supplemented with 10% FBS. SW480, CCD 841 CoN, FHC, and A549 were purchased from ATCC. All of the cells were cultured in a humidified incubation chamber at 37°C with 5% CO2. HT29 stable knockdown cells were established by infecting cells with Lentiviral Transduction Particles containing short hairpin RNA (shRNA) against CTNNB1 (Sigma-Aldrich, TRCN0000314991 and TRCN0000314990) or MISSION pLKO.1-puro Non-Mammalian shRNA Control (Sigma-Aldrich, SHC002) and selected with puromycin (1μg/ml). The lentiviral CTNNB1 shRNA plasmid or a non-mammalian shRNA control plasmid were transfected into HEK293T cells to produce lentiviral particles. Gene knockdown was confirmed by Western blot assays. All cell lines were validated based on short tandem repeats (Characterized Cell Line Core Facility, University of Texas M. D. Anderson Cancer Center, Houston, TX).
Cells were seeded onto the wells of 6-well plates at 1x104 cells per well and incubated for 12 hours. Exponential or stationery phase bacteria were washed with sterile phosphate buffered saline, pH 7.4 (PBS), resuspended in the appropriate cell culture media, added to the wells at 1x102 cfu/well, and incubated for 24 or 48 hours. Antibiotic trimethoprim was added at 50 μg/ml final concentration after 24 hours of incubation to prevent media acidification due to bacterial growth. Trimethoprim is generally bacteriostatic for Gram-positive bacteria [77]. We have also confirmed that at the concentration we used, trimethoprim is bacteriostatic and does not kill Sg. Cells were detached by trypsin treatment, stained with trypan blue and counted in a Cellometer® Mini automated cell counter (Nexcelome Biosciences, Lawrence, MA). Transwell experiments were performed as described above except that bacteria were cultured in Transwell inserts (Corning) with 0.4 μm pore size, which allows the passage of secreted factors and soluble metabolites but not bacteria. Each experiment was performed in duplicate and repeated at least three times. Treatment of cells with iCRT3 was based on a protocol previously described [78]. Briefly, 25 μM of iCRT3 (Sigma) was added to the cells 1 hour before the addition of bacteria. Fresh inhibitor was added every 4 hours subsequently until the end of the experiment. iCRT3 has no effect on the growth of L. lactis or Sg. To determine the effect of bacterial culture supernatants, stationary phase cultures grown in BHI were centrifuged and supernatants collected and filtered through a 0.2 μm filter to remove residual bacteria. 0.5 ml of a supernatant was added to each well in a 6-well culture plate.
To prepared heat killed bacteria, bacterial cells from stationary phase were incubated at 95°C for 30 mins. Bacterial lysate was prepared by using a cell disruptor (Constant Systems Ltd, UK). 100 ml of stationary phase bacterial culture was pelleted and resuspended in 10 ml PBS and then processed in the cell disruptor.
This was performed following a procedure described previously with slight modifications [79]. Cells were seeded onto the wells of 24-well tissue culture plates at 106 cells/well. Bacteria from a stationary phase culture were washed twice in PBS, resuspended in DMEM supplemented with 10% FBS, and added to the wells at a multiplicity of infection (MOI) of 10. The plates were incubated in a humidified incubation chamber at 37°C with 5% CO2 for 1 hour. Each well was washed three times with sterile PBS to remove unbound bacteria. To determine the number of associated bacteria, cells were lysed with sterile PBS containing 0.025% Triton X-100 and dilution plated. For the detection of internalized bacteria, after washing with PBS, cells were incubated in DMEM containing 10% FBS, gentamicin (200 μg/ml) and ampicillin (200 μg/ml) for 1 hr. Cells were washed again with PBS for three times, lysed and dilution plated. All experiments were performed in triplicate wells and repeated at least three times. Adherence and internalization was expressed as a percentage of total bacteria added.
Cells were cultured in the appropriate medium in the presence or absence of bacteria for 12 hours and washed with sterile PBS three times. To obtain total cell lysates, cells were lysed with a lysis buffer (1% Triton X-100, 50 mM HEPES, pH 7.4, 150 mM NaCl, 1.5 mM MgCl2, 1 mM EGTA, 100 mM NaF, 10 mM Na pyrophosphate, 1 mM Na3VO4, 10% glycerol, and phosphatase inhibitor cocktail (Sigma)). To extract nuclear proteins, cells were resuspended in 200 μl of a buffer consisting of 10 mM HEPES pH 7.9, 10 mM KCl, 0.1 mM EDTA, 0.1 mM EGTA and 10 μl of 1%NP-40. Cells were vortexed for 30 sec and then centrifuged at 2000g for 5 min. Pellets were then resuspended in 100 μl of a buffer consisting of 20 mM HEPES pH 7.9, 500 mM NaCl, and 1 mM EDTA and vortexed for 10 min. Lysates were centrifuged at 14,000 rpm for 10 min and supernatants were used for western blots. To extract protein from tumor tissues, tissues were homogenized using a Tissue LyserLT (Qiagen) and lysed with the lysis buffer used for total cell lysates. The lysates were subjected to SDS-gel electrophoresis and western blot. Rabbit polyclonal antibodies against β-catenin (1:4000), c-Myc (1:3000), PCNA (1:2000), cleaved caspase 3 (1:1000), lamin B1 (1:1000) and β-actin (1:5000) were all from Cell Signaling Technology (CST). Horse radish peroxidase (HRP)-conjugated anti-rabbit IgG (CST) was used as the secondary antibody. Signals were detected using HyGLO, chemiluminescent HRP (Denville, Mteuchen, NJ). Band intensity was quantified using Image J.
Total RNA was extracted from co-cultured cells or colon tissues using the RNeasy Kit (QIAGEN) or All-Prep DNA/RNA/Protein Mini kit (QIAGEN). cDNA was generated by using the Transcriptor First Strand cDNA Synthesis Kit (Roche). qPCR was performed using FastStart SYBR green master mix (Roche) in a Viia 7 Real Time PCR System (Applied Biosystems). The following cycle conditions were used: 95°C for 10 minutes followed by 40 cycles at 95°C for 30 seconds and 60°C for 1 minute. For cyclin D1, CT values were first normalized to GAPDH, then to cells cultured in media only (ΔΔCT). For cytokines in colon tissues, CT values were normalized β-actin (ΔCT).
Cells were co-cultured with bacteria for 12 hours and washed with sterile PBS three times. To detect proliferating cells, cells were incubated with 5-bromodeoxyuridine (BrdU) (BD Biosciences) at a final concentration of 10 μM in cell culture media for 30 mins. Cells were washed and stained for BrdU incorporation by using BrdU Flow kit (BD Pharmingen) according to manufacturer’s instructions. For the detection of cells undergoing apoptosis, cells were stained with prodium iodide (PI) and anti-Annexin V antibodies by using Annexin V-FITC Apoptosis Detection Kit (BD Phamingen). Flow cytometry analysis of samples was done using a LSRII flow cytometer (Becton-Dickinson), and data were analyzed using the FCS Express 3 software.
All animal experiments were performed according to prototocls approved by the Institutional Animal Care and Use committee at Texas A&M Health Science Center. (1) Xenograft model. HCT116 or SW480 cells (1 x 106) were incubated with TX20005 or L. lactis (MOI = 1) for 12 hours. The cells were immediately washed, trypsinized and mixed with Matrigel (Corning, MA) according to the manufacturer’s instructions and subcutaneously injected (100 μl) into the dorsal flap of 5-week-old nude mice (Jackson Laboratory, Bar Harbor, ME). Three hours after the injection, mice were administered a broad-spectrum antibiotic imipenem (MSD) by intraperitoneal (i.p.) injection (150 mg/kg body weight). Tumor diameters were measured with a digital caliper, and tumor volume calculated using the formula: Volume = (d1xd1xd2)/2, with d1 being the larger dimension [80]. (2) AOM-induced mouse model of CRC. Eight-week old female A/J mice (Jackson Laboratory, Bar Harbor, ME) were treated with AOM (10 mg/kg body weight) by i.p. injection once a week for 2 or 4 weeks. Mice were then given ampicillin (1g/L) in drinking water for one week and switched to antibiotic-free water 24 hours prior to bacterial inoculation. Mice were orally gavaged with saline, TX20005 or L. lactis using a feeding needle (~ 1 x 108 cfu/mouse) at a frequency of three times per week for 24 weeks, or once a week for 12 weeks and were euthanized one week after the final gavage. One hour before sacrifice, mice received an i.p. injection of BrdU at 100 mg/kg body weight. Colons were removed by cutting from the rectal to the cecal end and opened longitudinally for visual evaluation. Tumor number was recorded and tumor size measured using a digital caliper. Tumor burden was calculated as the sum of all the tumor volumes of one mouse. Visual evaluation was carried out by two blinded observers. Mice were fed with standard ProLab IsoPro RMH3000 (LabDiet).
At necropsy, colons from 3 randomly selected mice from each group were “Swiss rolled” from the rectal to the cecal end, fixed in Methcarn (60% methanol, 30% chloroform, and 10% glacial acetic acid), paraffin embedded, and cut into 5μm sections across. Every 10 sections were stained with hematoxylin/eosin (H&E) and histological evaluation performed by a pathologist in a blinded fashion. Pathological scores were given using the following scale [81]: 0, no dysplasia; 1, mild dysplasia characterized by aberrant crypt foci (ACF), +0.5 for small gastrointestinal neoplasia (GIN) or multiple ACF; 2, moderate dysplasia with GIN, +0.5 for multiple occurrences or small adenoma; 3, severe or high grade dysplasia restricted to mucosa; 3.5, adenocarcinoma (involvement through muscularis mucosa); 4, adenocarcinoma (through submucosa and into or through the muscularis propria). Inflammation was scored using the following scoring matrix [82]: 0, normal; 1 - </ = 1 multifocal mononuclear cell infiltrates in lamina propria accompanied by minimal epithelial hyperplasia and slight to no depletion of mucous from goblet cells; 2, involves more of intestine or more frequent, occasional small epithelial erosions, no submucosa involvement; 3, moderate inflammation plus submucosa neutrophils, crypt abscesses, ulcers; 4, most of colon; transmural; crowding of epithelial cells with elongated crypts, ulcers plus crypt abscesses [82].
Proliferating crypt cells were detected by staining every 10 sections with anti-BrdU antibodies and counting BrdU-positive cells. A total of ~50 crypts were counted per mouse and the percentage of BrdU+ cells vs. total crypt epithelial cells counted was calculated. Apoptosis was determined by performing TUNEL assay on every 10 sections. Crypts were counted in the same manner as for BrdU+ cells. Sections were also stained for β-catenin. A Leica DM2000 LED microscope was used for imaging. Paraffin embedding, sectioning, histochemistry and immunohistochemistry were performed by the Histology Core, Gulf Coast Digestive Diseases Center, Baylor College of Medicine, Houston, TX.
To detect Sg in the mouse colon, fecal pellets were collected from mice at the end of 12-week gavage with TX20005. DNA was extracted using QIAamp Fast DNA Stool Mini Kit (Qiagen) following manufacturer’s instructions. Sg and Sp-specific primers were designed based on genomic sequences unique to Sg and Sp, respectively. The specificity of the primers was tested against a panel of strains of closely related species within the S. bovis group (S16 Fig). The primers correctly identified Sg, and Sp strains respectively, and not closely related species. The sequence for the oligonucleotides is as follows: forward Sg-specific primer– 5’ TGACGTACGATTGATATCATCAAC 3’, reverse Sg-specific primer –5’CGCTTAACACATTTTTAGCTAATACG 3’, forward Sp-specific primer– 5’ ATGGATAGTCATAGAATTGA 3’, and reverse Sp-specific primer– 5’ GGACAATGCCCTCATCTAGC 3’. qPCR was performed using Fast Plus EvaGreen qPCR Master Mix (Biotium) in a Viia 7 Real Time PCR System (Applied Biosystems) with the following cycling condition: 95°C for 10 minutes followed by 40 cycles at 95°C for 30 seconds and 60°C for 1 minute. Melting temperature analysis was performed at the end the cycles. ΔCT was normalized to the results from qPCR reactions using universal 16S rRNA primers. DNA was extracted from mouse colon tissues spiked with serially diluted Sg cultures of known concentration and was used to generate a standard curve. The standard curve was then used to convert ΔCT to bacterial concentration.
The same primers and cycling conditions were used to detect Sg and Sp in DNA extracted from human tumor and adjacent normal tissues. Samples enriched with Sg (strongly positive) is arbitrarily defined as with a 5 CT cutoff from the mean, which corresponds to approximately 90-fold enrichment above the average and approximately 750 CFU/100 ng DNA. PCR products from randomly selected positive samples were purified and sequenced. All were found to have the correct DNA sequence.
Rabbit serum was raised against formalin killed TX20005 (Rockland Immunochemicals). The antiserum and pre-bleed serum were tested against TX20005, S. infantarius (TX20012), S. macedonicus (TX20026), S. pasterianus (TX20027), E. coli XL-1 Blue, and L. lactis MG 1363, to determine the specificity of the antibodies. The antiserum specifically recognized Sg not other bacterial strains under the experimental conditions (S15 Fig). Methcarn-fixed paraffin embedded colon sections (5 μm) from mice treated with two weekly injections of AOM and 24 weeks of oral gavage with bacteria were used to detect Sg. Human colon tissue sections were obtained from US Biomax. Briefly, sections were deparaffinized with xylene and rehydrated in an ethanol gradient. The slides were incubated in a citrate buffer at 95°C for 15 min, cooled to room temperature (RT), rinsed with PBS and incubated in blocking buffer (PBS containing 1% Saponin and 20% BSA) for 30 min. The slides were then incubated with rabbit anti-Sg serum (1:250 dilution) at 4°C overnight, washed with PBS, and incubated with donkey-anti-rabbit Alexa 594 (1:500 dilution in PBS) for 1 hr at RT. The slides were washed again, stained with DAPI, mounted and examined in a DeltaVision Elite microscope (GE Healthcare).
Comparisons of multiple groups were done by two-tailed one-way or two-way analysis of variance (ANOVA) followed by Student-Newman-Keuls (SNK) test. Comparison between two groups of data was done by unpaired, two-tailed t test. Pearson correlation analysis was performed to determine the correlation between Sg burden and tumor number and burden, respectively. Fisher’s exact test was used to compare the qPCR data from human tissues. Analyses were carried out using the Graphpad Prism 6 software.
Animal studies were performed in accordance with protocols (IACUC#2014-0370-IBT) approved by the Institutional Animal Care and Use Committee at the Texas A&M Health Science Center, Institute of Biosciences and Technology. The Texas A&M University Health Science Center—Institute of Biosciences and Technology is registered with the Office of Laboratory Animal Welfare per Assurance A4012-01. It is guided by the PHS Policy on Human Care and Use of Laboratory Animals (Policy), as well as all applicable provisions of the Animal Welfare Act.
Colon biopsy samples were collected from patients at the University of Texas M. D. Anderson Cancer Center (MDACC), Houston, TX. The patients had previously given written informed consent for use of their samples in future colorectal cancer research. Patient identifiers were anonymized. The mean age at surgery is 62.5. The cohort contains mostly stage II and III tumors. Collection and handling of patient samples were carried out in strict accordance to protocols approved by the institutional review board at MDACC and Texas A&M Health Science Center.
|
10.1371/journal.pgen.1001034 | Sex Reversal in Zebrafish fancl Mutants Is Caused by Tp53-Mediated Germ Cell Apoptosis | The molecular genetic mechanisms of sex determination are not known for most vertebrates, including zebrafish. We identified a mutation in the zebrafish fancl gene that causes homozygous mutants to develop as fertile males due to female-to-male sex reversal. Fancl is a member of the Fanconi Anemia/BRCA DNA repair pathway. Experiments showed that zebrafish fancl was expressed in developing germ cells in bipotential gonads at the critical time of sexual fate determination. Caspase-3 immunoassays revealed increased germ cell apoptosis in fancl mutants that compromised oocyte survival. In the absence of oocytes surviving through meiosis, somatic cells of mutant gonads did not maintain expression of the ovary gene cyp19a1a and did not down-regulate expression of the early testis gene amh; consequently, gonads masculinized and became testes. Remarkably, results showed that the introduction of a tp53 (p53) mutation into fancl mutants rescued the sex-reversal phenotype by reducing germ cell apoptosis and, thus, allowed fancl mutants to become fertile females. Our results show that Fancl function is not essential for spermatogonia and oogonia to become sperm or mature oocytes, but instead suggest that Fancl function is involved in the survival of developing oocytes through meiosis. This work reveals that Tp53-mediated germ cell apoptosis induces sex reversal after the mutation of a DNA–repair pathway gene by compromising the survival of oocytes and suggests the existence of an oocyte-derived signal that biases gonad fate towards the female developmental pathway and thereby controls zebrafish sex determination.
| Zebrafish has become an important model for understanding vertebrate development and human disease, yet the genetic mechanisms that regulate gonad fate to determine zebrafish sex remain elusive. In this work, we describe a mutation in the fancl gene that causes zebrafish to develop exclusively as male due to female-to-male sex reversal. Fancl is a member of the Fanconi Anemia/BRCA pathway involved in the repair of damaged DNA. We find that the sex-reversal phenotype is caused by an abnormal increase of programmed germ cell death during the critical period for zebrafish sex determination in which oocytes progress through meiosis. This abnormal increase in germ cell death compromises oocyte survival, gonadal somatic cells do not maintain the female gene expression profile, gonads become masculinized to testes, and mutants develop into fertile males. Remarkably, we show that the introduction of a mutated allele of the tp53 (p53) tumor suppressor gene into fancl mutants rescues the sex-reversal phenotype by reducing germ cell death. We conclude that Tp53-mediated germ cell death alters gonad fate selection in fancl mutants by compromising oocyte survival, possibly by eliminating a hypothesized oocyte-derived signal, which alters sex determination in zebrafish.
| The existence of two differentiated sexes is common among animals and yet the mechanisms that determine sex are amazingly diverse. Among vertebrates, for instance, some species use primarily genetic factors and others rely on environmental factors to cause embryonic gonads to become testes or ovaries. Genetic sex determination (GSD) includes monogenic as well as polygenic systems, and in monogenic systems the sex-determining gene is usually found on sex chromosomes that evolved from a pair of autosomes after acquiring a novel sex-determining allele (reviewed in [1]). Mammals have an XX/XY sex chromosome system with males as the heterogametic sex, but birds and many reptiles have a ZZ/ZW sex chromosome system with females as the heterogametic sex. Among fish, both sex chromosome systems have been described [2]–[7]. In environmental sex determination (ESD), factors in the environment, such as temperature, control sexual fate [2]. GSD and ESD have long been thought of as distinct mechanisms, but recent data show regulation by both genetic and environmental factors within a single species [8]. In such species, the integration of genetic and environmental factors ultimately tips the bipotential gonads towards the male or the female fate (reviewed in [9]). For example, in medaka, a teleost fish with an XX/XY sex determination system, high temperatures can sex reverse XX females [10].
Despite the vast diversity of primary sex-determining mechanisms, genes downstream in the sex determination pathway appear to be broadly conserved among vertebrates. It has been suggested that during evolution, different species recruited different downstream genes to be the major sex-determining gene, sometimes relatively recently, and that changes at the top of the sex-determining pathway appear to be better tolerated than changes at the bottom of the pathway because they are less likely to have deleterious effects [11]. In mammals, the Y chromosome gene SRY (Sex determining region Y) is at the top of the sex determination cascade [12]–[16] and acts as a genetic switch that triggers the bipotential gonad to initiate the male pathway (reviewed in [17]). SRY however, does not appear to exist beyond therian mammals [18]. In several groups, including mammals, Dmrt1 (doublesex and mab-3 related transcription factor 1) is a downstream gene in the male sex-determination pathway, but in medaka (Oryzias latipes), a duplicated copy of dmrt1 (called DMY or dmrt1by) is the major sex-determining gene [19], [20] and recent work has shown that dmrt1 is required for testis development in chickens [21]. Interestingly, dmrt1by is absent in most Oryzias species [22], showing that the upstream regulators of sex determination can change rapidly.
Teleost fish show a broad diversity of sex determining mechanisms that range from genetic to environmental, from monogenic to polygenic, and from hermaphroditism to gonochorism (two distinct sexes) [2]. Zebrafish, like many other teleosts, have no obvious heteromorphic sex chromosomes [23]–[25]. Adult zebrafish have two differentiated sexes, but have been described to develop initially as juvenile hermaphrodites because all juveniles develop gonads with immature oocytes regardless of their definitive sex [26]–[28]. Zebrafish juvenile gonads contain immature oocytes that progress through oogenesis in about half of the individuals, which become females, but that degenerate in the other half of the individuals, which become males [26]–[28]. Oocytes begin to degenerate in a window of time (20–30 days post-fertilization (dpf)) that lasts several days and varies among individuals and among rearing conditions [26]–[31]. Because the sex of the zebrafish gonad drives secondary sex characters, gonadal sex determines the definitive sex of the fish. Zebrafish depleted of germ cells develop as infertile males [31]–[34] and it has been shown that the presence of germ cells is essential to maintain female fate in zebrafish [31]. We do not yet know, however, the primary genetic mechanisms that cause some zebrafish to become females and others to become males.
To broaden our knowledge of the genetic mechanisms involved in zebrafish sex determination, we studied a fancl zebrafish mutant that develops exclusively as male. Fanconi Anemia complementation group L (Fancl, OMIM 608111), along with 12 other Fanconi Anemia proteins, facilitates cellular responses to a variety of stresses, including signals of DNA damage and apoptosis [35], [36] and belongs to the Fanconi Anemia/BRCA DNA repair pathway. In humans, mutations in any of these Fanconi genes can cause Fanconi Anemia (OMIM 227650), a recessive disease characterized by bone marrow failure, high risk of acute myeloid leukemia, development of squamous cell carcinomas of the head and neck, and developmental abnormalities in many organs including gonads, which causes hypogonadism, impaired gametogenesis, defective meiosis and sterility [37], [38]. Fancl is a member of the Fanconi Anemia core complex with a Plant Homeo Domain (PHD) that mono-ubiquitinates Fancd2 and Fanci [39], [40], which co-localize with BRCA1 and BRCA2 proteins in nuclear foci to stimulate DNA repair. A severe allele of human FANCL causes a clinical phenotype that includes hematopoietic and skeletal abnormalities that are similar to, or more severe than, those typically observed in patients suffering from a defect upstream in the Fanconi Anemia pathway (H. Joenje, personal communication). We previously identified the zebrafish ortholog of the human FANCL gene [41]. Here we show that fancl homozygous mutants develop solely as males and that the absence of fancl mutant females is not due to female-specific lethality but to female-to-male sex reversal. Results demonstrated that the sex reversal of fancl mutants is not due to the absence of germ cells, but to an abnormal increase of germ cell apoptosis that compromises survival of developing oocytes and masculinizes the gonads. We found that reducing germ cell apoptosis by introducing a Tp53 (p53 or tumor protein p53) mutation rescues the fancl sex reversal phenotype, and that many double mutants develop ovaries and become females. These results suggest the model that oocytes normally must progress through meiosis to signal the gonadal soma to maintain female development, and point to Tp53-mediated apoptosis of germ cells as a factor that could be targeted by environmental or genetic signals to modify zebrafish sex determination.
A zebrafish fancl mutant (allele HG10A, accession number AB353980) was generated by insertional mutagenesis in a Tol2 transposon-mediated enhancer trap screen [42]. Cloning and sequencing of genomic DNA surrounding the insertion revealed that the Tol2 construct was inserted into exon 12 of fancl, thereby disrupting the coding region of the PHD finger domain (Figure 1A and 1B), which is essential for Fancl function [39].
To determine whether the HG10A Tol2 insertion disrupts fancl transcription, we performed reverse transcriptase-PCR experiments on cDNA isolated from testes of a homozygous fanclHG10A mutant adult. To learn if the Tol2 insert formed part of the fanclHG10A transcript, we designed a forward primer in exon1 and a reverse primer in the insertion (F1 and R1 in Figure 1A). The sequence of the PCR product revealed a fanclHG10A transcript that contained the Tol2 construct inserted after codon Q318 in exon 12 (arrowhead in Figure 1B line 2). This insertion is predicted to insert seven novel amino acid residues and to introduce a premature stop codon (asterisk in Figure 1B line 2), resulting in the loss of 41 of the 57 residues of the PHD finger domain. This loss eliminates the crucial tryptophan-337 (W, double underlined in the wild type (WT) in Figure 1B line 1) that is conserved in all PHD finger-type E3 ligases, as well as histidine-330 and five of the seven cysteines (H and C, underlined in Figure 1B line 1) that are highly conserved in PHD finger domains [39], [43], [44].
To test if fancl HG10A mutants could produce fancl transcripts with an intact PHD domain due to elimination of the Tol2 insertion, we amplified the region encoding the PHD domain using primers flanking the Tol2 insertion (primers F2 in exon-11 and R2 in exon-13, Figure 1A). RT-PCR experiments revealed that fancl HG10A mutants lacked the expected 232 base pair (bp) PCR-product corresponding to the intact PHD domain found in wild-type siblings (WT in Figure 1C), but instead possessed a PCR-product of smaller size (174 bp) (fancl in Figure 1C). Cloning and sequencing of the F2-R2 products revealed that the small band from fancl mutants was a variant transcript that lacked both the first half of exon 12 and the Tol2 insertion (fanclΔTol2 in Figure 1B line 3). This fanclΔTol2 variant resulted from the joining of exon-11 to the second half of exon-12 due to a splice acceptor site that is newly created at the junction of the Tol2 insertion (Figure 1B line 3). The absence of the first half of exon-12 in the fanclΔTol2 transcript introduced a frameshift that generated an early stop codon (asterisk in Figure 1B line 3) leading to a predicted truncated protein lacking the entire PHD domain. These results show that homozygous fancl HG10A mutants have two variant transcripts, both of which encode products lacking an intact PHD finger domain shown to be essential for the ubiquitination function of Fancl [39].
To characterize the fancl HG10A phenotype, we crossed fancl +/HG10A heterozygotes (called fancl+/− below), and after genotyping the progeny by PCR, observed that all fancl HG10A/HG10A homozygous mutants (called fancl−/− below) developed exclusively as males, even though their wild-type and heterozygous siblings developed about as many females as males. Two alternative hypotheses could explain the lack of homozygous fancl mutant females: female-specific lethality or female-to-male sex reversal. To distinguish between these two hypotheses, we crossed female fancl+/− heterozygotes to male fancl−/− homozygotes. We raised 211 progeny to adulthood, determined their phenotypic sex according to sexually dimorphic characters including the color of the anal fin and body shape, and finally scored their fancl genotypes by PCR. Under normal conditions, this cross should give 50% heterozygotes (about half of which should be female), and 50% homozygous mutants (about half of which should be female), expecting a 1∶1∶1∶1 ratio of heterozygous females to heterozygous males to homozygous mutant females to homozygous mutant males. The fancl female death hypothesis predicts a 1∶1∶0∶1 ratio, or 66% heterozygotes and 33% homozygous mutants, but the sex reversal hypothesis, predicts a 1∶1∶0∶2 ratio, or equal proportions (50%∶50%) of homozygous mutants (all male) and heterozygotes (males plus females). Resulting genotypes revealed 46 fancl+/− females: 62 fancl+/− males: 0 fancl−/− females: 103 fancl−/− males, which showed that about half of the progeny were fancl homozygous mutants (103/211, 49%) and the other half were heterozygous for the fancl mutation (108/211, 51%) (Figure 2). These results had strong statistical support (chi-square likelihood ratio = 0.794, p-value >0.1), thus ruling out the hypothesis that homozygous fancl mutant females died. Results, however, were consistent with the hypothesis that animals that otherwise would have become females developed as males due to female-to-male sex reversal. Sex distributions within each genotype confirmed our previous observations that all fancl homozygous mutants developed as males (n = 103, 100%), and while approximately half of fancl heterozygous siblings developed as males (n = 62, 57%), the other half developed as females (n = 46, 43%) (Figure 2). These scores showed strong statistical support for the hypothesis that fancl mutants experienced female-to-male sex reversal (chi-square likelihood ratio = 73.946, p-value<0.0001). To exclude the possibility that some of the fancl mutants could have ovaries despite their external male phenotypic characters, we dissected the gonads of adult fancl homozygous mutants (n = 45), heterozygous females (n = 11) and heterozygous males (n = 29). In all cases, we found a perfect match between external sexual characters and gonadal sex. These results ruled out the possibility that fancl mutants masqueraded as males externally while having female gonads. We conclude that the HG10A Tol2 insertion into fancl induced a female-to-male sex reversal phenotype in zebrafish.
Because germ cells play a fundamental role in controlling female sex determination in zebrafish [31], [32], we wondered if fancl could play a role in zebrafish germ cell development. To address this question, we first tested whether fancl is expressed in germ cells of wild-type zebrafish. We analyzed the expression pattern of fancl by in situ hybridization on sections of gonads at seven developmental time points encompassing representative stages of gonad development (Figure 3), including sexually undifferentiated and presumptively still bipotential gonads (e.g. 10, 17 and 23 days post-fertilization (dpf)); transitioning gonads (e.g. 26 dpf), sexually determined but still immature gonads (33 and 37 dpf), and mature adult gonads (6 months post-fertilization). Results showed no detectable fancl expression in undifferentiated wild-type gonads at 10 dpf (data not shown), but weak expression signal appeared in immature gonads at 17 dpf and 23 dpf (arrows in Figure 3A and 3B). In transitioning gonads at 26 dpf, fancl expression increased in developing germ cells (arrows in Figure 3C and 3D), and signal was clearly detected in the ooplasm of oocytes in the ovary-like gonad (arrow in Figure 3C). At 33 dpf and 37 dpf, immature gonads showed a clear morphology of ovaries or testes, and fancl expression signal was maintained in developing oocytes and spermatocytes (arrows in Figure 3E–3H).
In adult gonads, fancl expression remained restricted to germ cells, but remarkably, the intensity of the detected signal differed depending on the stage of germ cell differentiation (Figure 3I and 3J). In ovaries, the weak fancl signal detected in early stage IB oocytes (eIB in Figure 3I) contrasted with the obvious strong signal in the ooplasm of late stage IB oocytes (lIB in Figure 3I). This result suggests that oocytes up-regulate fancl transcription before they transition into stage II. At later stages of oogenesis, fancl signal became less intense as oocytes progressed through oogenesis (Figure 3I). This reduction in staining intensity may be due to the dilution of transcript as oocytes increase in volume when cortical alveoli (also known as cortical granules in non-fish species) appeared in the ooplasm (stage II) and yolk began to accumulate (stage III) (Figure 3I). We detected low levels of fancl transcript at late stages of oocyte maturation (stage IV), suggesting that fancl is part of the maternal load of messenger RNA transcripts stored in the egg and passed along to embryos. This result agrees with our detection of fancl transcripts by RT-PCR and in situ hybridization experiments even at early developmental stages before the embryonic transcription machinery becomes active [41]. In testes, fancl expression appeared in spermatocytes (sc in Figure 3J), but not in more advanced stages of spermatogenesis, including spermatids and sperm (sp in Figure 3J). This result revealed the stage-specific expression of fancl during spermatogenesis.
The finding that fancl was expressed in zebrafish germ cells during the time-window critical for gonad differentiation and sex determination (17 to 33 dpf) and was up-regulated in early stages of gametogenesis is consistent with the hypothesis that Fancl plays a specific role in germ cell development and suggests that its disruption might lead to the female-to-male sex reversal phenotype displayed by fancl mutants.
Because zebrafish depleted of germ cells by dead end (dnd) morpholino (MO) knockdown [45], [46] develop exclusively as males [31], [32], and even though adult fancl mutants are fertile, we wondered if the female-to-male sex reversal of fancl mutants could be related to extremely low numbers of germ cells during stages of sex determination in juvenile mutants, or at least in those that otherwise would have developed as females and had been reversed to males. To answer this question, we performed gene expression analyses comparing gonads of fancl homozygous mutants (fancl), wild-type sibling controls (WT) and dnd-MO knockdown animals (dnd) at key stages in sex determination: 19 dpf (Figure 4A–4I), 26 dpf (Figure 4J–4X) and 33 dpf (Figure 4Y–4M'). Expression of the germ cell specific marker vasa [47] revealed the presence of germ cells in gonads of all fancl mutants sectioned (n = 15) (Figure 4D, 4P, 4S, 4E', and 4H') and sibling controls (n = 13) (Figure 4A, 4J, 4M, 4Y, and 4B'), while all germ-cell depleted animals injected with dnd-MO (n = 16) lacked vasa signal (Figure 4G, 4V, and 4K'). The presence of substantial numbers of germ cells in all fancl mutants tested even at early stages of gonad development rules out the possibility that the near absence of germ cells is the cause of the female-to-male sex reversal in fancl mutants.
Because all fancl mutants developed as males, we wondered if fancl mutants embark upon the male pathway from the beginning of gonad development, or whether they follow a normal bipotential pathway of development that later derails exclusively to the male pathway. To address these alternatives, we used the expression of cyp19a1a (cytochrome P450 family 19 subfamily A polypeptide 1a) and amh (anti-Mullerian hormone), which are the earliest sex-specific somatic gonadal cell markers known for ovary and testis, respectively, to monitor development before gonads were sexually differentiated at the morphological level [29], [31], [48].
In 19 dpf undifferentiated gonads, somatic cells of fancl mutants, as well as those of wild-type controls and dnd-MO animals, expressed both the female marker cyp19a1a and the male marker amh (Figure 4B, 4C, 4E, 4F, 4H, and 4I). This result showed no indication that fancl mutant gonads were developing abnormally, which suggests that fancl mutant gonads initially embark upon the normal bipotential pathway of development, and later derail into the male pathway. The fact that individual gonads in both fancl mutants and WT siblings expressed both cyp19a1a and amh, as did animals lacking germ cells, suggests that the onset of expression of these somatic cell markers is independent of germ cell derived signals. These results extend to a much earlier age than previously noted (19 dpf versus 35 dpf [31]) the time at which gonads depleted of germ cells express amh.
At 26 dpf, different individual WT juveniles showed different degrees of sexual differentiation, suggesting that this age is within the transitional period of sex determination. Some WT animals had gonads with few oocytes, low expression of cyp19a1a and up-regulation of amh (Figure 4K and 4L), while others had gonads with many developing oocytes, up-regulation of cyp19a1a and absence of amh signal (Figure 4N and 4O). In contrast to WT sibling controls, at 26 dpf, all fancl mutants had gonads with no ooctyes or just a few small oocytes, and most of them (4 out of 5) lacked expression of cyp19a1a and showed up-regulation of amh (Figure 4Q and 4R). Most juvenile fancl mutants at 26 dpf, therefore, had completed the transitional period of sex determination, and had embarked on the male pathway. Only one of the five fancl mutants analyzed retained a remnant of a few cyp19a1a-expressing cells despite the presence of a considerable number of amh-expressing cells (Figure 4T and 4U); this animal was probably still transitioning towards the male pathway. In 26 dpf dnd-MO animals, all gonads were depleted of germ cells, and like fancl mutants, showed no cells or few cells expressing cyp19a1a and many cells up-regulated for the male marker amh (Figure 4W and 4X). Therefore, most fancl mutants and dnd-MO animals tipped the fate of the bipotential gonad towards the male pathway earlier than WT controls.
At 33 dpf, WT juveniles had already passed the transitional period of sex determination. Males had immature testes with no oocytes, no cyp19a1a-expressing cells and many cells with high levels of amh expression (Figure 4Z and 4A'), and females had immature ovaries, with cyp19a1a-positive somatic cells surrounding oocytes but no amh-expressing cells (Figure 4C' and 4D'). In contrast to WT sibling controls, at 33 dpf, most fancl mutant gonads (6 of 8) showed clear testes morphology, including the absence of cyp19a1a expression and up-regulation of amh expression (Figure 4F' and 4G'). Interestingly, we found two fancl mutants that still had some oocytes; in contrast to WT controls, however, these individuals showed low cyp19a1a expression and high amh signal (Figure 4I' and 4J'), which would be expected if these two fancl mutants were putative females that were in the process of sex-reversing to males. At 33 dpf, all fancl (Figure 4G' and 4J') and dnd-MO animals (Figure 4M') showed the typical male-specific up-regulation of amh. In contrast to 33 dpf dnd-MO animals, all of which lacked cyp19a1a expression (Figure 4L'), fancl mutants that still retained some oocytes showed low levels of cyp19a1a expression (Figure 4I'). These results would be expected if the presence of oocytes is essential to maintain cyp19a1a expression, and suggested the hypothesis that the female-to-male sex reversal of fancl mutants is due to abnormal development of oocytes that leads to a failure of somatic cells of the gonad to maintain cyp19a1a expression and to down-regulate amh expression.
Because the Fanconi Anemia/BRCA system is involved in the repair of damaged DNA, such as that originating in meiotic recombination, we hypothesized that oocyte development is altered in fancl mutants. To test this hypothesis, we performed a histological analysis of fancl and wild-type gonad sections stained with hematoxylin and eosin at different stages of development to follow the progression of germ cells through meiosis (Figure 5).
At 19–22 dpf, WT sibling controls and fancl homozygous mutants had undifferentiated gonads with no obvious morphological differences between genotypes. Gonads of both genotypes contained stage IB perinucleolar oocytes (arrows in Figure 5A and 5B), as indicated by the presence of nucleoli at the periphery of the nuclei [49]. Shortly after the beginning of stage IB, chromosomes decondense and form lampbrush chromosomes [50], which occurs during the diplotene stage of meiosis I as the synaptonemal complex dissolves and recombination nodules keep homologous chromosomes together [51]. We define “early” perinucleolar oocytes (epo) as stage IB oocytes that have not yet decondensed their chromosomes, and “late” perinucleolar oocytes (lpo) as stage IB oocytes that have already formed lampbrush chromosomes and entered the diplotene stage of meiosis I. Gonads of fancl (10 individuals) and WT siblings (10 individuals) at 19–22 dpf both had early (epo in Figure 5A and 5B) but not late stage IB oocytes, indicating that at this time, oocytes had not yet entered the diplotene stage of meiosis I in either genotype.
At 26 dpf (Figure 5C–5F), most WT controls (7 of 9 individuals) showed late perinucleolar oocytes that had progressed through meiosis from early to late stage IB (lpo in Figure 5C), in which lampbrush chromosomes were visible, indicating that recombination had completed and oocytes had already entered the diplotene stage of meiosis I [51]. In contrast to WT, most fancl mutants (11 of 12) lacked oocytes at late stage IB (Figure 5F), indicating that oocytes in fancl mutants failed to progress through meiosis to the diplotene stage. Only one of the twelve fancl mutants showed late stage IB oocytes (lpo in Figure 5D), and this individual also contained pyknotic cells (pc in Figure 5D), some of which were identifiable as oocytes and some of which were of unclear origin due to their advanced stage in the process of degeneration. The fancl mutants that lacked oocytes (11 of 12) also had numerous pyknotic cells (pc in Figure 5F), and showed groups of spermatogonia (sg in Figure 5F), which were also found in WT animals (sg in Figure 5E) that had gonads with a testis-like morphology.
The difference between fancl and WT controls became accentuated at 32 dpf (Figure 5G–5I). At 32 dpf, all fancl gonads lacked oocytes and had become immature testes with spermatogonia and spermatocytes (sg and sc in Figure 5I), but only about half of WT siblings had immature ovaries with late stage IB oocytes (lpo in Figure 5G) while the other half had immature testes (Figure 5H).
At adult stages (Figure 5J–5L), consistent with results observed at 32 dpf, all fancl mutants lacked oocytes and had mature testes filled with germ cells at different stages of spermatogenesis (Figure 5L). In contrast, half of the WT controls had mature ovaries filled with oocytes at different stages of oogenesis (Figure 5J), and the other half had mature testes (Figure 5K).
This analysis of developmental histology revealed that in fancl mutants, oocytes failed to progress through meiosis and rarely reached the diplotene stage. Interestingly, in contrast to wild types, we observed abundant pyknotic cells in all fancl mutant gonads at 26 dpf (pc in Figure 5D and 5F), suggesting that the absence of oocytes in older fancl mutants could be related to increased germ cell apoptosis associated with the failure to complete meiosis.
To examine whether germ cell apoptosis could be the cause of both the abnormally high number of pyknotic germ cells in fancl juvenile gonads and the absence of oocytes at late stage IB, we used immunoassay to examine the activation of Caspase-3, an early marker of apoptosis [52], [53]. We scored the number of Caspase-3-positive cells in 70 gonadal cross-sections in each of 12 individuals: six wild-type sibling controls (Figure 6A) and six fancl homozygous mutants (Figure 6B) at 25 dpf, a stage within the time-window critical for sex determination. The morphology of the Caspase-3-positive cells detected in the immunoassay (shown in red in Figure 6B), and the subsequent staining of the same slides with hematoxylin and eosin (data not shown) confirmed that the Caspase-3-positive cells were germ cells and not somatic cells, and corroborated our earlier finding that germ cells that appeared to be pyknotic in our histological analysis are indeed apoptotic cells. In many cases, the shape and size of the apoptotic Caspase-3-positive cells was appropriate for oocytes, however, we cannot rule out the possibility that some Caspase-3-positive cells might be undifferentiated gonial cells (oogonia or spermatogonia). Results revealed that the average number of apoptotic germ cells in gonads of fancl−/− mutants was almost three fold higher than in gonads of wild-type sibling controls (Figure 6C) (t-test p = 0.0058, statistically significant at the p = 0.01 level). Therefore, these results suggest the hypothesis that the absence of oocytes in fancl mutants is caused by increased apoptosis of germ cells, especially oocytes, which ultimately leads to the sex reversal phenotype observed in fancl mutants.
The hypothesis that the female-to-male sex reversal of fancl mutants is caused by increased germ cell apoptosis predicts that blocking apoptotic pathways should rescue the sex reversal phenotype. Because tumor protein Tp53 (alias p53) is an important activator of apoptosis [54], we can inhibit apoptosis in fancl mutants by introducing a tp53 mutation into the fancl mutant line. To generate double mutants, we crossed a zebrafish female carrier of the hypomorphic mutation tp53M214K [55] to a male homozygous fancl mutant, identified double heterozygotes (fancl+/HG10A;tp53+/M214K called fancl+/−;tp53+/− below) among F1 progeny by PCR, and in-crossed double heterozygotes to obtain an F2 population containing double homozygous mutants. Among the F2 raised to adulthood, 44/171 (25.7%), or about a quarter, were fancl−/− homozygous mutants. Among these 44 fancl−/− homozygous mutants, 15 were also tp53−/− homozygous mutants, from which 11 developed as females and four as males (Figure 7A). All of the fancl homozygous mutant siblings (n = 29) that were either homozygous wild type for tp53+/+ (n = 8) or heterozygous for the tp53+/− mutation (n = 21) developed exclusively as males (Figure 7A). This result shows that the female-to-male sex reversal phenotype characteristic of fancl mutants was rescued in fancl−/−;tp53−/− doubly homozygous mutants (Figure 7A). The sex-ratio scores observed in the three genotypes showed strong statistical support (chi-square likelihood ratio = 32.088, p-value<0.0001) for the hypothesis that the presence of females in fancl−/−;tp53−/− double mutants and the absence of females in the other tp53 genotypes (fancl−/−;tp53+/− and fancl−/−;tp53+/+) is linked to the tp53 genotype. Histological analyses of fancl−/−;tp53−/− females corroborated the conclusion that external female sex characteristics were accompanied by ovaries filled with normal oocytes at all stages of development similar to fancl+/+; tp53+/+ wild-type female siblings (Figure 7B and 7C).
To determine whether the tp53 mutation rescues fancl sex reversal phenotype by reducing germ cell apoptosis, we studied the activation of Caspase-3 in histological sections of fancl homozygous mutants that were either homozygous for the tp53−/− mutation (n = 5) or wild-type for the tp53+/+ mutation (n = 5) at 25 dpf, a critical stage for sex determination (Figure 7D and 7F). Counts of Caspase-3-positive cells of 70 gonadal cross-sections per animal in these ten animals showed that double homozygotes (fancl−/−;tp53−/−) had an average number of apoptotic germ cells approximately three fold lower (Figure 7E and 7F) than their fancl−/− mutant siblings that were homozygous wild-type for the tp53+/+ mutation (Figure 7D and 7F) (t-test p = 0.1032, approaching statistical significance given the small sample size). These results support the hypothesis that the tp53 mutation rescues the fancl female-to-male sex reversal phenotype by decreasing the number of apoptotic germ cells, thereby counteracting the abnormally high frequency of apoptotic germ cells observed in fancl homozygous mutants. This result is consistent with the hypothesis that the fancl mutation causes the female-to-male sex reversal phenotype by increasing germ cell apoptosis during a critical time for sex determination.
Despite the broad use of zebrafish as a model for vertebrate development, its sex determination mechanism remains poorly understood. In this work, we characterize a zebrafish fancl mutation that causes homozygotes to develop exclusively as fertile males due to female-to-male sex reversal. We show that an increase of germ cell apoptosis in mutants compromises the survival of oocytes undergoing meiosis, which may imply an alteration of the signaling between germ cells and somatic cells of the gonads, masculinization of gonads to form testes, and the development of a male phenotype. We show that the mutant sex reversal phenotype can be rescued by reducing Tp53-mediated apoptosis, which allows oocyte survival, and suggests a pivotal role of germ cell apoptosis in zebrafish sex determination. Extending these results from fancl mutants to wild-type zebrafish, we propose a model in which genetic and environmental sex determining factors act to increase or decrease germ cell apoptosis and oocyte survival and thus alters the strength of a hypothetical oocyte-derived signal that maintains expression of female genes in somatic cells and hence determines sex in zebrafish.
Fancl protein helps mediate cellular responses to a variety of stresses, especially DNA damage and apoptosis [36]. Mutations in human FANCL lead to Fanconi Anemia (FA) [39], a disease of bone marrow failure, enormous risks of cancer, and hypogonadism and impaired fertility (reviewed in [38]). Likewise, the most consistent FA phenotype in murine FA gene knockout models (e.g. Fancc, Fancg, Fanca, Fancd1, Fancd2), is hypogonadism, impaired gametogenesis and infertility (reviewed in [56]). Our work shows that the disruption of fancl in zebrafish causes homozygous mutants to develop exclusively as males due to female-to-male sex reversal rather than female-specific lethality. This is the first demonstration, to our knowledge, that a mutation in a Fanconi gene can cause female-to-male sex reversal.
Our work revealed expression of fancl in germ cells during zebrafish gonad differentiation, which is consistent with a role of Fancl in germ cell development. Other species, such as mouse, also express fancl in their germ cells [57], [58], suggesting a conserved role of Fancl in vertebrate germ cell development. Previous work had shown exclusive male development in zebrafish lacking germ cells due to total loss-of-function of dead end, nanos, ziwi, or zili [31]–[34], [59]. We demonstrate here, however, that germ cells are present throughout the entire life in all individuals homozygous for the fancl mutation, which rules out the possibility that male development in fancl mutants that otherwise would have become females is due to lack of germ cells. Work presented here shows specifically that the mere presence of germ cells is insufficient to feminize gonads, but rather, it suggests that oocytes passing through meiosis are essential to support differentiation of ovaries. Our results are in agreement with previous suggestions that zili mutants all become phenotypic males probably due to the lack of oocytes at week 4 during the window of sex determination rather than due to the total loss of germ cells at week 8 [59]. Homozygous fancl mutants, in which germ cells are always present, provide a useful tool to better understand the role of germ cell-soma signaling that tips gonad fate towards the male pathway.
Comparison of sex-specific gonadal markers among fancl mutants, WT controls and dnd-MO animals, which lack germ cells, reveals that the onset of expression of the female marker cyp19a1a and the early male marker amh in individual undifferentiated gonads at 19 dpf is similar in all genotypes. This result supports the conclusion that the onset of early somatic makers is independent of germ cell signaling [31]. These results also show that undifferentiated gonads of fancl mutants initially develop as normal bipotential “juvenile ovaries” containing oocytes at early stage IB with no obvious histological differences from gonads of WT controls.
During the critical time-window for sex determination in zebrafish (e.g. 26 dpf), however, fancl mutant gonads become morphologically different from wild-type gonads. Wild-type animals have perinucleolar oocytes that progress through meiosis from early stage IB to late stage IB with obvious lampbrush chromosomes, indicating that recombination is complete and oocytes are at the diplotene stage of meiosis I, in which homologous chromosomes begin to separate but remain attached at chiasmata [51]. In contrast to wild types, most fancl mutants lack late stage IB oocytes, indicating that oocytes fail to progress beyond pachytene stage, when recombination occurs, and do not enter diplotene. Our results show that the levels of fancl transcripts are regulated during the process of gametogenesis because fancl expression up-regulates in oocytes transitioning from early to late stage IB (Figure 3I). Consistent with this result, a large-scale gene expression profiling study of developing ovaries in trout found fancl in a group of many genes that were over-expressed when the first oocyte meioses were observed [60]. In fancl zebrafish mutants, the failure of oocytes to transition from early to late stage IB suggests that Fancl might promote the successful progression of oocytes through meiosis I or the survival of meiotic oocytes. The FA pathway is apparently involved in meiosis because in mouse, Fanca is expressed in pachytene spermatocytes and Fanca knockout mice have elevated rates of mis-paired meiotic chromosomes and increased germ cell apoptosis [37]. Whether this effect on meiosis depends on the known role of FA proteins in homologous recombination in somatic cells [61] or some other aspect of meiosis is as yet unknown.
The failure of oocytes to progress through meiosis in fancl mutants correlates with the observation that most mutant gonads do not express the female somatic marker cyp19a1a, but instead up-regulate the male somatic marker amh. Interestingly, we found a few fancl mutants with some late stage IB oocytes accompanied by expression of cyp19a1a, but also showing high expression levels of amh; we interpret these animals as females whose progress towards ovary development was being derailed due to the mutation of fancl. These results would be expected if oocytes are essential to maintain cyp19a1a expression.
We hypothesize that in juvenile fancl mutants, the absence of oocytes progressing through meiosis alters oocyte signaling to the soma that maintains the female program. Without this signal, somatic cells do not maintain the expression of cyp19a1a, do not suppress amh expression, and as a result, gonads do not become ovaries but instead become masculinized and form testes. It is likely that this signal arising from meiotic oocytes is essential for somatic pre-granulosa cyp19a1a-expressing cells to proliferate and to differentiate as mature granulosa cells. In mammals, it has been suggested that meiotic oocytes reinforce ovarian fate by antagonizing the testis pathway [62], [63]. Studies on gonadal somatic cell lineages in mice and medaka, have shown that granulosa cells of the ovary and Sertoli cells of the testis develop from a common precursor [64]–[66]. It is possible that mammalian meiotic oocytes reinforce the ovarian pathway by preventing granulosa cells from trans-differentiating into Sertoli-like cells, because the loss of oocytes in mammals induces maturing follicular cells (or pre-granulosa cells) to acquire Sertoli-like cells characteristics [67]. We hypothesize that the action of meiotic oocytes in preventing pre-granulosa cells from trans-differentiating into Sertoli-like cells is an ancestral function that has been conserved in mammals and fishes. Although our experiments do not address the question of whether somatic cells trans-differentiate in fancl mutant gonads, our results are consistent with the hypothesis that fancl mutants, which lack oocytes at the diplotene stage of meiosis, can not prevent the trans-differentiation of pre-granulosa cyp19a1a-expressing cells into Sertoli-like amh-expressing cells. This hypothesized mechanism could explain the disappearance of cyp19a1a-expressing cells and the maintenance and proliferation of amh-expressing cells in fancl mutant gonads that results in gonad masculinization. Future transcription profiling analyses comparing wild-type animals and fancl mutants lacking oocytes will help to identify genes involved in oocyte-soma signaling essential for ovary development.
We observed that the loss of oocytes in fancl mutants during the time-window of sex determination (25 dpf) is accompanied by an abnormal increase of Caspase-3-mediated apoptosis of germ cells compared to wild-type siblings. This result suggests the hypothesis that the disappearance of meiotic oocytes in fancl mutants is due to an increase in germ cell apoptosis, which provides a cellular mechanism for the female-to-male sex reversal phenotype of fancl mutants. To test this hypothesis, we suppressed cell death in fancl mutants by making them homozygous for a tp53 mutation. We show that the reduction of apoptosis in fancl−/−;tp53−/− double mutants is sufficient to promote the survival of developing oocytes and to rescue the female-to-male sex reversal phenotype of fancl mutants. Our result showing that only fancl−/−;tp53−/− double mutants developed any females, while their fancl−/−;tp53+/− and fancl−/−;tp53+/+ sibling controls developed exclusively as males, indicates that the amount of germ cell apoptosis alters sex determination in fancl mutants.
The double mutant experiments further show that Tp53 activity mediates increased apoptosis associated with the fancl mutation. Doubly homozygous fancl−/−;tp53−/− rescued females were fertile and developed normal ovaries full of oocytes maturing through all stages of oogenesis. Active Caspase-3 results show that the amount of germ cell apoptosis is lower in double homozygous fancl−/−;tp53−/− individuals than in their fancl−/−;tp53+/+ mutant sibling controls, which further supports the hypothesis that the abnormal increase of apoptosis in fancl mutants that compromises the survival of meiotic oocytes is the mechanism responsible for the female-to-male sex reversal.
We did not notice a sex ratio biased towards females in the tp53M214K mutant line. This allele, however, is hypomorphic, and may possess levels of apoptosis compatible with the male pathway. This conclusion is supported by our finding that a few fancl−/−;tp53−/− double mutants developed as males. An alternative explanation is that mechanisms of apoptosis independent of Tp53 might occur in male gonads that promote oocytes to disappear in developing testes.
Our finding of increased germ cell apoptosis in fancl zebrafish mutants is consistent with the increase of apoptosis in a variety of cell types reported in Fanconi Anemia knockout mice. For instance, Fanca−/−, Fancc−/−, and Fancg−/− knockout mice show increased apoptosis of hematopoietic or neuronal cells, which might lead to a progressive loss of stem and progenitor cells [68]–[70]. Bone marrow failure in children with Fanconi Anemia is attributed to excessive apoptosis and subsequent failure of the hematopoietic stem cell compartment (reviewed in [56]). Interestingly, Fanca−/− knockout mice also show increased male germ cell apoptosis [37], suggesting that a role of the FA network related to apoptosis of germ cells might be a conserved feature in fish and mammals. Young Fancl−/− knockout mice, in contrast to fancl mutant zebrafish, do not show sex reversal but initially develop as sterile males and sterile females. Fancl−/− knockout male mice – but significantly, not Fancl−/− knockout female mice – can recover fertility and become fertile adult males. These results suggest that Fancl is necessary for germ cell proliferation in mouse embryos and for the maturation of oocytes, but not for the proliferation or maturation of spermatogonia in adulthood [58]. In zebrafish, the fact that fancl mutant males are fertile and that fancl−/−;tp53−/− rescued females are also fertile indicates that Fancl function is not essential for the maturation of zebrafish spermatogonia and oogonia to become sperm or mature oocytes, but rather that Fancl function affects specifically germ cell survival.
The loss of oocytes progressing through meiosis in fancl mutants suggests that Fancl function is involved in the survival of developing germ cells through meiosis, and that when Fancl is mutated, developing oocytes cannot survive due to an inappropriate increase of Tp53-dependent germ cell apoptosis. This idea is consistent with the fact that genetic deletion of Tp53 can rescue the TNF-alpha dependent apoptosis caused by accumulation of the pro-apoptotic protein kinase PKR resulting from a mutation of the human FANCC gene [68], reviewed in [56]. Therefore, inappropriate activation of Tp53-dependent apoptosis might be a common mechanism affecting cell survival in both zebrafish and human after alteration of the FA network. Given the fundamental similarity of the cellular mechanisms of the FA pathway in zebrafish and humans, the screening of small molecule libraries for compounds that can rescue the sex-reversal phenotype of zebrafish fancl mutants might identify compounds of therapeutic importance for Fanconi Anemia patients.
Our analysis of zebrafish fancl mutants suggests a model in which oocyte survival regulated by Tp53-mediated apoptosis is a central element that can tip gonad fate towards the male or the female pathway (gradient red box in Figure 8). Zebrafish develop initially as juvenile hermaphrodites, and have immature ovaries during the juvenile stage regardless of their definitive sex [26]–[28]. This immature ovary is bipotential, and expresses both female (cyp19a1a) and male (amh) specific markers (Figure 8A) [29], [31], [48]. During the fate decision period, some wild-type animals up-regulate cyp19a1a and suppress amh expression (Figure 8B) thereby tipping the fate of the gonad towards the female ovarian pathway (Figure 8C). Complementarily, other wild-type individuals suppress cyp19a1a and up-regulate amh expression (Figure 8D) and gonad fate tips towards the male testis pathway (Figure 8E). In this work, we show that oocyte survival is crucial to maintain the female gene expression profile of somatic cells that is essential for ovary development.
In wild-type zebrafish, juvenile bipotential gonads contain immature oocytes at early stage IB ([49]; and this work). In transitional stages, gonads that become ovaries possess oocytes that progress through meiosis to late stage IB and reach diplotene, where they arrest for the remainder of oocyte development [49]. In fancl−/− homozygous mutants, loss of oocytes at or before diplotene likely alters signaling from germ line to the soma, leading to loss of cyp19a1a expression, failure to down regulate amh expression, and consequent masculinization of the gonads to form testes (Figure 8G). The cyp19a1a gene encodes aromatase, the enzyme that converts testosterone to estrogen. It is known that aromatase is critical for female fate in zebrafish because pharmacological treatments with the aromatase inhibitor fadrozole masculinizes gonads [71]–[73] and because, complementarily, treatments with estrogen (estradiol) down-regulate amh expression and feminize the gonad [74]. We hypothesize that the apoptotic loss of oocytes in fancl mutants causes cyp19a1a gene expression to disappear and leads to the failure to maintain aromatase levels, which results in failure to produce and sustain high estrogen levels in the gonad, causing gonads to abandon the female fate and instead, enter the testis developmental program.
The presence of oocytes appears to be important for sex determination not only for zebrafish, but also for medaka. In contrast to zebrafish, in which all individuals begin oogenesis, in medaka only XX females start oogenesis while XY males suppress oogenesis and all germ cells remain undifferentiated (reviewed in [75]). A feature common to both species is that the number of developing oocytes is a key feature that tips undifferentiated gonads towards an ovary fate ([31], [75] and this work). In medaka, the partial removal of PGCs can reduce the number of developing oocytes below a threshold necessary for female development [76]. In addition, medaka hotei mutants, which have aberrant oocyte development [77], fail to maintain cyp19a1a expression and gonads develop into testes. Therefore, the survival of developing oocytes appears to be important for sex determination in both zebrafish and medaka. These considerations support the hypothesis that when the number of oocytes exceeds a threshold, sexual fate tips towards the female pathway, and alternatively, when the oocyte number fails to exceed that threshold, the sexual fate tips towards the male pathway, as we observed in zebrafish fancl mutants.
In zebrafish, presumptive juvenile males had more TUNEL signal in germ cells than presumptive females had suggesting the hypothesis that oocyte apoptosis could be the mechanism of testicular and ovarian differentiation in zebrafish [27]. Consistent with this hypothesis, analysis of ziwi null mutants showed that total loss of germ cells by apoptosis caused ziwi mutants to develop exclusively as sterile males [34]. Our results show that Tp53-mediated germ cell apoptosis is a mechanism that can tip gonad fate towards the female or male pathway, at least in fancl mutants. Because environmental factors such as high temperature (Figure 8H) or endocrine-disrupting chemical treatments can also increase oocyte apoptosis and cause sex reversal [71]–[73], it is plausible to suggest that the integration of genetic and environmental factors converge to modify the levels of Tp53-mediated germ cell apoptosis, which affect oocyte survival during the critical time window to determine the sexual fate of the gonad, and ultimately alter zebrafish sex determination.
Animals were handled in accordance with good animal practice as defined by relevant animal welfare bodies, and the University of Oregon Institutional Animal Care and Use Committee approved all animal work (Animal Welfare Assurance Number A-3009-01, IACUC protocol #08-13).
The zebrafish fancl mutation (HG10A; GenBank accession AB353980) was generated by insertional mutagenesis by Tol2 transposon-mediated enhancer trap [42]. The tp53 mutant line tp53zdf1 causing the amino acid substitution M214K was obtained from ZIRC (http://zebrafish.org/zirc/home/guide.php) [55]. Genotyping of tp53 animals was performed as described [55]. Genetic nomenclature follows guidelines from ZFIN (http://zfin.org/zf_info/nomen.html).
The full-length zebrafish fancl cDNA was previously described [44] (GenBank accession AY968598). Primer pairs used to amplify the fancl wild-type or mutant alleles were: WT_F:CTGGTCTTTATTGACTGTAATGGC; WT_R:TAGATAAGCTCCAGATTTGGCTTG; Mutant_F:GTCAGCCCATCCAGATCAGCAG; Mutant_R:CATGACGTCACTTCCAAAGGACC. PCR conditions were: 5′94°C; 32 cycles of: 30″94°C, 30″55°C, 1′72°C; followed by 10′72°C. Sizes of PCR-amplified bands: Wild type: 479 bp; Mutant: 370 bp.
Total RNA isolation from dissected adult testes and cDNA synthesis were performed as described [41]. Primers used for reverse transcriptase-PCR (RT-PCR) experiments were: F1:GACGGCTTCATCACAGTGCTG; R1:CATGACGTCACTTCCAAAGGACC; F2:GAACCCTGACTGCACTGTCCTAC; R2:GCTTTGGCGACTGGTTGGCAGAC. PCR conditions were: F1-R1: 3′94°C; 40 cycles of: 30″94°C, 30″58°C, 1′30″72°C; followed by 10′72°C; F2-R2: 3′94°C; 37 cycles of: 20″94°C, 30″60°C, 45″72°C; followed by 10′72°C. Sizes of PCR-amplified bands: F1-R1: 1239 bp F2-R2: 232 bp.
To obtain animals lacking germ cells, wild-type zebrafish embryos from the AB strain were injected at the 1–2 cell stage with antisense morpholino oligonucleotide (Gene Tools, Oregon) directed against dead end as described [46]. Sibling non-injected embryos and a fraction of dnd MO-injected embryos were fixed at 24 hours post-fertilization to confirm the presence or absence of germ cells by whole-mount in situ hybridization using vasa probe as described [47].
Animals were reared and collected under standard conditions [78]. In situ hybridization experiments on zebrafish cryosections were performed as described [29]. Adjacent sections of gonads were obtained by placing three consecutive sections of the gonad on three different slides. Probes for amh and cyp19a1a were made as described [29] and probe for vasa was made from its 3′end as described [47]. A fancl cDNA fragment of 786 nt containing the PHD domain (nucleotides 646-1431 of AY968598) was cloned in TOPO vector (Invitrogen) and used to synthesize DIG-labeled riboprobe (Boehringer Mannheim). For gonad histology, euthanized animals were fixed in Bouin's fixative for about 24–48 hours and washed repeatedly in 70% ethanol. Animals were dehydrated and embedded in paraffin, sectioned at 7 microns, and stained with hematoxylin and eosin.
Animals were fixed at 25 dpf in 4% PFA ON at 4°C, dehydrated, embedded in paraffin, and sectioned at 7 microns. Apoptotic cells were detected by immuno-fluorescence using anti-active Caspase-3 as primary antibody (1∶200, BD Pharmingen) and Alexa-Fluor594 goat anti-rabbit as secondary antibody (1∶1000, Invitrogen) following an immuno-histochemical protocol (S. Cheesman, personal communication). Gonads were screened for positive signal by DIC-fluorescence microscopy. The number of positive cells in gonads of fancl and wild-type animals was scored in 840 sections: 70 sections containing gonads per each animal (n = 12).
|
10.1371/journal.pbio.1001946 | Fission Yeast Pxd1 Promotes Proper DNA Repair by Activating Rad16XPF and Inhibiting Dna2 | Structure-specific nucleases play crucial roles in many DNA repair pathways. They must be precisely controlled to ensure optimal repair outcomes; however, mechanisms of their regulation are not fully understood. Here, we report a fission yeast protein, Pxd1, that binds to and regulates two structure-specific nucleases: Rad16XPF-Swi10ERCC1 and Dna2-Cdc24. Strikingly, Pxd1 influences the activities of these two nucleases in opposite ways: It activates the 3′ endonuclease activity of Rad16-Swi10 but inhibits the RPA-mediated activation of the 5′ endonuclease activity of Dna2. Pxd1 is required for Rad16-Swi10 to function in single-strand annealing, mating-type switching, and the removal of Top1-DNA adducts. Meanwhile, Pxd1 attenuates DNA end resection mediated by the Rqh1-Dna2 pathway. Disabling the Dna2-inhibitory activity of Pxd1 results in enhanced use of a break-distal repeat sequence in single-strand annealing and a greater loss of genetic information. We propose that Pxd1 promotes proper DNA repair by differentially regulating two structure-specific nucleases.
| Genome stability maintenance relies on DNA repair enzymes, among which are structure-specific nucleases that cleave DNA in a sequence-independent but structure-dependent manner. It is important to understand how the activities of such nucleases are controlled, because either insufficient or excessive cleavage of DNA could jeopardize genome integrity. In this study, we discovered a new regulator of two different structure-specific nucleases in the fission yeast Schizosaccharomyces pombe. The identified protein, which we named Pxd1, promotes the activity of the 3′ endonuclease Rad16, but restrains the activity of the 5′ endonuclease Dna2. In the absence of Pxd1, several Rad16-dependent DNA repair processes become defective. One of these processes is a DNA-repeat–mediated double-strand break repair pathway called single-strand annealing, which causes genomic deletions. When the Dna2-inhibitory activity of Pxd1 is impaired, Dna2-dependent end processing of double-strand breaks is enhanced and a more extensive deletion occurs during single-strand annealing. Thus, Pxd1 facilitates a potentially dangerous DNA repair process, but in the meantime minimizes its deleterious consequences. We propose that a dual-target regulator like Pxd1 is ideally suited for coordinating multiple enzymatic activities during DNA repair.
| Structure-specific DNA nucleases contribute to the maintenance of genome stability by processing DNA secondary structures during DNA replication and repair [1],[2]. The activities of these nucleases must be tightly controlled to prevent unintended cleavage; however, the molecular mechanisms underlying the regulation of these nucleases have not been fully elucidated.
The roles of several structure-specific nucleases in DNA repair are best understood in the single-strand annealing (SSA) pathway of DNA double-strand break (DSB) repair. SSA is a repair pathway for DSBs occurring between repeat sequences and has been most thoroughly studied in the budding yeast Saccharomyces cerevisiae [3]. SSA relies on the DNA resection process to generate 3′-ended single-stranded DNA (ssDNA) extending from the break to the repeat sequences [4]. Such long-range resection is mediated by two structure-specific nucleases, Exo1 and Dna2, which act in parallel to each other [5]. Upon annealing of the ssDNA of the repeat sequences, the intervening sequence between the repeats, which now becomes 3′ nonhomologous ssDNA tails, is removed by a nuclease complex Rad1-Rad10 in budding yeast (XPF-ERCC1 in mammals and Rad16-Swi10 in the fission yeast Schizosaccharomyces pombe) [6].
The function of Rad1-Rad10 in SSA requires two positive regulators, Saw1 and Slx4 [7]–[10]. Saw1 recruits Rad1-Rad10 to the DNA substrate during SSA [8],[11]; however, the exact role of Slx4 in SSA is not clear. Furthermore, it is not known whether the activities of the resection nucleases are regulated during SSA.
Here we show that a novel factor Pxd1 is a key regulator of SSA in fission yeast. It interacts with both the nonhomologous ssDNA cleavage nuclease Rad16XPF and the resection nuclease Dna2, thus influencing different aspects of SSA. Interestingly, Pxd1 regulates these two structure-specific nucleases in opposite ways: it promotes the completion of SSA by activating the nuclease activity of Rad16, while it minimizes genetic information loss by inhibiting RPA-mediated Dna2 activation.
A previously uncharacterized fission yeast protein, SPBC409.16c, has been predicted by PomBase as the ortholog of budding yeast Saw1 [12]. In budding yeast, Saw1 interacts with the Rad1-Rad10 nuclease [8],[11]. In an affinity purification coupled with mass spectrometry (AP-MS) experiment, we found that Rad16 and Swi10, the fission yeast counterparts of budding yeast Rad1 and Rad10, respectively [13],[14], co-purified with SPBC409.16c (Figure 1A), thus corroborating the PomBase orthology prediction. We will hereafter refer to SPBC409.16c as Saw1.
Intriguingly, Dna2, Cdc24, and an uncharacterized protein SPCC1322.02 also co-purified with Saw1 (Figure 1A). Dna2 and the fission-yeast-unique protein Cdc24 are known to form a heterodimer and are both required for Okazaki fragment maturation in fission yeast [15]. When SPCC1322.02 was used as bait for AP-MS analysis, the same six proteins were again isolated together (Figure 1B), suggesting that Rad16-Swi10-Saw1, Dna2-Cdc24, and SPCC1322.02 co-exist in a protein complex, which we named the PXD (pombe XPF and Dna2) complex. Accordingly, we named SPCC1322.02 Pxd1.
Pxd1 is annotated by PomBase as a “sequence orphan” with no apparent orthologs outside of the fission yeast clade, and it does not contain any known domains. To identify the regions of Pxd1 that participate in its interactions with Rad16-Swi10 and Dna2-Cdc24, we performed truncation analysis and found that its interaction with Rad16-Swi10 is mediated by the middle region of Pxd1 (residues 101–233), whereas its interaction with Dna2-Cdc24 is mediated by the C-terminal region of Pxd1 (residues 227–351) (Figure 1C).
Because distinct regions of Pxd1 mediate its interactions with Rad16-Swi10 and Dna2-Cdc24, we hypothesized that Pxd1 may act as a scaffold to bring these two nucleases together. We tested this idea by examining the association of the two nucleases in wild-type and pxd1Δ backgrounds. Cdc24 co-immunoprecipitated with Rad16 in the wild type, but this interaction was abolished in pxd1Δ (Figure 1D). Similarly, the interaction between Saw1 and Cdc24 was abolished in pxd1Δ (Figure 1E). These results suggest that, within the PXD complex, Pxd1 acts as a physical link between the Rad16-Swi10-Saw1 and Dna2-Cdc24 subcomplexes (Figure 1F).
To determine where Pxd1 binds on its binding partners, we performed yeast two-hybrid (Y2H) assay, immunoprecipitation using truncated proteins, and cross-linking mass spectrometry (CXMS) (Figure S1). Rad16, Dna2 and Cdc24, but not Swi10, exhibited positive Y2H interactions with Pxd1. An N-terminal fragment of Rad16 (residues 1–451), which contains a helicase-like domain, was sufficient to co-immunoprecipitate Pxd1 in the absence of Swi10. CXMS analysis of a Dna2-Cdc24-Pxd1(227–351) complex detected cross-links between the K148 residue of Cdc24 and two different residues of Pxd1 (K276 and K351). Consistently, Cdc24(80–245), which contains the K148 residue, is the smallest fragment of Cdc24 that could robustly co-immunoprecipitate Pxd1.
To understand the function of Pxd1, we generated a pxd1 deletion mutant, which exhibited no growth defect (Figure 2A). Thus, Pxd1 is unlikely to be important for the replication function of Dna2-Cdc24, which is essential for viability. We then examined the DNA damage sensitivity of deletion mutants of pxd1 and related nonessential genes. pxd1Δ showed mild sensitivity to ionizing radiation (IR) but displayed no obvious sensitivity to UV, methyl methanesulfonate (MMS), camptothecin (CPT), or hydroxyurea (HU) (Figure 2A). Consistent with the known role of Rad16-Swi10 in nucleotide excision repair (NER), rad16Δ and swi10Δ showed severe sensitivity to UV that was at a level similar to the mutant lacking another NER factor, Rhp14XPA (Figure 2A). These three mutants also showed similar sensitivity to MMS and HU. However, rad16Δ and swi10Δ were more sensitive to IR than rhp14Δ, which most likely reflected the non-NER functions of Rad16-Swi10, such as the removal of the 3′ nonhomologous ssDNA tails during homologous recombination (HR) repair [16],[17]. Surprisingly, saw1Δ displayed no sensitivity to any treatment (Figure 2A). In addition, deletion of saw1 did not enhance the DNA damage sensitivity of pxd1Δ (Figure 2B).
To test the epistatic relationship between pxd1Δ, rhp14Δ, and rad16Δ, we examined the sensitivity of their single, double, and triple mutants (Figure 2C). Deletion of pxd1, rhp14, or both in rad16Δ did not enhance the IR sensitivity. In contrast, the pxd1Δ rhp14Δ double mutant showed greater IR sensitivity than either single mutant, reaching a level similar to that of rad16Δ. These results suggest that Pxd1 acts with Rad16-Swi10 in the non-NER repair of IR-induced DNA damage.
To further delineate the role of Pxd1 in non-NER repair, we examined whether Pxd1 functions with Rad16-Swi10 in SSA. We constructed a strain in which an HO endonuclease-induced DSB is flanked by two direct repeats (Figure 3A). In such a system, homologous recombination between the two repeats may proceed through either the SSA or BIR mechanisms, but because the two repeats are only about 6 kb apart, SSA is expected to be the predominant pathway [18]. Regardless of which mechanism is used, two 3′ nonhomologous ssDNA tails, one 6,328 nt long and the other 29 nt long, must be removed by a nuclease such as Rad16-Swi10, resulting in the loss of the HO cleavage site and a leu1+ marker (Figure 3A). For simplicity, we will hereafter refer to this repair process as SSA.
When wild-type cells harboring the SSA system were shifted from an HO repression (+ thiamine) to an HO induction condition (−thiamine) in liquid media, no obvious growth arrest was observed, but the cells became Leu− (Figure S2A and B), indicating that SSA repair was highly efficient. In contrast, when HO was induced in rad16Δ and swi10Δ cells, their proliferation was retarded for approximately 20 h, suggesting a delay of the repair process (Figure S2A). Eventually most of the rad16Δ and swi10Δ cells survived and became Leu−, most likely due to backup nuclease activities (Figure S2B). On thiamine-free solid media, the repair defect of rad16Δ and swi10Δ also manifested as a growth delay (Figure 3B). pxd1Δ cells showed the same growth delay as rad16Δ and swi10Δ cells (Figure 3B). In addition, the double mutants rad16Δ pxd1Δ and swi10Δ pxd1Δ exhibited the same phenotype as the three single mutants, indicating that Rad16-Swi10 and Pxd1 function in the same process. In this assay, saw1Δ again behaved like the wild type. Moreover, deleting saw1 in pxd1Δ did not exacerbate the phenotype. Thus, unlike its budding yeast ortholog, fission yeast Saw1 does not appear to be important for SSA.
To more directly monitor SSA, we examined the elimination of the intervening DNA sequence between the repeats using qPCR (Figure 3C). The rate of DNA elimination in the pxd1Δ and swi10Δ mutants was significantly slower than in the wild type and the saw1Δ mutant (Figure 3C). In addition, we visualized Rad52 nuclear foci, which is an indication of ongoing DNA repair activity. In the wild-type and saw1Δ cells, the level of Rad52 foci transiently increased after HO induction but returned to the pre-induction level within 8 h (Figure S2C and D). In contrast, in pxd1Δ, rad16Δ, and swi10Δ cells, HO-induced Rad52 foci remained at a high level for more than 10 h. Thus, DNA repair in these three mutants failed to efficiently proceed to completion.
To test whether the interaction between Pxd1 and Rad16-Swi10 is required for SSA, we examined cells expressing truncated versions of Pxd1. Pxd1 missing either its N-terminal region or C-terminal region could rescue the defect of pxd1Δ, whereas Pxd1 without the middle region failed to rescue the phenotype (Figure 3D). Thus, the region of Pxd1 involved in Rad16-Swi10 binding is required for SSA.
During SSA, the role of Rad16-Swi10 is to remove the 3′ nonhomologous ssDNA tails. Given that the interaction between Pxd1 and Rad16-Swi10 is required for SSA, we hypothesized that Pxd1 is involved in the same step. To test this idea, we monitored the level of 3′ ssDNA using a qPCR assay. In this assay, the PCR template was genomic DNA pre-digested with a restriction enzyme, BstUI, that cuts double-stranded but not single-stranded DNA. Thus, the level of the PCR product reflects the amount of ssDNA (Figure 3E). In wild-type and saw1Δ cells, only a transient and small increase (approximately 10%) of ssDNA occurred after HO induction (Figure 3F). In contrast, in pxd1Δ and swi10Δ cells, ssDNA accumulated to a much higher level and persisted (Figure 3F). Thus, 3′ ssDNA removal is defective in pxd1Δ and swi10Δ, but not in saw1Δ, mutants.
Rad16 (also known as Swi9) and Swi10 are required for mating-type switching, presumably due to their involvement in resolving recombination intermediates of the HR process triggered by the programmed DSB at the mating type locus (Figure S3A) [19],[20]. To test whether Pxd1 also participates in mating-type switching, we performed an iodine-staining assay on h90 homothallic strains growing on a medium compatible with mating and sporulation (Figure S3B). Dark staining indicates efficient mating-type switching, whereas light or sectored staining indicates defects in mating-type switching. Wild-type and saw1Δ h90 colonies were darkly and homogenously stained (Figure 4A). In contrast, rad16Δ and pxd1Δ colonies showed much weaker and uneven staining patterns. This result suggests that pxd1Δ, like rad16Δ, is defective in mating-type switching. Consistent with the idea that a failure of the HR process underlies the mating-type switching defect of rad16Δ and pxd1Δ, we observed using ChIP-seq that, in heterothallic h− cells, Rad52 accumulated more strongly at the mating type locus in rad16Δ and pxd1Δ than in wild-type cells (Figure 4B). In h− cells, the programmed DSB also triggers an HR process, but the mating type does not switch because only one type of donor sequence is available.
When different truncated forms of Pxd1 were tested for their abilities to rescue the mating-type switching defect, the middle region-deleted version of Pxd1 failed to rescue the iodine-staining phenotype of pxd1Δ h90 colonies, suggesting that the interaction between Pxd1 and Rad16-Swi10 is important for mating-type switching (Figure 4C).
Covalent Top1-DNA adducts, referred to as Top1 cleavage complexes (Top1cc), arise spontaneously and can jeopardize cell survival if not removed. It was shown recently that Rad16-Swi10 and Tdp1 redundantly remove Top1cc in fission yeast [21]. We, therefore, tested whether Pxd1 also contributes to this process. Tetrad analysis showed that, like swi10Δ, pxd1Δ is synthetic lethal/sick with tdp1Δ, and the synthetic lethality/sickness can be rescued by the deletion of top1 (Figure 4D and Figure S4A). Further analysis showed that the C-terminally truncated version, but not the middle region-deleted version, of Pxd1 could rescue the synthetic lethality/sickness (Figure 4E and Figure S4B). These results suggest that Pxd1 acts with Rad16-Swi10 in the removal of Top1cc (Figure S4C).
To understand how Pxd1 acts with Rad16-Swi10, we tested whether its absence affects the nuclease activity of Rad16-Swi10 purified from fission yeast cells. For a positive control, we used a strain expressing C-terminally truncated Pxd1 as the only form of Pxd1, so that Dna2-Cdc24, which also has nuclease activities, does not co-purify with Rad16-Swi10. As described earlier, this truncated form of Pxd1 is sufficient for SSA, mating-type switching, and Top1cc removal. Consistent with the known substrate specificity of XPF family nucleases, Rad16 immunoprecipitated from such a strain showed robust nuclease activity toward 3′ overhang DNA and Y fork DNA but not 5′ overhang DNA (Figure S5A). The nuclease-dead mutant Rad16-D700A immunoprecipitated from the same Pxd1 C-terminal truncation background did not show nuclease activity toward any substrates, demonstrating that the nuclease activity we observed was Rad16-specific (Figure S5A). Rad16 immunoprecipitated from pxd1Δ cells had much weaker nuclease activity than the positive control (Figure 5A and Figure S5B). The expression level and stability of Rad16 were not affected by the loss of Pxd1 (Figure S5C). Thus, Pxd1 is required for a robust nuclease activity of Rad16-Swi10.
The middle region of Pxd1 is required for its interaction with Rad16-Swi10 and is needed for SSA, mating-type switching, and the removal of Top1cc. To identify functionally important residues within this region, we mutated the residues conserved between Pxd1 and its homologs in two other fission yeast species and found that a double point mutation, A155D/E172A, significantly weakened the interaction between a recombinant Pxd1 protein purified from E. coli and Rad16-Swi10 immunoprecipitated from pxd1Δ fission yeast cells (Figure 5B). When introduced into the pxd1 gene in fission yeast, this mutation impaired 3′ ssDNA removal during SSA (Figure 5C) and diminished the nuclease activity of Rad16-Swi10 purified from the Pxd1 C-terminal truncation background (Figure 5D). These data strongly suggest that the interaction between Pxd1 and Rad16-Swi10 is needed for Pxd1 to activate Rad16-Swi10.
When we added Pxd1 protein purified from E. coli to Rad16-Swi10 immunoprecipitated from pxd1Δ cells, we observed a dose-dependent enhancement of nuclease activity (Figure 5E). As a control, the A155D/E172A mutant form of Pxd1 purified from E. coli failed to activate the nuclease activity (Figure 5E). Thus, recombinant Pxd1 is sufficient for activating Rad16-Swi10.
To probe the role of the interaction between Pxd1 and Dna2-Cdc24, we overexpressed a Pxd1 C-terminal fragment, Pxd1(227–351), which encompasses the Dna2-Cdc24–interacting region. Remarkably, Pxd1(227–351) overexpression caused severe growth defect, and this defect could be suppressed by co-overexpression of both Dna2 and Cdc24, or Dna2 alone (Figure 6A). Two mutant alleles of the gene encoding the DNA helicase Pfh1 (Pif1 homolog), pfh1-R20 and pfh1-R23, which are suppressors of temperature-sensitive mutants of dna2 and cdc24 [22],[23], also suppressed the growth defect caused by Pxd1(227–351) overexpression (Figure S6A). Thus, the growth defect is likely due to a down-regulation of the functions of Dna2-Cdc24. To determine whether the interaction between Pxd1 and Dna2-Cdc24 is important for this down-regulation, we performed mutagenesis on the C-terminal region of Pxd1 and found that simultaneously mutating five residues conserved between Pxd1 and its homologs in two other fission yeast species, referred to as the 5A mutation, weakened the interaction between Pxd1 and Dna2-Cdc24 (Figure S6B). The overexpression of Pxd1(227–351)-5A did not cause any growth defect (Figure 6B), indicating that the Pxd1(227–351) overexpression phenotype is mediated by an interaction with Dna2-Cdc24.
To understand how Pxd1(227–351) down-regulates the functions of Dna2-Cdc24 when overexpressed, we investigated whether in vitro it influences the nuclease activity of Dna2-Cdc24. We found that Dna2 and Cdc24 co-overexpressed and purified from pxd1Δ cells were able to cleave a 5′ overhang DNA substrate (Figure 6C). The stability of Dna2 and Cdc24 was not affected by pxd1Δ (Figure S6C). Consistent with the results obtained with budding yeast and human Dna2 [24],[25], the addition of RPA markedly stimulated the nuclease activity of Dna2. Recombinant Pxd1(227–351) purified from E. coli did not affect the basal activity of Dna2; however, it significantly weakened the activation effect of RPA (Figure 6C). Pxd1(227–351)-5A failed to inhibit the RPA-mediated activation of Dna2 (Figure 6D). Thus, the interaction between Pxd1 and Dna2 impedes the activation of Dna2 by RPA.
RPA can enhance the nuclease activity of Dna2 by promoting the binding of Dna2 on ssDNA in budding yeast [24]; therefore, we hypothesized that Pxd1(227–351) may block RPA-mediated Dna2 binding to DNA substrates. To test this idea, we first investigated the ability of Pxd1 and Dna2-Cdc24 to bind a 5′ overhang DNA using a gel mobility shift assay. In this assay, DNA cleavage was prevented by using a buffer containing 1 mM EDTA and no divalent cations. Dna2-Cdc24 shifted the mobility of the DNA, whereas Pxd1(227–351) had no effect (Figure 6E, lanes 2–5). The addition of Pxd1(227–351) with Dna2-Cdc24 led to the formation of a complex that migrated faster than the Dna2-Cdc24-DNA complex (Figure 6E, lanes 6–8 and Figure 6F, lanes 3–5), most likely due to a higher negative charge of the Pxd1-Dna2-Cdc24-DNA complex because the recombinant Pxd1(227–351) has a low PI of 5.09. As a control, the addition of Pxd1(227–351)-5A, which cannot efficiently interact with Dna2-Cdc24, had much weaker ability to shift the Dna2-Cdc24-DNA complex (Figure 6F, lanes 6–8). These results show that, consistent with the lack of effect of Pxd1 on the basal nuclease activity of Dna2, Pxd1 does not appear to affect the ability of Dna2-Cdc24 to bind naked DNA.
When RPA was added to the DNA binding reaction with Dna2-Cdc24, a Dna2-Cdc24-RPA-DNA complex that migrated slower than the Dna2-Cdc24-DNA complex and the RPA-DNA complex was detected (Figure 6E, lanes 10–12). Addition of Pxd1(227–351) interfered with the formation of this higher-order complex and resulted in a form of DNA that appeared to be bound by only RPA (Figure 6E, lanes 14–16 and Figure 6F, lanes 11–13), suggesting that Dna2-Cdc24 was dissociated from the RPA-DNA complex in the presence of Pxd1. In comparison, Pxd1(227–351)-5A was weaker in its ability to disrupt the higher-order complex (Figure 6F, lanes 14–16). From these results, we conclude that Pxd1 inhibits the RPA-mediated activation of Dna2 by blocking the binding of Dna2-Cdc24 to RPA-coated DNA.
The Dna2-inhibitory effect of Pxd1 may influence the actions of Dna2 in either DNA replication or DSB resection. Because Pxd1 is down-regulated during the S phase of the cell cycle (our unpublished observation), we hypothesized that it may mainly regulate the resection function of Dna2. During resection, Dna2 is expected to act with Rqh1, a RecQ family helicase, in a pathway parallel to Exo1 [5]; therefore, in an exo1Δ background, the residual resection activity should be Rqh1- and Dna2-dependent. Using a qPCR-based assay to monitor resection from an irreparable HO-induced DSB (Figure 7A), we found that, as reported [26], the deletion of exo1, but not rqh1, strongly reduced long-range resection (Figure 7B). No obvious difference was found between pxd1Δ and the wild type. However, deletion of pxd1 in exo1Δ partially rescued the resection defect. Thus, consistent with the results of Pxd1(227–351) overexpression and the in vitro nuclease assay, Pxd1 appears to attenuate the Dna2- and Rqh1-mediated resection activity, at least in the exo1Δ background. Supporting this idea, the deletion of pxd1 did not rescue the DNA resection defect of rqh1Δ exo1Δ cells (Figure 7C). The DNA damage sensitivity of exo1Δ cells was not rescued by pxd1Δ (Figure S6D), probably due to Exo1 also playing nonresection roles in genome maintenance.
To determine which region of Pxd1 is involved in resection inhibition, we examined the effect of introducing truncated versions of Pxd1 into an exo1Δ pxd1Δ double mutant. The N-terminal–truncated and middle-region–deleted versions curtailed long-range DNA resection as strongly as the full-length Pxd1. In contrast, a C-terminally truncated version, Pxd1-Δ (302–348), which is defective in binding Dna2, failed to impede resection (Figure 7D). These results suggest that the interaction between Pxd1 and Dna2 is required for the inhibitory effect of Pxd1 on DNA resection. In addition, the C-terminal region of Pxd1 alone can inhibit DNA resection in the exo1Δ pxd1Δ background (Figure 7D).
During the SSA repair process, DNA resection is required for rendering the homologous repeats single-stranded [4],[18]. We hypothesized that the C-terminal region of Pxd1 may regulate the homologous partner choice during SSA repair when there are multiple homologous sequences on the same side of the DSB [4],[27]. To test this idea, we constructed an SSA competition system. In this system, one additional homologous sequence was inserted between the two repeats in the original SSA strain (Figure 7E). During SSA repair, the repeat sequence on the left side of the HO site can anneal with either potential homologous partner on the right side of the HO site. If partner1 is used, the postrepair cells will remain Leu+; however, if partner2 is used, cells will become Leu− and suffer a greater loss of genetic information (Figure 7E). We found that the DSB-proximal homologous sequence, partner1, was more frequently used in exo1Δ than in wild-type cells (Figure 7F), presumably because slower resection in exo1Δ cells reduces the chance of partner2 becoming single-stranded before a productive repair using partner1 has occurred. Removing the Dna2-inhibitory region of Pxd1 reversed the effect caused by exo1 deletion (Figure 7F), consistent with the rescue of the resection defect observed using the irreparable HO system. Interestingly, in an exo1+ background, the same Pxd1 truncation enhanced the use of the distal homologous sequence, partner2 (Figure 7F). These results suggest that Pxd1 restricts the use of break-distal homologous sequences during SSA repair to prevent excessive loss of genetic information.
In this study, we identified a novel fission yeast protein, Pxd1, which interacts with two structure-specific nucleases, Rad16-Swi10 and Dna2-Cdc24. Our data indicate that Pxd1 can activate the 3′ nuclease activity of Rad16-Swi10, but inhibit the RPA-mediated activation of the 5′ nuclease activity of Dna2-Cdc24. These two capacities of Pxd1 allow it to promote SSA and, at the same time, reduce the negative impact of SSA on genome integrity (Figure 7G).
Unlike the situations in budding yeast, in fission yeast, neither saw1Δ nor slx4Δ has an observable SSA defect (Figure 3 and Figure S5D). Among the two functionally important features of S. cerevisiae Saw1 [11], the R19 residue required for Rad1 binding is conserved in S. pombe Saw1, whereas the C-terminal positive amino acid stretch required for DNA binding is missing in S. pombe Saw1 (Figure S7). We suspect that S. pombe Saw1 may have lost its SSA-related function or become redundant.
Compared with Slx4 proteins in S. cerevisiae and metazoans, S. pombe Slx4 is much shorter and appears to have lost the region required for the interaction with XPF-ERCC1 [28]. On the other hand, the middle region of Pxd1 (residues 101–233), which mediates Rad16 binding, seems to possess sequence similarity to the XPF-binding region of metazoan Slx4 proteins, which has been referred to as the MLR (MEI9XPF-interaction-Like Region) (Figure S5E) [28]–[30]. Thus, we speculate that during evolution, in the lineage leading to the fission yeast, the ancestor Slx4 protein may have split into two proteins, one becoming Pxd1 and the other evolving into the current-day S. pombe Slx4, which is solely involved in the regulation of the Slx1 nuclease [31].
In budding yeast, CDK1-mediated phosphorylation promotes the resection function of Dna2 [32]. Here we show that the resection activity of fission yeast Dna2 is subject to a negative regulation by Pxd1. Thus, Dna2 appears to be a regulatory target used in diverse organisms for controlling the resection process. Intriguingly, pxd1 C-terminal truncation caused an overt phenotype in the SSA competition assay, but pxd1 deletion did not alter resection in the irreparable HO system, suggesting the possibility that the resection process may be regulated differently depending on whether strand annealing with a homologous partner has occurred.
Highly repetitive DNA elements, such as retrotransposons in yeasts and Alu elements in humans, mediate chromosome rearrangements through homologous recombination pathways including SSA [33]–[35]. The results of our SSA competition assay suggest that fine-tuning the resection activities may be a strategy that evolution has exploited to ameliorate the deleterious consequences of repeat-mediated recombination.
Are there evolutionary advantages of using one protein to exert opposite controls on two nucleases? One possibility is that Pxd1 may serve as a hub to integrate regulatory signals so that the up-regulation of one nuclease and the down-regulation of the other can be more precisely coordinated. The expression level of Pxd1 appears to decrease in S phase (our unpublished observation), suggesting that cell cycle control of these two nucleases is imposed through Pxd1. Thus, the activity of Dna2 is relieved from inhibition during S phase when it is needed for DNA replication. On the other hand, given that the activation of Rad16 by Pxd1 is important for removing the 3′ nonhomologous ssDNA, the decrease of Pxd1 during S phase may curtail HR repair events involving nonhomologous ssDNA. Further analysis will be needed to assess to what extent such a regulation affects DNA repair pathway choices.
The fission yeast strains used in this study are listed in Table S1, and plasmids used in this study are listed in Table S2. Genetic methods for strain construction and the composition of media are as described [36]. To construct an SSA system based on a strain in which an HO cleavage site is inserted at the arg3 locus [37],[38], we first cloned a 1.2-kb sequence immediately upstream of the arg3 ORF between the EcoRI and ClaI sites in the integrating vector pJK148 [39], resulting in plasmid pDB169. Then, a 0.6-kb sequence corresponding to cmb1 ORF, which is immediately downstream of arg3, was cloned into the BamHI site in pDB169, resulting in plasmid pDB174. A 0.3-kb sequence from the intergenic region between arg3 and cmb1 was cloned between the NotI and SacII sites in pDB174, resulting in plasmid pDB176. Integration of XbaI-cut pDB176 into the HO strain DY1012 resulted in the SSA strain DY2392. For monitoring the ssDNA tail removal, a BstUI restriction site was introduced into pDB176, resulting in plasmid pDB459. Integration of pDB459 into the HO strain DY4840 resulted in the SSA strain DY5999. To create the SSA competition system, a 400-bp sequence immediately upstream of the arg3 ORF was inserted into the AatII site in pDB176, resulting in plasmid pDB1637, which was then integrated into an HO strain. Protein overexpression in S. pombe was conducted using pDUAL vectors containing the strong nmt1 promoter [40],[41].
The lysate from 50 OD600 units of cells was prepared by glass bead beating in lysis buffer A (50 mM Tris-HCl, pH 8.0, 0.1 M NaCl, 10% glycerol, 0.05% NP-40, 1 mM PMSF, 1 mM DTT, 1× Roche Protease Inhibitor Cocktail). TAP-tagged and YFP-tagged proteins were immunoprecipitated with IgG Sepharose beads (GE healthcare) and GFP-Trap beads (Chromotek), respectively.
Rad16-YFH and Swi10 were co-overexpressed in an isp6Δ psp3Δ pxd1Δ fission yeast strain. Cells were lysed using a French press in lysis buffer A. YFH-tagged protein was enriched with anti-FLAG M2 affinity gel (Sigma) and eluted with 3× FLAG peptide.
Cdc24-YFH and Dna2 were co-overexpressed and purified as above.
His6-tagged RPA and Pxd1 were expressed in a BL21 E. coli strain. Cells were lysed using a French press in lysis buffer B (50 mM phosphate buffer, pH 8.0, 0.3 M NaCl, 10 mM imidazole, 10% glycerol, 1 mM PMSF), and purification was performed using Ni-NTA-agarose (QIAGEN). The eluate was dialyzed with storage buffer (50 mM Tris-HCl, pH 8.0, 0.1 M NaCl, 10% glycerol, 1 mM DTT) before freezing at −80°C.
For yeast two-hybrid analysis, we used the Matchmaker system (Clontech). Bait plasmids were constructed by inserting cDNAs into a modified pGBKT7 vector. Prey plasmids were constructed by inserting cDNAs into a modified pGAD GH vector. Bait and prey plasmids were co-transformed into the AH109 strain, and transformants were selected on the double dropout medium (SD/–Leu/–Trp). The activation of the HIS3 and ADE2 reporter genes was assessed on the quadruple dropout medium (SD/–Ade/–His/–Leu/–Trp).
Dna2-Cdc24-Pxd1(227–351) complex was prepared by incubating anti-FLAG beads bound by Cdc24-YFH and Dna2 from fission yeast with Pxd1(227–351) from E. coli, washing the beads, and eluting with 3× FLAG peptide. About 12 µg of purified complex in a volume of 20 µl was cross-linked by BS3 or DSS at a final concentration of 0.5 mM for 1 h at room temperature. The reactions were quenched with 20 mM NH4HCO3. Proteins were precipitated with ice-cold acetone, resuspended in 8 M urea, 100 mM Tris, pH 8.5. After trypsin digestion, the LC-MS/MS analysis was performed on an Easy-nLC 1000 UHPLC (Thermo Fisher Scientific) coupled to a Q Exactive-Orbitrap mass spectrometer (Thermo Fisher Scientific). Peptides were loaded on a pre-column (75 µm ID, 8 cm long, packed with ODS-AQ 12 nm–10 µm beads from YMC Co., Ltd.) and separated on an analytical column (75 µm ID, 11 cm long, packed with Luna C18 3 µm 100 Å resin from Phenomenex) using an acetonitrile gradient from 0–25% in 55 min at a flow rate of 200 nl/min. The top 10 most intense precursor ions from each full scan (resolution 70,000) were isolated for HCD MS2 (resolution 17,500; NCE 27) with a dynamic exclusion time of 60 s. Precursors with 1+, 2+, or unassigned charge states were excluded. pLink was used to identified cross-linked peptides with the cutoffs of FDR<5% and E_value<0.001 [42].
For MMS, CPT, and HU sensitivity analysis, five-fold serial dilutions of cells were spotted onto YES with or without the indicated concentration of the chemical. To measure UV sensitivity, after spotting on YES plates, the cells were exposed to the indicated dose of UV treatment. To measure IR sensitivity, the cells were irradiated in microfuge tubes using a Cesium-137 Gammacell 1000 irradiator and then spotted onto YES. The plates were incubated for 2 or 3 d at 30°C.
Genomic DNA was extracted from 3–5 OD600 units of cells collected at different times after HO induction. Five hundred nanograms of genome DNA was digested by 4 U of BstUI for 1.5 h. The amount of amplifiable DNA was determined by qPCR, using the actin gene, act1, as the normalization control. Primer sequences are listed in Table S3.
Genomic DNA was extracted from 3–5 OD600 units of cells collected at different times after HO induction. Five hundred nanograms of genome DNA was digested by 4 U of ApoI for 1.5 h. The amount of amplifiable DNA was determined by qPCR. Primers located at different distances from the HO site were used, and their sequences are listed in Table S3.
The following formula was used to calculate the percentage of DNA that was resected: %resected = (100/2ΔCt−1)/f. ΔCt is the difference in average cycles between digested template and undigested template, and f is the fraction of DNA that has been cut by HO.
Oligo461 (5′-CACGCTACCGAATTCTGACTTGCTAGGACATCTTTGCCCACGTTGACCC-3′) and oligo462 (5′-GTCAGAATTCGGTAGCGTG-3′) were used to prepare the 3′ overhang DNA structure. Oligo461 and oligo463 (5′-GGGTCAACGTGGGCAAAG-3′) were used to prepare the 5′ overhang DNA structure. Oligo461 and oligo464 (5′-TCGATAGTCTCTAGATAGCATGTCCTAGCAAGTCAGAATTCGGTAGCGTG-3′) were used to prepare the Y fork DNA structure. The oligos were annealed in 1× annealing buffer (50 mM Tris-HCl, pH 7.5, 100 mM NaCl). For radiolabeled substrates, oligo461 was radiolabeled at its 5′ end. For reactions analyzed with ethidium bromide (EB) staining, 30 pmol of nonradioactive substrate was used per reaction. For reactions analyzed with autoradiography, 30 pmol of nonradioactive substrate mixed with about 50 fmol of radioactive substrate was used per reaction.
Anti-TAP immunoprecipitates from 50 OD600 units of cells were incubated with substrate in 50 mM Tris-HCl, pH 7.5, 50 mM NaCl, 1 mM MnCl2, 1 mM dithiothreitol, and 0.1 mg/ml bovine serum albumin (BSA) at 30°C for 1 h. The products were separated in 15% denaturing or 10% native gels. The substrates used for denaturing gel analysis were radiolabeled, whereas the substrates for native gel analysis were not radiolabeled. The native PAGE gels were stained with EB, and the denaturing PAGE gels were analyzed by autoradiography.
The reaction mixtures (20 µl) contained 50 mM Tris-HCl, pH 7.5, 50 mM NaCl, 1 mM MgCl2, 1 mM dithiothreitol, 0.1 mg/ml BSA, and 30 pmol of substrate. Reactions were carried out at 30°C for 1 h, and the products were analyzed in a 15% denaturing gel.
The assay mixtures (10 µl) contained 50 mM Tris-HCl, pH 7.5, 1 mM EDTA, 1 mM dithiothreitol, 0.1 mg/ml BSA, 50 mM NaCl, 5% glycerol, and 15 fmol of radioactive 5′ overhang DNA. The assay mixtures were incubated at room temperature for 30 min, and then 2 µl of 6× native loading buffer was added. The products were separated in a 5% PAGE gel in 1× TBE at 3 W for 2 h and analyzed by autoradiography.
|
10.1371/journal.pntd.0003937 | Spatial Distribution of Dengue in a Brazilian Urban Slum Setting: Role of Socioeconomic Gradient in Disease Risk | Few studies of dengue have shown group-level associations between demographic, socioeconomic, or geographic characteristics and the spatial distribution of dengue within small urban areas. This study aimed to examine whether specific characteristics of an urban slum community were associated with the risk of dengue disease.
From 01/2009 to 12/2010, we conducted enhanced, community-based surveillance in the only public emergency unit in a slum in Salvador, Brazil to identify acute febrile illness (AFI) patients with laboratory evidence of dengue infection. Patient households were geocoded within census tracts (CTs). Demographic, socioeconomic, and geographical data were obtained from the 2010 national census. Associations between CTs characteristics and the spatial risk of both dengue and non-dengue AFI were assessed by Poisson log-normal and conditional auto-regressive models (CAR). We identified 651 (22.0%) dengue cases among 2,962 AFI patients. Estimated risk of symptomatic dengue was 21.3 and 70.2 cases per 10,000 inhabitants in 2009 and 2010, respectively. All the four dengue serotypes were identified, but DENV2 predominated (DENV1: 8.1%; DENV2: 90.7%; DENV3: 0.4%; DENV4: 0.8%). Multivariable CAR regression analysis showed increased dengue risk in CTs with poorer inhabitants (RR: 1.02 for each percent increase in the frequency of families earning ≤1 times the minimum wage; 95% CI: 1.01-1.04), and decreased risk in CTs located farther from the health unit (RR: 0.87 for each 100 meter increase; 95% CI: 0.80-0.94). The same CTs characteristics were also associated with non-dengue AFI risk.
This study highlights the large burden of symptomatic dengue on individuals living in urban slums in Brazil. Lower neighborhood socioeconomic status was independently associated with increased risk of dengue, indicating that within slum communities with high levels of absolute poverty, factors associated with the social gradient influence dengue transmission. In addition, poor geographic access to health services may be a barrier to identifying both dengue and non-dengue AFI cases. Therefore, further spatial studies should account for this potential source of bias.
| Dengue is influenced by the environment; however, few studies have investigated the relationship between neighborhood characteristics and the spatial distribution of dengue within small urban areas. We examined whether specific characteristics of an urban slum community were associated with dengue risk. From January 2009 to December 2010, we conducted community-based surveillance in a slum in Salvador, Brazil to identify patients with acute febrile illness (AFI) and to test them for dengue. We identified 651 (22.0%) patients with laboratory evidence of dengue infection among 2,962 AFI patients. All the four dengue serotypes were detected, but DENV2 predominated (DENV1 8.1%; DENV2 90.7%; DENV3 0.4%; DENV4 0.8%). Estimated risk of symptomatic dengue was 21.3 and 70.2 cases per 10,000 inhabitants in 2009 and 2010, respectively. We found that neighborhood poverty level and proximity to the health center were associated with higher risk of detection of dengue and other AFI. This study highlights the large burden of dengue in poor urban slums of Brazil and indicates that socioeconomic development could potentially mitigate risk factors for both dengue and non-dengue AFI cases. In addition, we found that residential proximity to a health care facility was associated with improved case detection. Therefore, further studies on disease distribution should consider household proximity to health care facilities when assessing risk.
| Approximately 2.5 billion people worldwide live in dengue-endemic areas and are at risk for acquiring the infection [1]. Every year, as many as 390 million dengue infections occur, resulting in an estimated 96 million symptomatic cases [2]. In the Americas, dengue incidence has continuously increased since the reintroduction of its vector, the mosquito Aedes aegypti, in the 1970s [3–5]. Brazil accounts for the largest number of dengue cases in the region. In 2013 alone, Brazil reported more than 1.46 million cases of dengue; 61.5% of the total number of cases recorded in the American continent [6–8].
Rapid urbanization, with subsequent increases in population density and poor living conditions, has been associated with the re-emergence of dengue [9]. Currently, approximately one third of the urban population in developing regions live in urban slums and, according to United Nations projections, about 2 billion people will reside in urban slums by 2030 [10,11]. In Brazil, a marked increase in the number of people living in impoverished urban slum communities occurred during the 20th century as a consequence of intense rural to urban migration and population growth [12]. The United Nations estimated that 26.4% of Brazilians lived in slums in 2010 [13]. In Brazil and elsewhere, several studies with ecological design have found associations between increased dengue risk and demographic, socioeconomic, and environmental characteristics, such as high population and household densities [14–17], wide social inequality and low socioeconomic status [18–25], low levels of population education [24–26], presence of a precarious sanitary system [16,17], lack of garbage collection [15,18,27], and low coverage of piped water [28,29].
The majority of these studies have examined large urban areas and compared dengue occurrence among states, counties, or cities. However, dengue transmission is highly focal in space, as the vector typically disperses within a short range (<100 meters) [30,31]. Up to now, it is unknown whether group-level factors are associated with dengue at smaller geographic scales, such as within a neighborhood. In addition, prior studies, particularly those performed in Brazil, used secondary data from the national dengue reporting system. As dengue usually presents with nonspecific clinical manifestations, the disease burden may have been underreported during interepidemic periods and over-reported during epidemic periods [7], a limitation of studies using official surveillance data.
In Salvador, Brazil, dengue has been transmitted endemically since 1995, with approximately 5,000 cases reported each year between 2008 and 2012 [32,33]. We estimated the spatial distribution of symptomatic dengue in an urban slum community in Salvador, and assessed whether group-level demographic, socioeconomic, and geographic factors influenced dengue distribution. Additionally, to investigate whether any associations were specific for dengue, we repeated the spatial distribution analyses and assessed group-level associated factors using cases of non-dengue acute febrile illness (AFI) as the outcome.
Between January 1, 2009 and December 31, 2010, we conducted enhanced community-based surveillance to detect patients with laboratory evidence of dengue infection among those seeking medical care for AFI at the only public emergency health unit (São Marcos Emergency Center [SMEC]; 38°26'09"W, 12°55'32"S) serving the Pau da Lima slum community in Salvador, Brazil (Fig 1A). The study site for the community-based surveillance was arbitrarily defined to have common boundaries with census tract territories, allowing use of official social and demographic population data to determine if AFI patients who sought medical attention at SMEC lived within the study site. In 2010, we performed a community survey and found that 84% (284 of 337) of the study site residents seek medical assistance for AFI at SMEC.
According to the 2010 national census, the population of Salvador was 2.7 million and 76,352 (3%) people lived in the Pau da Lima study surveillance site [34]. The study site was comprised of 98 census tracts (CTs) in an area of 3.7 km2 within the Sanitary District of Pau da Lima, a delimited administrative area with a population of 218,706 in 294 CTs [34]. The site’s topography is characterized by valleys and hills, with an elevation range of 60 meters (S1 Table). Population density was >215,000 inhabitants per km2 for 75% of the study site’s CTs [34]. On average, 71.9% of the families living within the study area had a per capita monthly income lower or equal to the Brazilian minimum wage (R$ 510.00; equivalent to US$289.77, in 2010) (S1 Table) [34]. Demographic and socioeconomic characteristics of the study site varied among the CTs. In general, CTs located around the study health unit presented higher population densities per household and higher percentages of younger inhabitants, black population, illiteracy, and poverty (S1 Fig). Lack of sanitation was more frequent among CTs located in the northeast region of the study site (S1 Fig).
The Zoonosis Control Center at the Municipal Secretary of Health conducted vector control actions within the study site, according to the national guidelines for dengue control and prevention [35]. Vector control activities included community education on vector control measures and bimonthly household visits for entomological surveys and vector control. These actions were routinely performed throughout the study period, except for three months between August 2 and November 3, 2010, when a strike of the dengue control agents interrupted their activities. Although we informed the Pau da Lima Health District about the participants’ laboratory dengue results, we were not able to provide this information in a timely enough fashion to guide the activities of the Zoonosis Control Center agents.
AFI surveillance was performed at SMEC from Mondays to Fridays, from 07h30 to 16h00. During surveillance hours, the study team used medical charts to prospectively identify patients with the following inclusion criteria: age of five years or more, reported fever or measured axillary temperature ≥37.8°C of up to 21 days of duration, and household address inside the study area. Patients who agreed to participate in the study and provided informed consent had an enrollment blood sample collected and were invited to return 15 days later for convalescent-phase blood sample collection. For patients unable to return to SMEC, a study team visited their domiciles to collect convalescent-phase blood samples. Blood samples were maintained under refrigeration and were processed on the same day of collection. Sera were stored at -20°C and -70°C for dengue serological and molecular testing, respectively. The study team retrospectively reviewed medical charts of enrolled participants to collect data on presumptive diagnoses, hospitalization, and death during hospitalization at SMEC. We also reviewed medical charts for every patient attended to at SMEC in 2009 and 2010 to ascertain the number of patients who were eligible for but were not enrolled in the study. Residential addresses for enrolled patients were confirmed by household visits and their positions were marked onto hard copy 1:1,200 scale maps, which were then entered into an ArcGIS database [36]. This database was merged with a cartographical database provided by IBGE [37] to identify the CT of residence of the study patients.
CT-level aggregated data for demographic, socioeconomic, and geographic variables were obtained from the 2010 national census [34]. Topographical data were obtained from IBGE [38]. Demographic variables examined were the mean age of CT population, percentage of residents ≤15 years of age, household density in hundreds (households/100/km2), and population density in hundreds (inhabitants/100 km2). Socioeconomic variables analyzed were the percentage of households with monthly per capita income ≤1 times the national minimum wage, percentage of illiteracy among residents ≥15 years of age, mean number of residents per household, percentage of residents who self-identified as black, percentage of households with inadequate sewage disposal (households without a closed connection to the sewage system or without a closed septic tank), percentage of households without public water supply, and percentage of households not covered by a garbage collection service. We also examined the following geographic variables: mean elevation in relation to sea level, range of elevation (measured as the difference between the highest and lowest points of the CT), and two-dimensional linear distance from the CT centroid to the health unit, which was used as a proxy for health care access.
Acute-phase sera were tested by enzyme linked immunoassay (ELISA) for detection of dengue NS1 antigen and dengue IgM antibodies (Panbio Diagnostics, Brisbane, Australia). Convalescent-phase sera were also tested by dengue IgM ELISA to identify seroconversions. Acute-phase sera from patients who were positive by NS1 ELISA or by IgM ELISA in either the acute- or the convalescent-phase sera were also tested by reverse transcriptase polymerase chain reaction (RT-PCR) [39] to identify the infecting serotype. We defined a dengue case as an AFI patient with a positive NS1 ELISA, acute- or convalescent-phase IgM ELISA, or RT-PCR.
We estimated the population risk of symptomatic dengue as the ratio between the number of dengue cases detected by our surveillance and the area population. We also estimated the risk of non-dengue AFI using the number of enrolled patients without laboratory evidence of dengue as the numerator. Risks were estimated for each CT. Dengue and non-dengue AFI standardized morbidity ratios (SMR) were calculated indirectly by dividing the estimated risk for each CT during the two-year study period by the estimated risk for the overall study area. The SMRs were plotted in study site maps.
Bivariate and multivariable regression analyses to assess associations with estimated risk of dengue were performed using Poisson log-normal models [40]. This model is an extension of the Poisson model, which allows for data overdispersion. Demographic variables associated with dengue risk in the bivariate analysis (P value ≤0.20) were entered into a demographic multivariable backwards regression model. The same approach was used to build socioeconomic and geographic multivariable regression models. Variables with P values ≤0.10 in the demographic, socioeconomic, or geographic multivariable regression models were selected for entry into a final backwards Poisson log-normal model to identify significant associations (P ≤0.05). A conditional autoregressive model (CAR) [40,41] was then used to account for the presence of spatially correlated residuals. The CAR model is an extension of the Poisson log-normal model, with the addition of a spatial component that is dependent on the neighboring structure of the spatial units of analysis. This component assumes that neighboring areas have similar risks, which often results in a smoothed risk map. In our CAR model, we assumed adjacency-based neighborhood spatial weights. A Bayesian approach with non-informative prior distribution for all parameters was applied in the model. Calculations were made using integrated nested Laplace approximation (INLA) [42]. A backwards selection method was also applied to the CAR model to select associated variables (P ≤0.05). Relative risks and 95% confidence intervals (95% CI) were calculated for all the models. Model fitness was assessed by the deviance information criterion (DIC) [43]. We repeated all the steps previously described to identify demographic, socioeconomic, and geographic variables associated with the estimated risk of non-dengue AFI. Risks of dengue and non-dengue AFI, predicted by the final multivariable Poisson log-normal and by the CAR models, were used to calculate adjusted SMRs for each CT, which were plotted in maps. Statistical and spatial analyses were performed using Maptools and INLA packages in the R software (The R Project for Statistical Computing) [42]. The dataset was imported to Quantum GIS software to produce the maps [44].
This project was approved by the Research Ethics Committee at the Gonçalo Moniz Research Center, Oswaldo Cruz Foundation, the Brazilian National Council for Ethics in Research, and the Institutional Review Board of Yale University. All adult subjects provided written informed consent. Participants <18 years old who were able to read provided written assent following written consent from their parent or guardian. All study data were anonymized before analysis.
During the two-year study period, a total of 12,958 study site residents ≥5 years old received medical care for an AFI at SMEC. Among these residents, 3,459 (26.7%) were evaluated for study inclusion (Fig 2). Age and sex distributions for the groups of patients who were and were not evaluated were similar (both groups had median age of 18 years old and were 47% male). Of the assessed patients, 2,962 (85.6%) were enrolled in the study. Patients who were enrolled in the study were older (19 versus 13 years) and were more likely to be male (48% versus 44%) compared to those not enrolled.
An acute-phase blood sample was collected from 2,874 (97.0%) enrolled patients. Paired blood samples were obtained from 2,523 (85.2%) patients. Laboratory testing identified 651 (22.0% of 2,962) patients with evidence of dengue infection; the remaining 2,311 (78.0%) patients were classified as having non-dengue AFI. Among the dengue cases, 380 (58.4%) were acute-phase IgM ELISA positive, 505 (77.6%) were convalescent-phase IgM ELISA positive, 103 (15.8%) were NS1 ELISA positive, and 247 (37.9%) were RT-PCR positive. IgM seroconversion was observed for 207 (31.8%) of the dengue cases. For RT-PCR confirmed cases, 20 (8.1%) were infected with DENV1, 224 (90.7%) with DENV2, 1 (0.4%) with DENV3, and 2 (0.8%) with DENV4. Dengue was less prevalent among patients enrolled in 2009 (152 of 1,466; 10.4%) compared to patients enrolled in 2010 (499 of 1,496; 33.4%). The estimated risk of symptomatic dengue in the study site was 21.29 and 70.23 cases per 10,000 inhabitants in 2009 and 2010, respectively.
The socio-demographic and clinical characteristics of dengue and non-dengue AFI patients are shown in Table 1. Compared to non-dengue AFI, dengue cases more frequently presented with myalgia, retro-orbital pain, arthralgia, and rash. Only 16% of the dengue cases had a presumptive diagnosis of dengue recorded in their medical charts. Yet, the likelihood of dengue suspicion was 7.5 times higher among dengue cases than among non-dengue AFI patients (P <0.001).
We were able to locate the census tract of residence for 570 (87.6%) of the 651 dengue cases and for 1,948 (84.3%) of the 2,311 non-dengue AFI patients. The estimated risks for both dengue and non-dengue AFI were higher for the population living in the census tracts located in the central region of the study site (Fig 1B and 1C).
Multivariable Poisson log-normal models identified the following CT-level factors associated with increased risk of dengue: a shorter linear distance between the centroid of the census tract and the emergency unit, a higher percentage of inhabitants who self-identify as black, and a higher percentage of families earning lower or equal to one Brazilian minimum wage per household inhabitant per month (Table 2). Estimated risk for non-dengue AFI was independently associated with the same CT-level factors, with higher mean age of the CT population as an additional risk factor (Table 2).
Addition of a spatial component to the multivariable models for both dengue and non-dengue AFI improved their fitness and seemed to capture the spatial pattern of dengue and non-dengue AFI, since the non-structured residuals of both models were randomly distributed in space. Dengue risk, assessed by the CAR model, increased 2% (RR: 1.02; 95% CI: 1.01–1.04) for each 1% increase in the percentage of CT families with a monthly income ≤1 times the Brazilian minimum wage per household inhabitant, and decreased 13% (RR: 0.87; 95% CI: 0.80–0.94) for each 100 meter increase in the linear distance between the CT centroid and the surveillance health unit (Table 2). Non-dengue AFI risk also increased as the percentage of families with a monthly income of ≤1 times the minimum wage per household inhabitant increased (RR: 1.03; 95% CI: 1.01–1.04) and decreased as the linear distance between the CT centroid and the surveillance health unit increased (RR: 0.87; 95% CI: 0.80–0.93). In addition, non-dengue AFI was positively associated in the CAR model with higher mean age of the CT population (RR: 1.10; 95% CI: 1.02–1.19), and with a higher percentage of inhabitants who are black (RR: 1.02; 95% CI: 1.01–1.04) (Table 2). Although the spatial distribution of SMRs adjusted by the final Poisson log-normal and CAR models were smoother than the non-adjusted SMRs for both dengue and non-dengue AFI, the CTs located in the central region of the study site maintained a higher relative risk for both conditions (Fig 3).
This enhanced surveillance study highlights the large burden of symptomatic dengue in a poor urban slum community of Salvador, the third largest city in Brazil. Even though the study site was relatively small and characterized by high levels of absolute poverty, the spatial distribution of the detected dengue cases was not homogenous, being influenced by neighborhood characteristics; namely, the gradient of social status and proximity to health services. These findings were not specific for dengue, as the spatial distribution of non-dengue AFI presented the same pattern.
During the study period, the case definition for suspected dengue in Brazil was a patient who lived or traveled to endemic areas and presented with fever up to seven days of duration plus two of the following symptoms: headache, retro-orbital pain, myalgia, arthralgia, rash, or prostration [45]. However, underreporting of patients fulfilling clinical and epidemiological criteria for dengue is common in Brazil and elsewhere [46,47]. Furthermore, dengue reporting tends to be influenced by disease severity and the availability of dengue laboratory testing [48]. We used an enhanced surveillance design to detect AFI patients with laboratory evidence of dengue infection. This approach allowed identification of dengue cases that were unlikely to be reported, as only 16% of the detected cases had a clinical suspicion of dengue recorded in their chart, and also provided more complete epidemiological disease data.
Enhanced surveillance was only conducted during working hours, resulting in an underestimation of dengue and other non-dengue AFI risks. Furthermore, enrolled patients were older than those evaluated but not enrolled, which also might have influenced dengue risk estimation as dengue risk is not equal for all age groups. However, compared to the incidence of reported dengue in the whole Sanitary District of Pau da Lima, the study detected greater dengue risk in 2009 (17.3 and 21.3 cases per 10,000 inhabitants, respectively) and in 2010 (44.1 and 70.2 cases per 10,000 inhabitants, respectively) [33]. This finding is noteworthy, since the study only enrolled 22.9% of the 12,958 AFI subjects from the study site seeking medical attention at SMEC. As the AFI patients who were assessed for study inclusion had comparable age and sex distribution to those not assessed, we can assume that dengue prevalence was similar between these two groups and infer that the actual risk for a dengue episode requiring medical attention in the study site was about four times greater than estimated.
We found a higher risk of dengue associated with poorer areas in the Pau da Lima slum community. Although some population-level studies based on reported dengue cases have also shown an association of symptomatic dengue risk and lower socioeconomic status [21–23,49], others have found an inverse association, where greater incidence occurred in areas of higher income [25,50], or even no association [51]. Discrepancies have been observed in individual-level studies, where dengue occurrence has not been associated [52,53] or was positively [54–56] or negatively [24,57] associated with income and socioeconomic status. It has been speculated that these contradictory results were due to the specificities of each study location, such as level of dengue susceptibility in the population, implementation and coverage of vector control measures, as well as differences in the study spatial unit or the socioeconomic variables considered [18,22,29]. However, poor communities typically have environmental characteristics that facilitate Aedes spp. breeding, including presence of refuse deposits and containers for water storage [58,59]. Therefore, the social gradient we found in association with increased risk of dengue may have acted as a surrogate for other proximal factors involved in dengue transmission.
Proximity of the CTs to the health unit was the variable most strongly associated with detection of dengue. This finding may be due to the fact that CTs located around the health unit had higher population densities per household, and higher percentage of inhabitants <15 years of age (a proxy for susceptibility to dengue infection) (S1 Fig); together these facts might favor dengue transmission as they increase opportunities for interactions between infected and susceptible hosts via the mosquito vector. In bivariate, but not in multivariable and CAR analyses, both population density per household, and percentage of inhabitants <15 years of age were associated with dengue detection. However, the distance between the CTs and the health unit was also positively associated with non-dengue AFI cases detection, suggesting that this association was not specific for a vector-born disease. Therefore, CTs proximity to the study health unit most likely indicates increased opportunity for case detection. Measured distances between households and health facilities have previously been associated with dengue occurrence [28], as well as with poorer colorectal cancer survival [60], lower clinic attendance and a higher degree of dehydration due to diarrhea [61], and decreased use of antenatal healthcare [62], among others. Geographic accessibility to health care is usually observed on a broader scale, especially in developing countries where greater inequalities in health care access are observed in smaller towns distant to large urban centers [63,64]. Our study demonstrates that this phenomenon may also be present at finer geographic scales, such as within urban communities. This finding may be particularly important in spatial distribution studies that use reported cases of mild and self-limited diseases, and that rely on passive surveillance. In this context, areas of higher disease risk may actually represent areas of greater provision of health services and greater opportunity for case detection rather than a true difference in disease frequency.
Other studies have identified a higher occurrence of dengue in areas that lack or have infrequent garbage collection [15,27,65], that have a lower coverage of closed sewer systems [17], and those with low coverage or irregular water supplies [28,57,66]. These associations may be explained by ecological preferences of the mosquito vectors, which find more favorable larval development sites in areas with poorer sanitation infrastructure. In bivariate analysis, we found an association between increased dengue risk and inadequate garbage collection; however, a significant independent association was not observed after adjusting for other covariates. We were unable to identify associations with the coverage levels of piped water supply and sewer provision. The divergence between our findings and those from other studies may be explained by the low variability in the characteristics of the CTs comprising our study site, or by colinearity with other socioeconomic variables included in the model (S2 Table). Alternatively, the inclusion of the distance from the health care unit and the CT centroid in the model may have overshadowed weaker associations.
This study has several limitations. Despite SMEC being the sole public emergency unit in the community, with the second closest public emergency unit located >1.5 km outside from the study site’s boundaries, Pau da Lima residents may have sought care elsewhere. In addition, we were not able to investigate dengue in all AFI patients seeking medical assistance at SMEC. However, AFI patients who were and were not evaluated for study inclusion were similar regarding age and sex distribution, suggesting that selection bias had a minor influence on our results. The CT of residence was not identified for all study participants, but we georeferenced the majority of them (87.6%) and ensured accuracy of the locations of CTs through household visits. We used different laboratory approaches to identify dengue cases. Even though it is likely that we missed some dengue cases by only performing RT-PCR on patients who were NS1 or IgM ELISA positive, the method we used to simultaneous test dengue by IgM and NS1 assays has been shown to increase diagnostic sensitivity [67]. Use of IgM ELISA to confirm dengue is consistent with the Brazilian Ministry of Health guidelines [68,69]; however, dengue IgM may remain detectable up to two months after an infection, and we may have classified patients with recent dengue infection as dengue cases. To account for the possible inclusion of recent asymptomatic infections among dengue cases, we repeated the multivariable Poisson-log normal and the CAR regression analyses using only patients confirmed by IgM ELISA seroconversion, NS1 ELISA, and RT-PCR and found similar associations (S3 Table). Finally, in our model, we could not include data from the Larval Index Rapid Assay for Aedes aegypti (LIRAa), a national survey for positive mosquito breeding sites in a random sample of dwellings [70], because this index is recorded in spatial units that do not align with CTs boundaries.
Strengths of this study include the laboratory testing of all enrolled patients and the assessment of group-level characteristics associated with non-dengue AFI. Additionally, we used conditional auto-regressive models, which increased model fitness by adding a spatially structured component. The increases in model fit indicate that there were residual spatial variations in the risk distributions that had not initially been captured by the studied variables.
Official surveillance systems based on passive reporting underestimate dengue burden; thus, enhanced surveillance is a useful tool to provide more accurate estimates of disease occurrence and its spatial distribution. According to the World Health Organization guidelines for dengue prevention and control, estimating the true burden of the disease is a critical step to achieve the goal of reducing dengue disease burden [71]. Our findings corroborate those of other studies showing that implementation of sentinel health unit-based enhanced surveillance for dengue is feasible and may be employed to obtain high quality information on disease trends and circulating serotypes as well as increase opportunities for timely detection and intervention during epidemics, which may not be achieved by passive surveillance [46,72].
In several settings, low socioeconomic status has been observed to impact dengue transmission [21,54,73], emphasizing that the disease burden is likely to be greatest in vulnerable populations such as urban slum dwellers, and as we found in this study, the poorest segments of such populations. Until initiatives address social inequity and the underlying poverty-associated environmental determinants of dengue transmission, specific vector control actions that are difficult to apply citywide, such as biological larvae control, may target groups at higher disease risk, such as those living in the poorer areas of urban communities.
Finally, studies aiming to assess spatial distribution and group-level associated factors of diseases should account for potential detection bias. With the popularization of GIS and spatial analysis tools, the distance between each area unit and the closest health service is a viable proxy for health care accessibility and its use may help explain the spatial distribution of health and disease.
|
10.1371/journal.pmed.1002645 | Crowdsourcing to expand HIV testing among men who have sex with men in China: A closed cohort stepped wedge cluster randomized controlled trial | HIV testing rates are suboptimal among at-risk men. Crowdsourcing may be a useful tool for designing innovative, community-based HIV testing strategies to increase HIV testing. The purpose of this study was to use a stepped wedge cluster randomized controlled trial (RCT) to evaluate the effect of a crowdsourced HIV intervention on HIV testing uptake among men who have sex with men (MSM) in eight Chinese cities.
An HIV testing intervention was developed through a national image contest, a regional strategy designathon, and local message contests. The final intervention included a multimedia HIV testing campaign, an online HIV testing service, and local testing promotion campaigns tailored for MSM. This intervention was evaluated using a closed cohort stepped wedge cluster RCT in eight Chinese cities (Guangzhou, Shenzhen, Zhuhai, and Jiangmen in Guangdong province; Jinan, Qingdao, Yantai, and Jining in Shandong province) from August 2016 to August 2017. MSM were recruited through Blued, a social networking mobile application for MSM, from July 29 to August 21 of 2016. The primary outcome was self-reported HIV testing in the past 3 months. Secondary outcomes included HIV self-testing, facility-based HIV testing, condom use, and syphilis testing. Generalized linear mixed models (GLMMs) were used to analyze primary and secondary outcomes. We enrolled a total of 1,381 MSM. Most were ≤30 years old (82%), unmarried (86%), and had a college degree or higher (65%). The proportion of individuals receiving an HIV test during the intervention periods within a city was 8.9% (95% confidence interval [CI] 2.2–15.5) greater than during the control periods. In addition, the intention-to-treat analysis showed a higher probability of receiving an HIV test during the intervention periods as compared to the control periods (estimated risk ratio [RR] = 1.43, 95% CI 1.19–1.73). The intervention also increased HIV self-testing (RR = 1.89, 95% CI 1.50–2.38). There was no effect on facility-based HIV testing (RR = 1.00, 95% CI 0.79–1.26), condom use (RR = 1.00, 95% CI 0.86–1.17), or syphilis testing (RR = 0.92, 95% CI 0.70–1.21). A total of 48.6% (593/1,219) of participants reported that they received HIV self-testing. Among men who received two HIV tests, 32 individuals seroconverted during the 1-year study period. Study limitations include the use of self-reported HIV testing data among a subset of men and non-completion of the final survey by 23% of participants. Our study population was a young online group in urban China and the relevance of our findings to other populations will require further investigation.
In this setting, crowdsourcing was effective for developing and strengthening community-based HIV testing services for MSM. Crowdsourced interventions may be an important tool for the scale-up of HIV testing services among MSM in low- and middle-income countries (LMIC).
ClinicalTrials.gov NCT02796963
| HIV testing remains low among key populations, including MSM, especially in low- and middle-income countries.
Although crowdsourcing has been recommended as a tool for developing public health interventions, few studies have formally evaluated crowdsourcing as a way to promote health in a real-world setting.
Previous studies on crowdsourcing have suggested that it may be a promising approach to increase HIV testing among MSM in China.
We investigated the effectiveness of a crowdsourced intervention in promoting HIV testing among Chinese MSM in eight cities distributed in two provinces.
We found that the crowdsourced intervention was effective in promoting HIV testing. Compared to the control period, the intervention period showed an 8.9% absolute increase (and a 43% relative increase) in HIV testing.
The intervention was particularly effective in promoting HIV self-testing.
In this setting, we found crowdsourcing was a way to develop innovative and effective HIV testing services. Further research is needed to examine this approach in other settings.
A crowdsourcing approach can help build and sustain community engagement related to HIV testing.
Programs aiming to expand HIV testing should consider crowdsourcing as a tool to design tailored services.
| Approximately 14 million people living with HIV have yet to be tested, compromising the effectiveness of HIV treatment and prevention programs [1]. Testing rates are particularly poor among key populations (e.g., men who have sex with men [MSM], 25%–32%) in low- and middle-income countries (LMIC) [1,2]. Entrenched community norms that marginalize key populations, limited HIV resources, and insufficient community awareness all contribute to low levels of HIV testing around the world, including China [3–5].
In China, HIV infection rates are still increasing among MSM, highlighting the need for innovative HIV testing interventions [6,7]. Recent global literature suggests that community involvement and social media can be important tools in reaching high-risk populations and improving HIV testing rates [8,9]. In China, community engagement has also been shown to be an important predictor of HIV testing [10], and previous community-based interventions have shown promise [11,12].
Crowdsourcing can be a useful tool to develop community-driven HIV testing services in LMIC [13]. It allows a group, including experts and nonexperts, to solve problems and share solutions with the public [13]. Crowdsourcing is a scalable, cost-effective tool to aggregate community wisdom [14]. Crowdsourcing approaches have been used to organize half a million people to help identify viral protein structures [15], deployed laypeople for cardiopulmonary resuscitation in out-of-hospital cardiac arrest [16], and solicited videos that promote HIV testing [14].
However, the potential for crowdsourcing as a tool to develop community health services, such as HIV testing, is uncertain; crowdsourcing for health improvement has focused on developing single components of health campaigns [17], rather than creating a comprehensive service. Pilot trials have suggested that crowdsourcing may be a useful tool for developing images and videos to promote HIV testing [14,18]. Further research is needed to examine whether crowdsourcing can effectively improve HIV testing services among key populations in LMIC.
The purpose of this stepped wedge cluster randomized controlled trial (RCT) was to evaluate a comprehensive crowdsourced intervention to increase HIV testing uptake among MSM in China.
The study protocol provides a more detailed description of the trial design and analysis plan (S1 Text) [19]. Briefly, we developed an intervention consisting of a multimedia HIV testing campaign, an online HIV testing service, and local testing promotion campaigns tailored for MSM. The intervention was developed using crowdsourcing and consisted of three components: a nationwide open contest call for images and concepts that encourage HIV testing, a 72-hour regional designathon for developing HIV testing strategies [20], and local participatory contests soliciting HIV testing messages. We then conducted a closed cohort stepped wedge cluster RCT in eight Chinese cities to evaluate the impact of this crowdsourced intervention compared to conventional programs routinely provided by local Centers for Disease Control (CDCs) and community-based organizations (CBOs). The intervention was implemented over 12 months. We followed standard guidelines for reporting stepped wedge cluster RCTs (S1 CONSORT Checklist) [21].
The steps for intervention development are shown in Fig 1 and further explained in S2 Text. Three participatory activities shaped the content and structure of the intervention: a nationwide open contest, a regional strategy designathon, and local participatory contests.
We selected four cities each from Guangdong province (Guangzhou, Shenzhen, Zhuhai, and Jiangmen) and Shandong province (Jinan, Qingdao, Yantai, and Jining). Each city had existing infrastructure for MSM HIV surveillance led by the local CDC and capacity to deliver new HIV testing services. Participants were eligible if they were born biologically male, age 16 or older, currently living and planning to live in one of the eight cities for 12 months post-enrollment, HIV-negative or unknown HIV status, had not had HIV testing within the past 3 months, had anal sex with a man at least once during their lifetime, and were willing to provide their cell phone numbers for follow-up. We recruited participants through China’s largest MSM social networking mobile phone application (app), Blued, by sending a survey invitation to registered users in the eight selected cities.
We randomly assigned the order of intervention for each of the four cities in Guangdong province and Shandong province, then paired the cities by order of intervention (S1 Table). Prior to receiving the intervention, cities were considered to be in the control state. We initiated the intervention for each pair at 3-month intervals, and each pair of cities received the intervention for 3 consecutive months. In total, we collected data at baseline followed by four data collection points over 12 months.
After electronically signing an informed consent form, all participants were asked to fill out a survey at baseline, and at every 3 months thereafter. Any participant who returned a photo of a positive test result was counseled to seek out confirmatory testing from their local CDC or CBO and was no longer eligible for subsequent follow-up surveys. Conditions of the control state, intervention, and post-intervention period are shown in Fig 3.
To determine the necessary sample size for recruitment, we assumed that a crowdsourced intervention would be more effective than a conventional method for promoting HIV testing [14]. We assumed an HIV testing rate of 25% during the control period and 35% during the intervention period. These assumptions were made based on existing levels of HIV testing under conventional HIV testing promotion and pilot data evaluating crowdsourced HIV testing videos in China [14]. With eight clusters (cities), four intervention time lines, a coefficient of variation of 0.4, a 0.05 significance level, 90% power, and 30% loss to follow-up, we planned to recruit at least 1,040 men from the eight cities (130 from each city).
The primary outcome of this study was the proportion of participants who tested for HIV over the previous 3 months. This was assessed based on self-reported data from each follow-up survey. We also measured the following secondary outcomes: facility-based HIV testing; HIV self-testing; syphilis testing; condomless sex; using WeChat, Weibo, QQ, or mobile phone applications to give/receive information about HIV testing; anticipated HIV stigma; HIV testing social norms; HIV testing self-efficacy; and community engagement in sexual health. We adapted validated scales to measure anticipated HIV stigma [22], HIV testing social norms [23], HIV testing self-efficacy [24], and increase in community engagement in sexual health during the follow-up period (as compared to baseline) [10]. The primary and secondary outcomes were measured at baseline and in each of the four follow-up surveys at 3, 6, 9, and 12 months after baseline. We also measured incident HIV testing during the study period, defined as a participant’s first HIV test during the study period, as well as incident HIV testing among participants who had never tested for HIV at baseline.
Descriptive analysis was used to summarize the demographic characteristics and HIV testing proportions. To investigate the effect of the intervention, we evaluated the difference in probability of HIV testing in the control period and in the intervention period (including the post-intervention period). We applied intention-to-treat analysis utilizing generalized linear mixed models (GLMMs) to evaluate the primary and secondary study outcomes. Intervention status and time indicators allowing for piecewise secular trends were considered fixed effects, while sites and individual participants with multiple measurements across the four follow-ups were considered random effects [25]. We treated all intervention periods and post-intervention periods in the same manner. The estimated intervention effect sizes (risk ratio [RR] for binary outcomes, and mean difference for continuous outcomes) were each reported with a 95% confidence interval (CI) and p-value. We encountered non-convergence problems when a log-binomial GLMM was used to estimate relative risk. Thus, we employed a log Poisson GLMM model [26].
Sensitivity analyses were conducted to evaluate a per-protocol effect and a city/cluster-level effect of the intervention on HIV testing proportion in the past 3 months. The per-protocol analysis only included participants who reported viewing the intervention materials. For the city-level intervention effect, we used the HIV-testing proportion of each city across the four follow-ups as the outcome, and fit a normal linear mixed model that accounted for the random effect of sites to estimate the difference in HIV testing proportions. An intention-to-treat sensitivity analysis was conducted to evaluate the effectiveness of the intervention while adjusting for the province in the model. According to our prespecified analysis plan, we tested interactions with age (>30 years old versus ≤30 years old) and for cities with in-person community activities during the intervention development phase (Jinan, Qingdao, Guangzhou, and Shenzhen). We also used multiple imputations to examine the intervention effect. Using multiple imputations by chained equations and 30 imputed datasets, missing HIV testing data were imputed at each follow-up period using a logit model that included baseline variables (age, marital status, province, and income) and intervention status during the respective follow-up period. A log Poisson GLMM was fit to estimate relative risk. Rubin’s rules were employed to pool the parameter estimates using the MIANALYZE Procedure in SAS [27]. All data analyses were completed using SAS 9.4 (Cary, NC).
Ethical approval was obtained from the institutional review committees at the Guangdong Provincial Center for Skin Diseases and STI Control (Guangzhou, China), Shandong University (Jinan, China), University of North Carolina at Chapel Hill (Chapel Hill, NC), and the University of California, San Francisco (San Francisco, CA). The trial is registered with ClinicalTrials.gov (NCT02796963).
Participants were recruited from July 29, 2016, and followed until August 21, 2017. Overall, the study link was clicked 39,764 times. Of these, 16,193 withdrew from the survey prior to reading the consent form and 21,187 people did not meet eligibility requirements. Among the remaining, 1,003 did not sign the informed consent form. (S1 Fig)
A total of 1,381 participants were enrolled. Of these, 203, 139, 134, and 203 were recruited from Guangzhou, Jiangmen, Zhuhai, and Shenzhen, respectively, in Guangdong province, while 180, 189, 182, and 151 were recruited from Yantai, Jinan, Qingdao, and Jining, respectively, in Shandong province (Fig 4, S1 Table).
Based on the predetermined randomization schedule, 383 participants from Guangzhou and Yantai were assigned to the first intervention group (Group 1), 328 participants from Jiangmen and Jinan were assigned to the second group (Group 2), 316 participants from Zhuhai and Qingdao were assigned to the third group (Group 3), and 354 participants from Shenzhen and Jining were assigned to the fourth group (Group 4) (Fig 4, S1 Table).
Among the 1,313 HIV-negative or status unknown participants after the third follow-up, 306 did not finish our last survey, with a loss-to-follow-up rate of 23% (306/1,313, Fig 4). Loss-to-follow-up rates were similar between the four intervention groups. Characteristics of participants lost to follow-up differed in age and income from participants who completed the last follow-up (S2 Table).
A total of 1,219 participants completed at least one follow-up survey. Participants who missed one follow-up survey but rejoined the study at a subsequent survey date were assumed not to have tested during the missed follow-up period.
Demographic characteristics of the participants were similar across the four groups (Table 1). The majority of participants were 30 years old or younger (82%), had never married (86%), had a college degree or higher (65%), had an annual income below US$9,500 (74%), and had disclosed their sexual orientation (65%). Most participants (95%) identified as male and the rest identified as transgender. Seventy-one percent reported their sexual orientation as “tongxinglian,” gay. In addition, 73% of the participants reported that they had engaged in condomless sex in the past 3 months and 5% tested for syphilis in the past 3 months. At baseline, 57% of participants had never tested for HIV.
Overall, the proportion testing for HIV in the last 3 months increased after the intervention and this trend was maintained during the post-intervention periods (Table 2). Similar results were also found for each study city (S1 Table). The number of incident testers in each follow-up period and the number of incident testers who had never tested for HIV prior to baseline are shown in S3 Table. Generally, incidence of HIV testing was high during the intervention period.
The proportion of individuals receiving an HIV test within a city was 8.9% (95% CI 2.2–15.5) greater during the intervention periods (Table 3). In addition, results from the intention-to-treat analysis showed that the probability of an individual HIV testing during the intervention periods (including post-intervention periods) was higher than in the control periods (estimated RR = 1.43, 95% CI 1.19–1.73) (Table 3). Multiple imputation analysis produced a similar result. In the per-protocol analysis, the estimated effect size was larger (RR = 1.49, 95% CI 1.21–1.83).
The model with interaction term suggested that the intervention effect was similar among MSM who were 16 to 30 years old (RR = 1.41, 95% CI 1.16–1.72) compared to MSM over 30 years old (RR = 1.57, 95% CI 1.12–2.21, interaction test p = 0.52). In addition, the intervention had a similar effect in cities with in-person community activities (RR = 1.56, 95% CI 1.24–1.96) compared to cities that did not have in-person activities (RR = 1.35, 95% CI 1.06–1.73, interaction test p = 0.27) (Table 3).
Out of 1,219 participants who completed at least one follow-up survey, 755 (62%) unique participants self-reported that they tested for HIV at any point during the study period. Of these, 642 (85%) tested during the intervention and post-intervention periods (S3 Table). Of the 699 participants who had never tested for HIV at baseline, 390 (56%) tested at any point during the study period (S3 Table). A total of 395 participants reported testing only once during the follow-up period, whereas 360 participants tested during more than one follow-up period (211 people tested twice, 107 people tested three times, and 42 people tested four times [S4 Table]).
A total of 593 participants (49% of 1,219 participants who completed at least one follow-up) received HIV self-testing. Of these, 442 (75%) used our self-testing platform, and 132 (30%) returned their self-test results via WeChat. Among these individuals, seven (5%) confirmed results were positive. We found that 93.9% of self-reported results were consistent with the returned photo of results (S5 Table).
By the end of the study period, 99 participants (50 from Guangdong province; 49 from Shandong province) reported a positive HIV test result. The cumulative HIV prevalence among those who tested for HIV was 13.1%. Of these, 32 individuals seroconverted during the 1-year study period. Seroconversion was defined has having an initial negative HIV test and then a subsequent positive HIV test.
Table 4 reports the results for secondary outcomes using an intention-to-treat approach. Our intervention increased HIV self-testing (RR = 1.89, 95% CI 1.50–2.38, S6 Table; S4 Text lists further statistical considerations). There was no difference in facility-based HIV testing between the intervention and control periods (RR = 1.00, 95% CI 0.79–1.26). The crowdsourced intervention also did not improve condom use, syphilis testing, or anticipated HIV stigma (Table 4).
HIV testing is an essential first step in the HIV care continuum. Although HIV testing is a major global health priority, in key populations large numbers of people remain untested [28]. We recruited MSM from eight Chinese cities and followed individuals longitudinally for 12 months to evaluate the effect of an intervention developed through crowdsourcing on promoting HIV testing. We found that the crowdsourced intervention was effective in promoting HIV testing compared to the control period, showing an 8.9% absolute increase (and a 43% relative increase) in HIV testing during the intervention period. The intervention was particularly useful in promoting HIV self-testing. Our study extends previous research on crowdsourcing by using it to develop a comprehensive HIV testing service, evaluating its effectiveness in a pragmatic trial, and assessing the long-term effect of the intervention [17]. In contrast with our study, nearly all of the limited crowdsourcing health research studies have been observational to date [17].
Within our study cohort, 62% of participants self-reported that they received HIV testing at least once during the study period, and 56% of previously untested MSM received HIV testing. This is consistent with the literature on global community-based HIV testing promotion [14,28,29] but breaks new ground in formally evaluating crowdsourcing as an approach for developing HIV testing services to identify untested individuals. The HIV prevalence we observed was higher than that reported among studies from MSM in Shandong province [30,31], suggesting a higher burden of HIV among previously untested MSM populations. However, it also suggests that community-based, crowdsourced interventions are capable of reaching these populations. Given the relatively low cost of crowdsourcing [14], these types of approaches may be useful in a number of LMIC settings.
The World Health Organization now recommends HIV self-testing [32]. While there is evidence suggesting that HIV self-testing can increase HIV testing among MSM [33–37], studies have not previously measured longitudinal effects in China or described community input. In our study, 49% of participants underwent HIV self-testing, 75% of whom used our self-testing platform, further supporting the acceptability of HIV self-testing. Our study also showed that the crowdsourced intervention increased HIV self-testing but not facility-based testing, highlighting the need for further research to enhance linkage to HIV prevention and other services.
Our data indicated that the intervention was effective among both young and older MSM. Young MSM were heavily invested in all of the participatory activities that shaped the design and implementation of the intervention. Given the high rates of WeChat/Weibo use among youth, we speculate that crowdsourcing may be particularly effective among young MSM who use social media. Considering that our intervention engaged substantial youth in each city to provide feedback about HIV services [17,20], crowdsourcing may help to develop more youth-friendly HIV services.
Our study has several research and policy implications. It suggests that crowdsourcing could be used to design tailored HIV services, providing direction for subsequent crowdsourcing research. Also, crowdsourcing provides an inclusive, effective way to solicit community input; hence, it might be used to inform health policy [38,39]. Finally, when planning HIV interventions for MSM, researchers and policy makers should consider social media interventions to expand dissemination [9].
Our study has several limitations. First, most HIV testing data were based on self-report. Nevertheless, previous studies have demonstrated that HIV test self-report reliably correlates with operational HIV testing data [40,41]. Within the sample of participants who returned images of their self-testing kit results, the vast majority of returned results was consistent with the self-reported results (S5 Table). Second, 23% of the non-seroconverted men did not finish our last survey, a relatively large loss to follow-up. Characteristics of participants lost to follow-up differed from participants who completed the last follow-up in age and income (S2 Table). However, our findings were robust when adjusted for age and income, and also when using multiple imputation analysis. Third, this study was limited in its generalizability. The RCT only included MSM recruited online. However, a previous study suggested that findings from online MSM studies in China are comparable to national MSM data [42]. Additionally, this study recruited mostly young men from urban China. Further research is needed among groups of different ages, cultures, and locations. Fourth, the implementation of HIV self-testing was delayed in two cities because of logistical problems. This may explain why HIV testing rates were lower in the earlier groups, which would bias our effect estimates towards the null, suggesting that reported results are even more conservative than the true effect. Fifth, we did not collect data on linkage to care. Linkage to care is a key component for HIV testing. Previous work has shown that CBOs can adopt innovative methods to promote linkage to care, which can be incorporated into future interventions [43]. Finally, there may have been contamination between the intervention and control periods, especially among participants of the crowdsourcing contest and designathon who may have viewed intervention materials in advance. We did not collect information on whether men participated in contests used to develop the intervention. Because it would be impossible to determine if control groups had inadvertently seen the intervention without exposing them to the intervention, we did not collect information on potential spillover effects. However, given that the intraclass correlation for participants within each city was low, we anticipate that the impact of the spillover would also be small.
In summary, crowdsourcing can help spur the development of new HIV testing services. Our data demonstrate that a crowdsourcing approach effectively increased HIV testing in Chinese cities, especially HIV self-testing among MSM. While the crowdsourcing approach was implemented across cities, each city’s local contests helped shape and contextualize testing messages for local communities. Crowdsourcing approaches may be an important tool for localizing and differentiating HIV services.
|
10.1371/journal.pntd.0003794 | Harnessing Case Isolation and Ring Vaccination to Control Ebola | As a devastating Ebola outbreak in West Africa continues, non-pharmaceutical control measures including contact tracing, quarantine, and case isolation are being implemented. In addition, public health agencies are scaling up efforts to test and deploy candidate vaccines. Given the experimental nature and limited initial supplies of vaccines, a mass vaccination campaign might not be feasible. However, ring vaccination of likely case contacts could provide an effective alternative in distributing the vaccine. To evaluate ring vaccination as a strategy for eliminating Ebola, we developed a pair approximation model of Ebola transmission, parameterized by confirmed incidence data from June 2014 to January 2015 in Liberia and Sierra Leone. Our results suggest that if a combined intervention of case isolation and ring vaccination had been initiated in the early fall of 2014, up to an additional 126 cases in Liberia and 560 cases in Sierra Leone could have been averted beyond case isolation alone. The marginal benefit of ring vaccination is predicted to be greatest in settings where there are more contacts per individual, greater clustering among individuals, when contact tracing has low efficacy or vaccination confers post-exposure protection. In such settings, ring vaccination can avert up to an additional 8% of Ebola cases. Accordingly, ring vaccination is predicted to offer a moderately beneficial supplement to ongoing non-pharmaceutical Ebola control efforts.
| Public health efforts for controlling the 2014–2015 Ebola outbreak in West Africa have focused on contact tracing and isolation of symptomatic individuals. In addition, substantial resources have been committed to scaling up the production of experimental vaccines. Ring vaccination—the vaccination of the contacts of an infected individual—was successfully implemented to achieve smallpox eradication. Ring vaccination is particularly feasible and effective in settings where the supply of vaccines is limited and disease incidence is low. Using a disease transmission model, we evaluated the benefit of adding ring vaccination to case isolation in Liberia and Sierra Leone. We found that ring vaccination could have averted up to 126 cases in Liberia and 560 cases in Sierra Leone, thereby saving lives and intervention resources.
| The Ebola outbreak in West Africa has resulted in unprecedented morbidity and mortality. As of March 4 2015, the World Health Organization (WHO) had reported 23,914 cases and 9,792 fatalities in countries with widespread transmission [1], with Liberia and Sierra Leone having been most profoundly impacted.
Ebola transmission occurs via direct human-to-human contact with body fluids from symptomatic patients. An elevated viral load in late-stage symptomatic or deceased victims can also put family members and funeral attendees at risk of post-mortem disease transmission [2, 3]. The health ministries in Liberia and Sierra Leone have been implementing intensive contact-tracing procedures, where patients or their relatives are interviewed to identify people with whom they came into close contact after developing symptoms. Contacts who are healthy but might have been exposed are monitored for 21 days, the maximum duration of the Ebola incubation period [4]. Contacts who present with Ebola symptoms, such as fever, are transported to isolation clinics [5]. Given that Ebola is transmitted directly between close contacts, the social clustering of individuals can be fundamental to the success of intervention strategies [6–8].
Several Ebola vaccine candidates have been developed in the past decade [9, 10], some of which have already been found to be safe and immunogenic in Phase 1 clinical trials [11]. One of these, a recombinant vesicular stomatitis viruses (rVSV) vaccine, conferred protection to non-human primates when administered immediately following exposure to an otherwise lethal dose of Ebola virus [12]. An alternate vaccine formula based on the chimpanzee adenovirus type 3 (ChAd3), developed in partnership between the National Institute of Allergy and Infectious Disease and GlaxoSmithKline, together with the rVSV vaccine are currently entering Phase 2/3 clinical trials in West Africa [13]. In addition, the WHO has deemed it ethical to use experimental vaccines in the current Ebola emergency situation [14].
Even with the scale up in production [15], the supply would be insufficient for mass vaccination of affected countries, given a combined population of over twenty million people. Consequently, the judicious prioritization of vaccine recipients is essential to maximize vaccine impact. Aside from vaccinating healthcare workers who are at high occupational risk of contracting Ebola [16, 17], a vaccination strategy to reduce community-wide transmission has yet to be evaluated. Evaluating Ebola vaccination strategies is pertinent not only to the current epidemic, but also in mitigating future outbreaks.
The targeting of the exposed contacts of infected individuals, a strategy known as ring vaccination [18, 19], is efficient for controlling rare pathogens [20]. For example, ring vaccination proved to be an effective strategy for smallpox eradication [18, 19]. Furthermore, ring vaccination could be seamlessly incorporated into the contact tracing efforts underway in the affected countries.
To evaluate the effectiveness of Ebola ring vaccination in West Africa, we developed a mathematical model that approximates disease progression in a realistic contact network. We predicted that the marginal benefit was greatest in settings where there are more contacts per individual, greater clustering, or insufficient resources for effective contact tracing. However, we found that ring vaccination provides moderate marginal benefit beyond current non-pharmaceutical interventions.
To determine the effectiveness of ring vaccination and case isolation of Ebola in West Africa, we modeled the transmission between close contacts by using the pair-approximation methodology which simulates disease propagation through a network (Fig 1, S1 Text) [6, 21]. Specifically, we tracked both the number of susceptible individuals ([S]), latently infected individuals ([E]), infectious individuals ([I]), removed individuals ([R]), as well as the number of contacts between epidemiological states. For example, [SI] denotes the number of contacts between susceptible and infectious individuals. We denoted the average number of contacts per individual in the network by k. To account for empirical mixing patterns between individuals during the Ebola outbreak in West Africa [22], we considered clustering of individuals. Clustering can be defined as the extent to which individuals who are in contact with each other share other contacts in the network and is quantified by the clustering coefficient (ϕ).
A susceptible individual becomes latently infected (E) at rate β[SI] per day, where β is the transmission rate. An individual remains in the latent period for an average duration of 1/σ days until becoming symptomatic and infectious (I). An infectious individual will transition to the removed state (R) (i.e. recovered or deceased) at rate δ, where 1/δ days is the average duration of the infectious period. In addition, we incorporated case isolation by removing a percentage of infectious individuals (ψ) from the community at rate γ per day (Table 1).
For our base case analysis, we used k = 5.74 as the mean number of contacts and ϕ = 0.21 for the clustering coefficient, derived from contact tracing data collected by the Liberian Ministry of Health and Social Welfare [22]. We also considered clustering coefficients of 0.10 and 0.40 (Table 1), consistent with previous studies on human contact networks [23–26]. In addition, we accounted for possible under-reporting of contacts by considering higher values of k (Table 1).
During contact tracing, an infected contact in the latent period moves to the observed state (TE) while a contact in the symptomatic period enters the isolated state (TI) at a daily contact tracing rate of τ per isolated case. Once an isolated individual has recovered and is no longer infectious, they transition to the TR state (S1 Text), while the individuals contacts are followed for an additional 1/ω days (Table 1). We defined the contact tracing efficacy by the probability of identifying an infected individual before transmission occurs (τ/(τ+ν)), where 1/ν is the average serial interval (Table 1). We assumed that the base case contact tracing efficacy was 40%, consistent with empirical estimates [27, 28]. We deemed the current Ebola epidemic to be eliminated at the point where incidence became lower than 0.025 cases per day, which corresponds to no new cases over a 42 day period [29].
d [ S ] d t = - β [ S I ] d [ E ] d t = β [ S I ] - τ ( [ E T I ] + [ E T R ] ) - σ [ E ] d [ I ] d t = σ [ E ] - ( 1 - ψ ) δ [ I ] - ψ γ [ I ] - τ ( [ I T I ] + [ I T R ] ) d [ R ] d t = ( 1 - ψ ) δ [ I ] + ω [ T R ] d [ T E ] d t = τ ( [ E T I ] + [ E T R ] ) - σ [ T E ] d [ T I ] d t = τ ( [ I T I ] + [ I T R ] ) - δ [ T I ] + σ [ T E ] + ψ γ [ I ] d [ T R ] d t = δ [ T I ] - ω [ T R ] d [ S I ] d t = - β ( [ S I ] + ( 1 - ϕ ) k - 1 k [ S I ] 2 [ S ] + ϕ k - 1 k N k [ S I ] 2 [ I I ] [ I ] 2 [ S ] ) - τ ( ( 1 - ϕ ) k - 1 k [ T I I ] [ S I ] [ I ] + ϕ k - 1 k N k [ T I I ] [ S I ] [ S T I ] [ S ] [ I ] [ T I ] ) -τ ( ( 1 - ϕ ) k - 1 k [ T R I ] [ S I ] [ I ] + ϕ k - 1 k N k [ T R I ] [ S I ] [ S T R ] [ S ] [ I ] [ T R ] ) +σ [ S E ] - ( 1 - ψ ) δ [ S I ] - ψ γ [ S I ]
We assumed that a proportion of the contacts of isolated cases identified through contact tracing are vaccinated. Experimental studies indicate that several vaccine candidates facilitate recovery of latent infection in non-human primates if administered within two days post-infection [12, 30]. We used a vaccine efficacy (ɛ) of 100% as our base case scenario and consider a range of vaccine efficacies from 5% to 95% (S1 Text). We investigated two scenarios: 1) vaccination must be administered pre-exposure in order to confer protection such that only susceptible individuals (S) who are uninfected can be protectively vaccinated, and 2) vaccination is efficacious with both pre- and post-exposure administration such that vaccine protection can be conferred to individuals in both the susceptible (S) and latently infected (E) [30]. The use of a post-exposure vaccine in the equations below is represented by χ = 1, otherwise χ = 0. Susceptible and latently infected individuals are vaccinated at the same daily contact tracing rate τ per isolated individual (S1 Text). When a susceptible individual is vaccinated they remain unprotected (P) until vaccine-mediated immunity is acquired 1/υ days later.
d [ S ] d t = - β [ S I ] - τ ( [ S T I ] + [ S T R ] ) d [ E ] d t = β [ S I ] - τ ( [ E T I ] + [ E T R ] ) - σ [ E ] d [ I ] d t = σ [ E ] + σ [ F ] - ( 1 - ψ ) δ [ I ] - τ ( [ I T I ] + [ I T R ] ) - ψ γ [ I ] d [ R ] d t = ( 1 - ψ ) δ [ I ] + χ τ ( [ E T I ] + [ E T R ] ) + ω [ T R ] + υ [ P ] d [ P ] d t = τ ( [ S T I ] + [ S T R ] ) - β [ P I ] - υ [ P ] d [ F ] d t = β [ P I ] - τ ( [ F T I ] + [ F T R ] ) - σ [ F ] d [ T E ] d t = ( 1 - χ ) τ ( [ E T I ] + [ E T R ] ) + τ ( [ F T I ] + [ F T R ] ) - σ [ T E ] d [ T I ] d t = τ ( [ I T I ] + [ I T R ] ) - δ [ T I ] + σ [ T E ] + ψ γ [ I ] d [ T R ] d t = δ [ T I ] - ω [ T R ]
To account for the recent decline in Ebola incidence [31, 32], we used a piecewise approach to calibrate our model, fitting both the early growth phase of the epidemic and the later phase of reduced transmission (S1 Fig, S1 Table, S1 Text). We estimated the rate of transmission per infectious contact (β), the date of Ebola emergence into the population (t0), the initiation of intervention scale up (tS), and reduced transmission mediated by factors other than intervention, such as behavior change (ξ). From the date of intervention scale up, we assumed that case isolation was implemented at our base case contact tracing efficacy. Using a least squares fitting algorithm, we estimated these four parameters from weekly confirmed incidence data in Liberia and Sierra Leone, respectively (S1 Table and S1 Text). We fit the model to confirmed incidence data from June 8, 2014 to January 4, 2015 for Liberia [31] and from May 11, 2014 to January 4, 2015 for Sierra Leone [32] (S1 Table and S1 Text). However, on December 20, 2014, the deployment of international aid began in Sierra Leone and considerably reduced transmission [33–35]. Thus, we considered a second reduction in transmission from this time point (S1 Text). To validate the calibration of Liberia and Sierra Leone, we forecasted the incidence of Ebola to March 8, 2015 and calculated the correlation fit value to the observed incidence (S1 Fig, S1 Table).
We estimated that intervention efforts were initially scaled up on September 5, 2014 in Liberia and October 8, 2014 in Sierra Leone. Our model predicts that under status quo intervention the total number of confirmed Ebola cases in Liberia will be 3,899 cases and 10,425 cases in Sierra Leone (Fig 2). If ring vaccination using a prophylactic vaccine had been combined with the initial scaling up of non-pharmaceutical interventions, four additional cases could have been averted in Liberia and 36 cases could have been averted in Sierra Leone, corresponding to relatively low marginal benefits of 0.21% and 0.48% additional cases averted, respectively (S2 Fig). By contrast, for a vaccine that can confer post-exposure protection, we estimate that 107 cases could have been averted in Liberia and 477 cases could be averted in Sierra Leone (Fig 2), corresponding to marginal benefits of 4.7% and 6.5%, respectively (S3 Fig).
We found that a prophylactic vaccine had minimal impact on the number of symptomatic individuals identified by contact tracing (S4 Fig). However, if the vaccine confers post-exposure protection, then ring vaccination is predicted to reduce the number of symptomatic individuals identified by contact tracing by 40 in Liberia and 137 in Sierra Leone compared to case isolation alone, thereby reducing requirements of isolation units in hospitals (S4 Fig).
We conducted sensitivity analysis with respect to contact tracing efficacy, average number of contacts per individuals (k), and the clustering coefficient (ϕ) (S2 Text). Varying contact tracing efficacy from 5% to 50%, the predicted epidemic size without ring vaccination ranged from 3,874 to 3,988 confirmed cases in Liberia and 10,246 to 11,048 confirmed cases in Sierra Leone (Fig 2). Implementing ring vaccination with a prophylactic vaccine is expected to achieve the greatest marginal benefit when contact tracing efficacies are between 20% and 28% in both Liberia and Sierra Leone (S2 Fig, S2 Text). Specifically, the greatest marginal benefit of ring vaccination can be up to 0.29% in Liberia and up to 0.64% in Sierra Leone, depending on the vaccine efficacy (S2 Fig). However, a vaccine with post-exposure protection is estimated to provide up to 6% marginal benefit in Liberia and 8% in Sierra Leone (S2 Text, S3 Fig). These results suggest that ring vaccination provides moderate benefit to case isolation under a range of contact tracing efficacies and vaccine efficacies.
Our sensitivity analysis of the social mixing patterns (k and ϕ) show that the marginal benefit of adding ring vaccination rises with more contacts per individual or increasing clustering (S2 and S5 Figs, S2 Text).
We demonstrated that a combination of contact tracing, case isolation and ring vaccination could effectively reduce Ebola transmission. In Liberia, as well as in past Ebola outbreaks, implementing case isolation along with change in human behavior dramatically reduced transmission [36–39], which commonly occurs when people perceive infection risk associated with specific behaviors. Our results suggest that in this current context of reduced transmission, ring vaccination offers only moderate additional benefit, which is consistent with findings from a previous cholera model [40]. Although the incremental benefit of ring vaccination could be relatively moderate, the marginal benefit of ring vaccination is expected to increase with the average number of contacts per individual, the clustering coefficient or lower contact tracing efficacies. Therefore, ring vaccination is particularly useful in regions where contact tracing is logistically challenging. In addition, a vaccine that can confer post-exposure efficacy is much more effective for use in ring vaccination strategies than a vaccine that can only confer protection prophylactically. An additional benefit of ring vaccination that is more pronounced for a vaccine that confers post-exposure protection than for an exclusively prophylactic vaccine is the reduction in the number of symptomatic individuals who need hospitalization.
The 2014 Ebola outbreak in West Africa has affected both rural and urban communities [41], which differ in accessibility for contact tracing and hospital facilities for case isolation. There is also regional variation in the number and clustering of contacts [42]. Our analysis demonstrates that the benefit of adding ring vaccination to case isolation is influenced by the extent of clustering in a population. Specifically, we found ring vaccination provided the greatest marginal benefit to case isolation in highly clustered populations, such as crowded homes or schools [23–26, 43]. As social networks can be highly complex systems consisting of densely connected communities [44–46], future studies should evaluate not only transmission patterns within communities but also transmission patterns between rural and urban areas in order to tailor ring vaccination to specific locations.
Clinical trials are underway to evaluate Ebola vaccine candidates, but there is currently considerable uncertainty regarding the efficacy post-exposure and prophylactically. Our analysis of vaccine efficacy, ranging between 5–100%, demonstrates a diminishing marginal contribution with declining vaccine efficacy. For example, at a hypothetical efficacy of 50% ring vaccination could have averted up to 214 cases, whereas a vaccine with an efficacy of 75% up to 360 cases of Ebola could have been averted.
Our modeling approach is based on a deterministic approximation to a stochastic network model. Consequently, our model did not capture the stochastic effects that become pronounced in the eradication phase of an outbreak [6] and for ring vaccination strategies that include second-order ring vaccination, where contacts of exposed individuals are vaccinated in addition to the exposed contacts of the isolated case. We would expect an increase in the marginal benefit of ring vaccination if both first-order and second-order vaccination were implemented because vaccination would be administered ahead of the wave of transmission. Thus, our results are conservative with regard to second-order ring vaccination.
Ebola has similar family and household transmission as smallpox. During the final eradication phase of smallpox, public health authorities relied on case isolation and ring vaccination [18, 19, 47, 48]. Because smallpox is more infectious than Ebola [2, 49, 50], we conservatively expect that case isolation and ring vaccination would likewise be a practical approach to eliminate Ebola and contain future outbreaks, especially with limited vaccine supplies.
The combined implementation of ring vaccination and case isolation can be an effective approach in curtailing an Ebola outbreak. Our model predicts that ring vaccination offers moderate benefit to case isolation, with the greatest benefit occurring where there are more contacts per individual or greater clustering among individuals, and when contact tracing has low efficacy or vaccination can confer post-exposure protection.
|
10.1371/journal.pntd.0007219 | Fast diagnosis of sporotrichosis caused by Sporothrix globosa, Sporothrix schenckii, and Sporothrix brasiliensis based on multiplex real-time PCR | The accurate diagnosis of sporotrichosis and identification at the species level are critical for public health and appropriate patient management. Compared with morphological identification methods, molecular diagnostic tests are rapid and have high sensitivity and standardized operating processes. Therefore, we designed a novel multiplex real-time polymerase chain reaction (PCR) method based on the calmodulin (CAL) gene for the identification of clinically relevant Sporothrix species: S. globosa, S. schenckii s. str., and S. brasiliensis. We evaluated the assay with clinical and spiked samples and assessed its diagnostic performance by comparing the results to those of culture and species-specific PCR. Thirty-three DNA templates were used to detect assay specificity, and three plasmids were constructed to create a standard curve and determine the limits of detection (LODs). For nucleic acid detection, the sensitivity and specificity reached 100%. The LODs were 10 copies, 10 copies and 100 copies for S. globosa, S. schenckii s. str and S. brasiliensis, respectively. For the clinical samples, the positive detection rates by culture, species-specific PCR and the multiplex real-time PCR assay were 87.9% (29/33), 39.4% (13/33), and 93.9% (31/33), respectively. For the spiked samples, the positive detection rates were both 100% for S. schenckii s. str and S. brasiliensis. Based on the above results, compared with culture and other molecular diagnosis methods, the novel multiplex real-time PCR assay is effective, fast, accurate, and highly sensitive. It has a lower reaction cost and lower sample volume requirements, can detect co-infections, and allows for standardized operation and easier interpretation of results. In the future, this assay could be developed into a commercial kit for the diagnosis and identification of S. globosa, S. schenckii s. str, and S. brasiliensis.
| Sporotrichosis is a subacute or chronic infectious disease caused by dimorphic fungi of Sporothrix spp. The genus Sporothrix consists of several species with different geographic distributions, virulence, and antifungal susceptibilities, making species-level identification necessary. S. brasiliensis, S. globosa, S. schenckii s. str and S. luriei make up the “pathogenic clade” of the genus Sporothrix. Importantly, S. luriei has a low clinical-epidemiological impact within this genus. Therefore, we designed a novel multiplex real-time PCR method using fluorescent probes for the identification of S. globosa, S. schenckii s. str, and S. brasiliensis. We designed a pair of primers based on the conserved sequence of the calmodulin gene of Sporothrix spp. and probes with different fluorescent signals based on the divergent sequences of S. globosa, S. schenckii s. str and S. brasiliensis. Through the verification of nucleic acid, clinical and spiked sample detection, the multiplex real-time PCR could quickly and accurately identify the three clinically relevant species of Sporothrix spp. with high sensitivity. This new assay could be applied in epidemiology, clinical diagnosis, and experiments with sporotrichosis to control new outbreaks, reduce diagnostic and identification time, and improve test efficiency.
| Sporotrichosis is a subacute or chronic infectious disease caused by dimorphic fungi of Sporothrix spp., which are distributed worldwide, especially in tropical and subtropical regions[1–2]. The infection generally occurs through traumatic inoculation of contaminated plant debris[3–4] or through bites and scratches from infected animals, mostly felines[5–6]. The disease can occur sporadically or in outbreaks[7–8]. The primary lesion is usually restricted to the skin and subcutaneous tissue but can subsequently affect adjacent lymphatic vessels[9]. Rarely, this fungus can disseminate through the blood or the lymphatic system and eventually lead to a systemic infection[3]. The pathogen of sporotrichosis, S. schenckii sensu lato, was recognized as the sole agent for more than a century following its first isolation in 1898 by Benjamin Schenck[10]. However, based on macroscopic characteristics and physiological and molecular aspects, S. schenckii sensu lato is considered to be a complex of several distinct species, including S. brasiliensis, S. mexicana, S. globosa, and S. schenckii s. str.[11], S. luriei[12], S. pallida[13–14], and S. chilensis[15]. Further, S. globosa, S. schenckii s. str, S. brasiliensis are medically important species within the Sporothrix genus. These species differ in ecology, distribution and epidemiology. Furthermore, widespread variations in virulence and antifungal susceptibility among these species have been demonstrated. S. brasiliensis, which is associated with severe clinical forms of sporotrichosis[16], is considered the most virulent species, followed by S. schenckii s. str. and S. globosa. Therefore, identification at the species level is critical for public health and appropriate patient management[17,18].
Sporotrichosis can be diagnosed through a combination of clinical manifestation and epidemiological and laboratory tests, including direct examination, culture, histopathological examination, molecular detection, sporotrichin skin tests and antibody detection[18]. It is difficult to detect the parasitic budding yeast cells by direct examination, likely because they are too small (2 to 6 μm in diameter). Yeast cells can be observed in tissue by histopathological examination using Haematoxylin and eosin (H&E), Gomori methenamine silver (GMS) or Periodic acid-Schiff (PAS) stain[1]. However, neither direct examination nor histopathological examination can identify the pathogen at the species level. The “gold standard” for diagnosing sporotrichosis is based on conventional culture of clinical specimens obtained from active lesions, pus, secretions or biopsies. After culture on Sabouraud agar (SDA) for 5 to 7 days at 28°C, filamentous hyaline colonies start to grow and then develop a dark colour, especially from the centre of the colonies[19] Positive cultures provide the strongest evidence for sporotrichosis, and isolates obtained from culture can be tested for antifungal susceptibility and phenotypic characterization.
With the development of molecular biology, an increasing number of methods based on nucleic acid detection have been applied for the rapid diagnosis of infectious disease. Many molecular diagnostic tests, including DNA sequencing of “barcoding” genes[20–23], nested PCR[24–25], PCR fingerprinting[26], restriction fragment length polymorphism (RFLP) of different gene targets[27], random amplified polymorphic DNA (RAPD)[7], amplified fragment length polymorphism (AFLP) [8], rolling circle amplification (RCA) [28] and species-specific primers[29], have been developed for Sporothrix spp. detection. However, there are still some shortcomings, such as time-consuming procedures (PCR sequencing for at least 12 h); complicated operation steps, which can increase the chance of contamination (nest PCR); insufficient sensitivity (RCA, 3 × 106 copies); and so on. Most of them can only identify isolates from culture, and only a few methods have been evaluated with clinical samples[24–25]. In addition, none of the above methods can detect co-infection simultaneously.
In the present study, we developed a novel multiplex real-time PCR assay to identify the mainly clinical pathogenic species S. globosa, S. schenckii s. str. and S. brasiliensis, and we evaluated the assay with clinical and spiked samples.
The analytical specificity was examined using 33 DNA templates, including from fungi (28), bacteria (3), a human (1) and a mouse (1) (S1 Table). None of the 33 templates yielded positive signals in the assays; furthermore, nonspecific amplification was not detected. Excluding the negative controls, all 25 Sporothrix templates, including S. globosa (21), S. schenckii s. str. (3) and S. brasiliensis (1), yielded positive signals, and the positive detection rate for the nucleic acid templates of the assay was 100%. Standard curves (Ct vs. log CFU) for S. globosa, S. schenckii s. str. and S. brasiliensis were constructed using the plasmid DNA template by serial 10-fold dilution (S1 Fig). In addition, the results indicated that the LODs of this assay were 10 copies, 10 copies and 100 copies for S. globosa, S. schenckii s. str. and S. brasiliensis, respectively.
For the mixed templates, the multiplex real-time PCR assay could detect the corresponding fluorescent signals in the same tube without nonspecific amplification. The Ct values of different amount templates are shown in Table 1.
The Ct values of the single fluorescence real-time PCR for S. globosa, S. schenckii s. str., and S. brasiliensis were 21.67±0.15, 24.64±0.15, and 27.00±0.11, respectively, while under the same templates and condition, the Ct values obtained from the multiplex real-time PCR were 21.80±0.07, 24.77±0.07, and 27.32±0.08, respectively. There was no significant difference in Ct values between multiplex and single fluorescence real-time PCR except for S. brasiliensis (t-test, p = 0.03).
A total of 40 samples from patients suspected of sporotrichosis were collected (S2 Table). Seven samples were eliminated based on culture and histopathological examination results. The results of the multiplex real-time PCR and species-specific PCR were negative for these 7 samples. Of the 33 selected samples, the positive detection rates of the culture, species-specific PCR and multiplex real-time PCR assays were 87.9% (29/33), 39.4% (13/33), and 93.9% (31/33), respectively (Table 2). The positive detection rates of the culture and multiplex real-time PCR assays were not significantly different (paired χ2, p = 0.4142). Differences were observed between the multiplex real-time PCR assay and species-specific PCR (paired χ2, p<0.0001).
The isolates from culture were identified by sequencing the CAL gene, and all 29 strains were S. globosa. Among the 33 samples detected by the multiplex real-time PCR assay, the Ct values of 31 were less than 40 and were judged as positive; the Ct values of 2 samples were greater than 40 and were judged as negative. Only FAM fluorescence was detected; therefore, all of the samples were identified as S. globosa infections, consistent with the sequencing results from the cultured isolates. Among the 33 samples, 11 were positive by all three methods, and 27 were positive by both culture and multiplex real-time PCR. In addition, culture was positive and multiplex real-time PCR was negative for 2 samples, and culture was negative and multiplex real-time PCR was positive for 4 samples, of which two were also positive by species-specific PCR.
No false positives were detected from the negative control group of spiked samples, and the positive detection rates of S. schenckii s. str., and S. brasiliensis were both 100% (6/6), while the Ct values were 33.03–38.57 (S. schenckii s. str.) and 30.23–34.84 (S. brasiliensis).
In 2006, Marimon et al. [21] reported Sporothrix complex for the first time, which led to many studies of the differences between species in the Sporothrix genus[16–17,22,30–32]. These studies showed the variations between different species and highlighted the need to identify the species level of Sporothrix spp.. Calmodulin (CAL), internal transcribed spacer (ITS), and elongation factor (EF), which are recognized as fungal "barcoding" genes, are widely applied in the identification of Sporothrix spp.[20–23]. However, all of the methods are based on conventional PCR. Compared to conventional PCR, real-time PCR has many extraordinary advantages, such as rapidity, sensitivity and low risk of contamination. Due to these strong points, real-time PCR is widely used in pathogen detection. However, the application of real-time PCR to Sporothrix spp. identification has not yet been reported [18]. In this study, by comparing sequences of the various “barcoding” genes, a target sequence, which can be used to design the primers and probes for three pathogenic species, was found on the calmodulin gene of Sporothrix spp.. Based on this finding, we established a multiplex real-time PCR to detect S. globosa, S. schenckii s. str and S. brasiliensis simultaneously, which not only can improve the sensitivity of Sporothrix spp. detection but also could save detection time and costs.
Furthermore, the detection ability of co-infections was assessed by 9 combinations of different amounts of plasmids. The results showed that, when the amounts were different, the amplification of the smaller amount plasmid was affected, and because of the competitive inhibition, the greater the difference in the proportion of plasmids was, the greater the effect was on the Ct value of the small amount of plasmid. For the same template, the Ct values were not significantly different between multiple and single fluorescence for S. globosa and S. schenckii s. str. Only the intensity of the fluorescent signal was weakened, but this decrease did not affect the judgement of the result. However, a difference in Ct value was observed for S. brasiliensis (p = 0.03), likely due to the weak amplification efficiency compared to the other two. Considered together, these results demonstrated that the assay was capable of detecting co-infections and that there was almost no mutual interference between the primers and probes in the multiplex system.
In addition, the clinical samples were used to evaluate the performance of the assay compared with the “gold standard” diagnostic and the latest molecular diagnosis methods of culture and species-specific PCR. Culture, as the “gold standard” for sporotrichosis diagnosis, is widely used in clinical practice. However, species-level identification requires further phenotypic identification and physiological tests, which require at least 3–4 weeks. During this period, contamination by fast-growing fungi or bacteria is likely happening. Some patients undergoing antifungal treatment can also have negative culture results. Therefore, histopathological examination results, such as detection of Sporothrix spp. yeast cells or typical pathological structures, are also combined with culture results for clinical diagnosis. In this experiment, these diagnostic criteria were followed when the clinical samples were collected. Results were considered negative only if the culture and histopathological examinations both excluded the diagnosis. In the present study, the positive rate of culture was 87.9% (29/33). Of the 4 negative results, 2 samples had contamination from other microorganisms and were judged to be negative, while the multiplex real-time PCR and species-specific PCR results of the 2 samples were both positive. Another 2 samples did not grow after 30 days in culture, while the multiplex real-time PCR results were positive. It is speculated that these 2 cases might have been treated with antifungal therapy or affected by the location of the biopsy. Based on the different lengths of the CAL gene sequences, species-specific PCR[29], reported by Rodrigues et al., could identify S. brasiliensis, S. schenckii s. str, S. globosa, S. mexicana, S. pallida, and its relative, Ophiostoma stenoceras. The reaction conditions for this species-specific PCR included 35 cycles, and amplification could not be detected after 35 cycles. In this experiment, according to the results of the multiplex real-time PCR, which was performed for 45 cycles, most of the Ct values were greater than 35 (22/33). In this situation, the results of species-specific PCR were often negative.
Until now, no reports of S. brasiliensis have appeared, and there have been only four known isolates of S. schenckii s. str. found in China[33]. To compensate for the singularity of pathogens in clinical specimens, an evaluation of the spiked samples was performed. Yeast cells, the pathogenic phase of Sporothrix spp., were selected for mixing with negative tissue samples, and positive results were obtained from all of the spiked samples, which were mixed with different numbers of yeast cells. In addition, the standard curves of assay were established on direct dilution of plasmids, whereas the clinical samples were prepared using DNA extraction kit. Since DNA lost is unavoidable during the extraction process, the sensitivity of the multiplex real-time PCR was applied to the extracted DNA samples, not the original clinical samples. Considering that different extraction methods result in varying degrees of DNA lost, it is recommended to use a method with high yield of DNA for clinical samples extraction.
In conclusion, the novel multiplex real-time PCR assay was effective, fast, accurate, and highly sensitive. It had a lower reaction cost and sample volume requirements, could detect co-infections and allowed for standardized operation and easier interpretation of results. However, the assay must still be validated with clinical samples of S. schenckii s. str and S. brasiliensis. In the future, the number of clinical samples used to validate the assay must be increased, and the assay could be further verified using pus or secretions from active lesions as the templates.
Twenty-five Sporothrix spp. isolates (including 21 S. globosa, 3 S. schenckii s. str, and 1 S. brasiliensis), twenty-eight other fungal strains (including agents of superficial, subcutaneous, and systemic mycoses in humans and animals), three bacterial strains, one human genomic DNA sample and one BALB/c-mouse genomic DNA sample were used to develop the PCR assays (S1 Table). The fungi were obtained from the Collection of Pathogenic Fungi at the Research Centre for Medical Mycology, Peking University (BMU, Beijing, China), the bacterial DNA were obtained from the National Institute for Communicable Disease Control and Prevention, Chinese Centre for Disease Control and Prevention (Beijing, China), the human DNA was obtained from a healthy volunteer, and the mouse DNA was obtained from a BALB/c mouse. All of the fungal strains were previously characterized at the species level via morphological analysis and sequence analysis of the rDNA operon (ITS1-5.8S-ITS2) and the CAL gene.
A total of 40 tissue biopsies were collected between September 2017 and August 2018; the clinical data are shown in S2 Table. These samples were collected from patients in the Dermatology Department of the Second Hospital of Jilin University, whose clinical manifestations indicated suspected sporotrichosis. The clinical symptoms of the subjects were examined by professional physicians. In addition, 6 other negative human samples were collected from different volunteers and were used as artificially contaminated samples to simulate clinically infected specimens of S. schenckii s. str and S. brasiliensis. All of the specimens were skin and subcutaneous tissue harvested by surgery. When the specimens were collected, informed consent was obtained based on the guidelines and agreements of the institutional ethics committee.
All of the fungal isolates were subcultured on 2% potato-dextrose agar (PDA) slide medium at 28°C for 7–14 days. All of the tissues were cut into small pieces with sterile scissors, and then all of the pieces were placed in liquid nitrogen and ground thoroughly with a mortar and pestle. DNA was extracted and purified with the QIAamp DNA Mini Kit (QIAGEN, Hilden, Germany) in accordance with the manufacturer’s instructions; detailed steps are provided in the supplemental methods (S1 Text). The quality of the extracted DNA was assessed by amplification of part of the rDNA operon or β-globin using universal primers[34]. The amplified products were visualized via agarose gel electrophoresis and UV detection. A single amplification product indicated that the sample was free of PCR inhibitors.
Three plasmids were constructed to create a standard curve and to determine the LODs of the multiplex real-time PCR assay. The CAL regions of S. globosa (BMU 09028), S. schenckii s. str (CBS498.86T) and S. brasiliensis (CBS 120339T) were cloned into pMD-18T vectors (Takara, Dalian, China). The plasmids were transformed into E. coli DH5a competent cells, and the cells that contained recombinant plasmids were cultivated in lysogeny broth for 24 h. The plasmids were then extracted from the cultured E. coli suspension with a Qiagen Plasmid Mini Kit (QIAGEN, Hilden, Germany). Plasmid concentrations were measured with a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific, USA), and the copy numbers of the plasmids were calculated from their total base lengths and DNA concentrations using the equation of Godornes et al. [35]. The DNA samples and plasmids were stored at -20°C until use.
All of the CAL sequences belonging to Sporothrix spp. were selected from the National Center for Biotechnology Information (NCBI) database to develop specific primers targeting the conserved sequence of Sporothrix spp. and probes marked by different fluorescent signals targeting the divergent sequences of S. globosa, S. schenckii s. str and S. brasiliensis (details in Table 3). Primer Express software (version 3.0; Life Technologies-Applied Biosystems) was used to design the primers and probes and to evaluate melting temperatures, GC content, dimers, and mismatches in the candidate primers and probes.
Each PCR mixture consisted of 2.5 μL of 10x Platinum Buffer (Life Technologies-Invitrogen), 4.0 μL of MgCl2 (50 mM), 0.2 μL of each primer (25 μM), 0.1 μL of each probe (25 μM), 0.25 μL of Platinum Taq DNA polymerase (5 U/μL; Life Technologies-Invitrogen), 1.5 μL of PCR nucleotide mix (10 mM), 5 μL of DNA template, and nuclease-free water to achieve a final volume of 25 μL. Multiplex real-time PCR was performed in a CFX96 Real-time PCR Detection System (Bio-Rad, Hercules, CA, USA) under the following conditions: predenaturation at 95°C for 3 min, followed by 45 cycles of 95°C for 15 s and 60°C for 30 s. The data were analysed with CFX Manager software (version 3.1; Bio-Rad).
The analytical specificity of the assays was tested by analysing 33 DNA samples derived from other pathogenic fungi, bacteria and tissues from a human and a mouse. The analytical sensitivity, standard curves and LODs of the assays were determined by using three 10-fold dilutions of the previously constructed plasmids, ranging from 2.0×105 copies/μL to 0.2 copies/μL. S. globosa, S. schenckii s. str and S. brasiliensis were detected by FAM, VIC, and CY5 fluorescence, respectively. The detection limit was noted for each probe. Each dilution of the plasmids was assayed in triplicate.
The detection ability for a mixed template was determined by analysing the Ct values from 9 compositions of plasmid mixtures. The amount of detected plasmid was set at one gradient larger than LOD (i.e., 100 copies for S. globosa; 1000 copies for S. brasiliensis; 100 copies for S. schenckii s. str). The amount of the other two plasmids were 10-fold, 100-fold, and 1000-fold greater than that of the detected plasmid. Each template mixture was comprised of three plasmids of different proportions (details in Table 1). The mixed templates were detected by multiplex real-time PCR. The obtained Ct values were compared with those of single plasmid detection in the same reaction system under the same conditions. The detection ability of the multiplex and single fluorescence assays was tested by comparing the Ct values from four different reaction systems (one multiplex and three single fluorescence values of FAM/VIC/CY5) with the same templates under the same conditions. Each reaction was assayed in triplicate.
A total of 40 specimens from biopsies were collected, and each was divided into three parts. One part was used for culture (PDA, 28°C), one for histopathological examination (HE and PAS) and one for DNA extraction. The performance of the multiplex real-time PCR assay was evaluated by comparison with the culture method and the species-specific PCR[29]. After 4 weeks of culture, no fungal growth or growth of contaminating microorganisms was judged as negative[18]. The results of the histopathological examination showed a mixed suppurative and granulomatous inflammatory reaction in the dermis and subcutaneous tissue, and the detection of asteroid bodies or Sporothrix spp. yeast cells by PAS was suggestive of sporotrichosis. The clinical diagnosis was made by combining the results of the culture and histopathological examination.
To evaluate the assay detection ability for S. schenckii s. str and S. brasiliensis infectious samples, we used artificially contaminated (spiked) samples to simulate clinically infected specimens. Each of the 6 negative human samples was divided into three parts, two of which were used to simulate infected specimens and one of which was used for negative controls. S. schenckii s. str (CBS498.86T) and S. brasiliensis (CBS 120339T) were subcultured on brain heart infusion (BHI) agar medium and incubated at 35°C for 7 days to obtain the yeast cells of Sporothrix spp. The yeast cell suspensions of S. schenckii s. str and S. brasiliensis were prepared with sterile saline solution, and the OD was adjusted at 520 nm to 0.2, approximately corresponding to a concentration of 106 cells/mL.[28]. Then, 10, 50, 100 μL each of the suspensions were mixed with 2 different negative human samples, respectively. DNA extraction procedures were the same as described above.
A no-template control (NTC), a negative control (NEG) and a positive control (POS) were established for each test of the multiplex real-time PCR assay. When the amplification result showed NTC (-), NEG (-), and POS (+), the test was considered a valid amplification. A Ct value from the valid amplification of less than 40 was judged as positive; otherwise, it was negative. The species-specific PCR was performed according to the literature[29], and a single clear band shown by gel electrophoresis and UV detection was judged as positive; otherwise, it was negative.
The quantitative data are presented as the mean ± standard deviation (SD). The differences in Ct values between multiplex and single fluorescence were tested with the independent samples t-test. The differences in positive detection rates between the multiplex real-time PCR and culture and between the multiplex real-time PCR and species-specific PCR were tested with the paired chi-square test. Statistical significance was defined as a p value <0.05. All of the calculations were performed with the Statistical Analysis System software package (version 9.3; Cary, NC, USA).
This study was performed in strict accordance with recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Ethics in Research Committee of the Second Hospital of Jilin University, protocol number 2018–018. Written informed consent was obtained from patients or at least one guardian of the patient before enrolment.
|
10.1371/journal.pbio.1000131 | Management Effectiveness of the World's Marine Fisheries | Ongoing declines in production of the world's fisheries may have serious ecological and socioeconomic consequences. As a result, a number of international efforts have sought to improve management and prevent overexploitation, while helping to maintain biodiversity and a sustainable food supply. Although these initiatives have received broad acceptance, the extent to which corrective measures have been implemented and are effective remains largely unknown. We used a survey approach, validated with empirical data, and enquiries to over 13,000 fisheries experts (of which 1,188 responded) to assess the current effectiveness of fisheries management regimes worldwide; for each of those regimes, we also calculated the probable sustainability of reported catches to determine how management affects fisheries sustainability. Our survey shows that 7% of all coastal states undergo rigorous scientific assessment for the generation of management policies, 1.4% also have a participatory and transparent processes to convert scientific recommendations into policy, and 0.95% also provide for robust mechanisms to ensure the compliance with regulations; none is also free of the effects of excess fishing capacity, subsidies, or access to foreign fishing. A comparison of fisheries management attributes with the sustainability of reported fisheries catches indicated that the conversion of scientific advice into policy, through a participatory and transparent process, is at the core of achieving fisheries sustainability, regardless of other attributes of the fisheries. Our results illustrate the great vulnerability of the world's fisheries and the urgent need to meet well-identified guidelines for sustainable management; they also provide a baseline against which future changes can be quantified.
| Global fisheries are in crisis: marine fisheries provide 15% of the animal protein consumed by humans, yet 80% of the world's fish stocks are either fully exploited, overexploited or have collapsed. Several international initiatives have sought to improve the management of marine fisheries, hoping to reduce the deleterious ecological and socioeconomic consequence of the crisis. Unfortunately, the extent to which countries are improving their management and whether such intervention ensures the sustainability of the fisheries remain unknown. Here, we surveyed 1,188 fisheries experts from every coastal country in the world for information about the effectiveness with which fisheries are being managed, and related those results to an index of the probable sustainability of reported catches. We show that the management of fisheries worldwide is lagging far behind international guidelines recommended to minimize the effects of overexploitation. Only a handful of countries have a robust scientific basis for management recommendations, and transparent and participatory processes to convert those recommendations into policy while also ensuring compliance with regulations. Our study also shows that the conversion of scientific advice into policy, through a participatory and transparent process, is at the core of achieving fisheries sustainability, regardless of other attributes of the fisheries. These results illustrate the benefits of participatory, transparent, and science-based management while highlighting the great vulnerability of the world's fisheries services. The data for each country can be viewed at http://as01.ucis.dal.ca/ramweb/surveys/fishery_assessment.
| Fisheries play an important role in the global provision of food, directly accounting for at least 15% of the animal protein consumed by humans and indirectly supporting food production by aquaculture and livestock industries [1],[2]. Demand for fish is expected to grow given escalating animal protein demands in developing countries and the rapidly increasing human population [1]–[4]. However, reported global marine fisheries landings have declined by about 0.7 million tonnes per year since the late 1980s [5], with at least 28% of the world's fish stocks overexploited or depleted, and 52% fully exploited by 2008 [1]. Severe reductions in abundance can change population genetic structure [6], harm the recovery potential of stocks [7], trigger broader ecosystem changes (e.g., [8]–[10]), threaten livelihoods [1], and endanger food security [11] and efforts towards the reduction of hunger [11],[12]. Given the different ecological and socioeconomic consequences of a global fisheries crisis, a number of international efforts have sought to improve management in the hope of moving towards sustainable marine fisheries (sensu Pauly et al. [13]). Some of these initiatives, which incorporated to varying degrees the improvement of marine fisheries management, include the United Nations Code of Conduct for Responsible Fisheries from the Food and Agriculture Organization [14], the Convention on Biological Diversity (http://www.cbd.int/), and the Millennium Ecosystem Assessment (http://www.millenniumassessment.org). Although these initiatives have received broad acceptance, the extent to which corrective measures are implemented and effective remains poorly known [15]–[17]. Using a survey approach, validated with empirical data and enquiries to fisheries experts, we quantified the status of fisheries management in each nation worldwide that has an exclusive economic zone (EEZ). We also related our measurements of management effectiveness to a recently developed index of fisheries sustainability. To our knowledge, these results represent the first global assessment of how fisheries management attributes influence sustainability, while providing a baseline against which future changes can be quantified.
We evaluated the effectiveness of national fisheries management regimes by quantifying their degree of compliance with a well-recognized set of conditions necessary for sustainable fisheries: (1) robust scientific basis for management recommendations, (2) transparency in turning recommendations into policy, (3) capacity to enforce and ensure compliance with regulations, and minimizing the extent of (4) subsidies, (5) fishing overcapacity, and (6) foreign fishing in the form of fisheries agreements [8],[14]. The extent to which individual countries met or were affected by these conditions was quantified using a set of normative questions assembled in an Internet survey, which was systematically distributed to fisheries experts worldwide. Over 13,000 experts were contacted as part of this survey, of which 1,188 responded from each country bordering the ocean (i.e., EEZ; see Materials and Methods for additional details on areas surveyed). Experts were mostly fisheries managers, university professors, and governmental and nongovernmental researchers. Despite these diverse backgrounds, responses were highly consistent within each country (i.e., where multiple responses were given, 67% of experts chose the same answer to any given question and 27% chose the next closest response; Figure 1A and 1B) and in accordance with independent empirical data (we found a strong correlation between experts' opinions and empirical data [r = 0.74, p<0.00001, n = 28 countries; Figure 1C]). Justification, extended results, and discussion on the reliability and validity of the experts' data are presented in Materials and Methods. We also used a Monte Carlo simulation approach to include score uncertainty estimates in the results. We provide the main results and general conclusions in the text; full results are presented in Figures S1, S2, S3, S4, S5 and http://as01.ucis.dal.ca/ramweb/surveys/fishery_assessment/.
Critical to the success of fisheries management is the scientific basis on which management recommendations are made [18],[19]. Preventing the collapse of fisheries and ecosystem-wide impacts requires scientific advice in which uncertainty is minimized by using skilled personnel, models that include, not only the dynamics of fished stocks, but also their embedded ecosystems, and high-quality and up-to-date data (such that reliable recommendations can be adapted as conditions and stocks fluctuate). Alternatively, the effects of uncertainty can be minimized by applying precautionary approaches in the face of limited knowledge [18],[20]. Of the world's 209 EEZs analyzed, 87% have scientific personnel who are qualified (e.g., with Ph.D.- or Masters-level education, or have participated in training courses or relevant conferences) to perform fisheries assessments and provide science-based management advice (Figure S1A), approximately 7% use holistic models as the basis of management recommendations (i.e., including a broad set of biological and environmental data on fisheries to enable ecosystem-wide understanding of fisheries drivers and impacts; see Figure S1B), 61% carry out frequent assessments to ensure the effectiveness of existing management measures (Figure S1C), and 17% implement precautionary approaches for at least some species (Figure S1D). We summarized all responses that pertain to “scientific robustness” on a linear scale using multidimensional scaling. (Multidimensional scaling is an ordination method that uses the similarities and dissimilarities among responses to reduce the number of variables analyzed. This facilitates the assessment and visualization of patterns from several dimensions into one. Very simplistically, this is analogous to calculating an average of the different scores for each country; see Materials and Methods.) The resulting scale ranged from 0 to 1, and we divided it into four quarters (i.e., from 0 to 0.25, from 0.25 to 0.5, from 0.5 to 0.75, and from 0.75 to 1, with the lowest quarter indicating the worst combination of attributes and the top the best). We found that 7% of all EEZs rank in the top quarter of such a scale (Figure 2, countries depicted in Figure 3A), which account for approximately 9% of the world's fisheries catches and approximately 7% of the world's fished stocks (data are for 2004; see details in Figure S2). Distinguishing between high- and low-income countries using per capita Gross Domestic Product (i.e., 2007 per capita Gross Domestic Product larger or smaller than US$10,000, respectively), we found that high-income countries ranked significantly higher on the scale of scientific robustness (Mann-Whitney U test: p<0.00001, Figure S1E).
We note that a recent study indicated the success of catch shares, as individual transferable quotas, in preventing fisheries collapses [21]. This strategy has been implemented primarily in the EEZs of New Zealand, Australia, United States, Iceland, Chile, and Peru, which are all countries with robust scientific capabilities (Figure 3A). Our results indicate that the global adoption of individual transferable quotas should be considered with caution given that their underlying success rests on the scientific robustness of the implemented quotas and that few countries meet that condition (Figure 3A).
Guidelines to improve the acceptance and compliance with fishing regulations recommend that decisions be based on the best available scientific evidence and follow a transparent and participatory process [8],[14],[22],[23]. Unfortunately, the process of policymaking can be subjected to substantial political pressures, perhaps including corruption. In our survey, management authorities from 92% of the EEZs consider scientific recommendations in formulating policies (Figure S1F), and in 87%, all stakeholders are consulted or their opinions considered (Figure S1G). Yet in 91% of all EEZs, regulations commonly face economic or political pressures to increase allowable catches or to implement regulations that err on the side of risk rather than caution (Figure S1I), whereas a surprising 83% of EEZs are thought to face corruption or bribery (Figure S1H). Of all EEZs, 26% rank in the top quarter of a scale of “policymaking transparency,” which summarizes, through multidimensional scaling, the attributes of considering scientific advice, participation, pressures, and corruption (Figure S1J, countries depicted in Figure 3B). Only 1.4% of all EEZs are in the top quarter on the combined scales of scientific robustness and policymaking transparency (Figure 2), which together accounted for 0.85% of the world's fisheries catch and 1.1% of the world's fished stocks (Figure S2). There were no significant differences between low- and high-income countries with respect to policy transparency (Figure S1J). However, the underlying mechanism was different, with low-income countries facing more corruption (p<0.00001, Figure S1H) and less commonly incorporating scientific advice (p<0.005, Figure S1F), whereas high-income countries faced slightly more political pressures (p<0.05, Figure S1I).
One of the biggest challenges in fisheries management lies in the implementation and enforcement of regulations [23]. Poverty, unemployment, available infrastructure for control and surveillance, the severity of penalties for violations, and participation in policymaking are all likely influencing the level of compliance with regulations. Proper enforcement through (1) adequate funding and equipment for the managing authorities, (2) patrolling of fishing grounds, and (3) tough penalties for infringements, occurs in 17% of all EEZs (Figure S1K; note that only ∼6% of all EEZs impose penalties that are sufficiently tough to deter violators). Not surprisingly, no EEZ was free of the effects of poaching (Figure S1L, see also [24]). On a scale of “implementation capability,” which summarizes, through multidimensional scaling, poaching and the different attributes of enforcement, we found that only approximately 5% of all EEZs are in the top quarter of such a scale (Figure S1M, countries depicted in Figure 3C). Only two relatively small EEZs, those of the Faeroe and Falkland Islands, were in the top quarter for all three indicators of scientific robustness, policymaking transparency, and implementation capability (Figure 2), which combined, accounted for 0.80% of the world's fisheries catch and 0.48% of the world's fished stocks (Figure S2). Better “implementation capability” is frequently more common among high- than low-income countries (p<0.0001, Figure S1M), which is mainly a consequence of better enforcement (p<0.00001, Figure S1K) and reduced poaching in the former (p<0.002, Figure S1L).
When the structure of a management regime is weak, fisheries will be prone to overexploitation due to several factors. Three that have received particular attention are fishing capacity, subsidies, and access to foreign fishing fleets [8],[23],[25],[26]. Open access to fishing (because of lack of effective management) leads to a “race for fish” that commonly increases fleet size and fishing power. This should reduce fish stocks, at which point fishing capacity should stabilize given decreasing profits from reduced catches [8]. Subsidies can override this mechanism by keeping fisheries profitable and encouraging overexploitation [8],[13]. The picture is further complicated by fisheries agreements that allow foreign fleets to catch fish that are not caught by national fleets [25],[26]. Unfortunately, such agreements are commonly made between developing coastal and island states (often with low capacity to assess stocks and to enforce regulations) and developed and heavily subsidized nations [25]. Recent analyses of current agreements indicate a high risk of overexploitation due to several reasons, including selling fishing rights on highly migratory stocks under bilateral agreements, selling access rights without specified catch limits, excessive by-catch, and distortion of reported catches, among others [25],[26]. Such agreements are thought to develop coastal economies through monetary gains and local employment. In certain instances, revenues are also used to generate management plans; their effectiveness, however, is unclear given chronic weaknesses in fisheries governance and management systems [25].
Our assessment of the extent of fishing capacity, subsidies, and access to foreign fishing fleets yielded the following results. We found that fleet sizes are quantified and regulated in 20% of the world's EEZs (Figure S1N), although in 93% of EEZs, fishing fleets face some level of modernization to catch fish more efficiently or cheaply (Figure S1O). Thus, although fishing capacity may be reduced in terms of fleet size, fishing power may remain constant or even increase due to technological improvements (i.e., fewer improved boats being more effective at catching fish). Effective controls on fleet size were more common among high-income than low-income EEZs (p<0.02, Figure S1N), but the former modernized their fleets more often than the latter (p<0.00001, Figure S1O). Using multidimensional scaling to summarize the results pertaining to “fishing capacity” (i.e., fleet size controls and fleet modernization), we found high-income EEZs having significantly higher fishing capacity than low-income ones (p<0.02, Figure S1P, countries depicted in Figure 3E). Fisheries sectors that rely to some degree on subsidies occurred in 91% of the world's EEZs (Figure S1Q; countries depicted in Figure 3D), and more commonly among high- than low-income EEZs (p<0.00001, Figure S1Q) (see also [27]). Access to foreign fishing is granted in 51% of all EEZs (Figure S1R, countries depicted in Figure 3F), and is more frequent in low- than high-income EEZs (p<0.00001, Figure S1R). In fact, our survey indicated that in 33% of the EEZs that are classified as low income (commonly, countries in Africa and Oceania), most fishing is carried out by foreign fleets from either the European Union, South Korea, Japan, China, Taiwan, or the United States (Figure S3). No single EEZ meets the best standards (i.e., top quarter of the scales) of scientific robustness, policymaking transparency, and implementation capability while being free of the effects of excess fishing capacity, subsidies, or access to foreign fishing (Figure 2).
The notion that industrialized fishing practices are solely responsible for the global fisheries crisis has been challenged by evidence of the significant effects of recreational and small-scale commercial or subsistence fisheries (e.g., [28],[29]). Although less intensive per unit area, small-scale and recreational fisheries can be far more extensive spatially. Small-scale and recreational fisheries are important in 93% and 76% of the world's EEZs, respectively (Figure S4), and small-scale fisheries are increasingly more predominant among low-income EEZs whereas recreational fisheries are more predominant in high-income countries (p<0.0001, Figure S4). Of the world's EEZs, 40% collect at least some data on small-scale fishing, and 13% on recreational fishing; 30% impose regulations on the size of fish caught in small-scale fishing, and 29% do so for recreational fishing, 7% regulate the number of fish caught in small-scale fishing, and 15% do so for recreational fishing, whereas 10% limit the number of fishers in small-scale fisheries, and 3% do so for recreational fishing (Figure S4). These management measures are more frequent in high- than low-income EEZs (Figure S4). Measures to regulate small-scale and recreational fishing are clearly limited and could prove detrimental to food supply and sustainability if they continue to operate outside the control of fisheries management institutions.
To provide a general overview of fisheries management effectiveness, we averaged all scores on the scales of scientific robustness, policymaking transparency, implementation capability, fishing capacity, subsidies, and access to foreign fishing. We excluded the effects of small-scale and recreational fisheries, recognizing that their lack of management would extensively reduce the scores. Only 5% of all EEZs were in the top quarter of this scale (Figure S1S, countries depicted in Figure 4), with high-income EEZs having significantly better overall management effectiveness than low-income ones (p<0.00001, Figure S1S). A sensitivity analysis indicated that the difference between high- and low-income EEZs was driven mainly by foreign fishing agreements, which disproportionally reduced the average score of low-income EEZs. Excluding foreign fishing access leads to similarly low average scores between high- and low-income EEZs (Figure S1S). Similar average scores are, however, explained by different mechanisms, namely excessive fishing capacity and subsidies in high-income EEZs and deficient scientific, political, and enforcement capacity in low-income EEZs (Figure S1).
One final question that we addressed in this study is to what extent the different attributes of fisheries management analyzed here relate to the actual sustainability of fisheries. We addressed this question using a recently developed method to quantify the probability that ecosystems are being sustainably fished (Psust). This metric assesses the probability that the ratio between the biomass losses due to fishing (i.e., total catch) expressed in primary production equivalents and the primary production of the area in which the catch was taken is sustainable (see Materials and Methods, [30],[31]). We found that this metric is particularly useful to differentiate misinterpretations in landings data when used as an indicator of fisheries status (Figure S5). The metric, for instance, differentiates between countries in which increasing landings (a possible symptom of good fisheries status) are sustainable or not, and between countries in which declining landings (a possible symptom of overfishing or enhanced management [32]) are indicative of the sustainability of fisheries or not (Figure S5). We used classification/regression tree analysis to identify the most likely management attributes that affect the probability of fisheries sustainability; we also included country wealth (i.e., the distinction between high and low income) in the classification tree to analyze differences in fisheries sustainability due to this factor.
Of all management attributes analyzed (i.e., scientific robustness, policymaking transparency, implementation capability, fishing capacity, subsidies, and access to foreign fishing) plus taking into account country wealth, we found that variations in policymaking transparency led to the largest difference in fisheries sustainability. We found that EEZs ranked in the upper best quarter on the scale of transparent policymaking (i.e., EEZs where scientific advice is considered and followed, all parties are consulted and considered, and where corruption and external economic and political pressures are minimal [see Figure S1F–S1I]) show the largest probability of having sustainable fisheries compared to EEZs ranked in any of the other three quarters (Figure 5). The probability of sustainability in policy transparent EEZs was 88% compared to 73% in others (Figure 5). We also found that subsidies have an additional negative effect on fisheries sustainability among EEZs with nontransparent policy systems. We found that the probability of fisheries sustainability in nontransparent EEZs was reduced from 78% to 67% due to the effects of even modest subsidies (Figure 5) (i.e., EEZs ranked in the first three quarters on the scale of subsidies or EEZs in which fisheries sectors are dependent minimally to almost entirely on subsidies).
The significant effect of policymaking transparency on fisheries sustainability likely relates to the fact that this particular attribute forms the core of the fisheries management process. Firstly, it determines the extent to which scientific advice will be translated into policy, whereas transparent and legitimate participation of involved parties is likely to promote compliance with regulations [22]. Our findings indicate that policymaking transparency is likely to work as a “sustainability bottleneck” through which other positive attributes of fisheries management are filtered. For instance, we found that scientific robustness did not influence the sustainability of fisheries. This may be because, in the process of policymaking, scientific advice may be overridden due to socioeconomic costs and political or corruption pressures. The recent catch quotas for Mediterranean Bluefin tuna (Thunnus thynnus) established by the International Commission for the Conservation of Atlantic Tunas may serve as an example. In this particular case, robust and well-founded scientific advice recommended to maintain catches at 15,000 tonnes per year and to close the fisheries during two spawning months; yet the policy was set at 22,000 tonnes per year, with fishing allowed during critical spawning months. This is a case in which scientific robustness may not necessarily result in sustainability due to significant pressures in the process of policymaking. We also found that variation in implementation capabilities did not have much effect on fisheries sustainability. This result can also be explained by the effect of policymaking transparency. If the policymaking process is participatory and legitimate, it is likely that even poorly enforced systems will move towards sustainability because of voluntary compliance [22]. In contrast, some systems may strongly enforce regulations, but if the regulations were flawed during the process of policymaking, good enforcement may not bring about sustainability either. If the establishment of regulations includes scientific advice and follows a participatory mechanism, it is likely that fisheries will be tightly regulated, regardless of who carries out the fishing, which may also explain the lack of significance of fishing capacity and international fisheries agreements on fisheries sustainability. This is not to say that fishing capacity and foreign fishing access do not have impacts on fisheries sustainability but rather that their effects are moderated by the policymaking process (i.e., fishing capacity and access agreements may have different effects on sustainability in situations that are tightly regulated compared to those that are not). Finally, our results indicate how deficiencies in the process of policymaking can leave fisheries vulnerable to overexploitation due to the effect of subsidies. It is known that subsidies can override possible fishing controls exerted by economic benefits (see section above on subsidies; [8],[13],[27]). We presume, however, that this effect is likely to be more pervasive in nontransparent systems given that fishing remains poorly controlled or regulated and allowed to fluctuate more freely, depending largely on subsidies.
Improvements to fisheries management have been incorporated into international initiatives, which have received broad acceptance (e.g., [14],[15]). Unfortunately, our study shows that there is a marked difference between the endorsement of such initiatives and the actual implementation of corrective measures. The ongoing decline in marine fisheries catches [5],[9],[33]–[36] and the ecological and socioeconomic consequences of a fisheries crisis call for a greater political will of countries worldwide if further fisheries declines and their wider consequences are to be prevented. Effective transfer of improved scientific capacities to policy, achieved through a transparent and participatory process, will be more important than ever in stabilizing our food supply from the sea and preventing unnecessary losses due to management deficiencies. Current projections suggest that total demand for fisheries products is likely to increase by approximately 35 million metric tonnes by 2030 (∼43% of the maximum reported catch in the late 1980s) [3],[4] and by approximately 73% for small-scale fisheries by 2025 [35]. This contrasts sharply with the 20% to 50% reduction in current fishing effort suggested for achieving sustainability [30],[36], and implies that regulators may face increasing pressures towards unsustainable catch quotas. Given that the demand for fish lies outside the control of conventional fisheries management, other national and international institutions will have to be involved to deal with poverty alleviation (inherently improving management, Figure S1) and stabilization of the world's human population (to soften fisheries demand), if pressures on management are to be prevented and sustainability achieved.
We considered factors broadly recognized as critical for the sustainable management of fish stocks (by sustainability, we mean sustainable catches and not social, economic, or institutional sustainability and the like, which at times are also associated with fisheries management and often dominate policy decisions). The factors considered in the present analysis were categorized into those related to the robustness of scientific recommendations, transparency in the process of converting recommendations into actual policy, the capability to enforce and ensure compliance with regulations, and the extent of fishing capacity, subsidies, and access to foreign fishing. Each of these attributes was evaluated with a set of questions whose answers could be categorized in a hierarchical order from worst- to best-case scenarios. In cases where several questions applied to the same attribute, we summarized all responses into a single scale using multidimensional scaling. Multidimensional scaling is an ordination method that uses similarities and dissimilarities among variables to reduce them to a specific number of dimensions. Here, we used the anchored multidimensional scaling method developed by Pitcher and Preikshot [37]. In this method, hypothetical countries are generated with the worst- and best-case scenarios for each question and used as normative extremes of a scale on which real countries are ranked. The approach also incorporates uncertainty using a Monte Carlo simulation tool based on the maximum and minimum possible for each score [38]. A copy of the software is available on request.
We focused our assessment on fishery management conditions for all ocean realms under the sovereignty of a defined coastal territory. Under the United Nations Convention on the Law of the Sea [39], the protection and harvesting of coastal resources rest within the 200-nautical mile EEZ of each coastal state. There are, however, exceptions, such as the European Union, whose fisheries regulations are mandated by the Common Fisheries Policy but whose enforcement is the responsibility of the member states; member states also differ in their fishing capability and possibly in their compliance with regulations. Similarly, many countries have overseas territories, which may or may not have autonomous control of the regulation of their fisheries, and consequently, there may be variations in the effectiveness of their management regimes. For instance, Saint Pierre and Miquelon, French Guiana, French Polynesia, French Southern and Antarctic lands, New Caledonia, Saint Martin, Reunion, Guadeloupe, and Martinique all are under the sovereignty of France, which furthermore has direct control over its own Atlantic and Mediterranean coast; yet all of these zones have different management conditions. To consider these differences in fishery management regimes, zones managed under the same entity (e.g., the European Union) or zones in different parts of the world belonging to the same sovereignty (e.g., overseas territories of France, United Kingdom, and United States) were analyzed separately. We also included zones that may not be technically defined or recognized as EEZs under the United Nations (e.g., division among coastal states of the Baltic Sea and Black Sea). In total, 245 such zones exist in the world (see Figure 3), which excludes conflict zones (e.g., the Paracel Islands, Spratly Islands, and Southern Kuriles). Out of those 245 zones, we were unable to gather data for isolated islands under the sovereignty of the United Kingdom (i.e., Ascension, Pitcairn, Saint Helena, South Georgia, and the South Sandwich Islands and Tristan da Cunha) and France (Clipperton Atoll) for which neither contacts nor information was available. We also excluded Monaco and Singapore; interviewees at local authorities (Coopération Internationale pour l'Environnement et de Développement in Monaco and the Agri-Food and Veterinary Authority in Singapore) in both of these countries claimed that although marine fishing occurs, it was minimal and considered insufficient to motivate governmental regulation. The final database contained complete data for 236 zones. Although all data are reported in Figures 3 and 4, the statistics reported in the text were based on 209 inhabited zones for which per capita Gross Domestic Product data exist; that excluded uninhabited and isolated atolls to prevent biases due to the fact that we could not get data for all such areas (i.e., United Kingdom and France, see above).
For each of the attributes analyzed (i.e., scientific robustness, policymaking transparency, enforcement capability, fishing effort control, subsidies, and access to foreign fishing), we created a set of questions whose answers could be ranked on a scale from worst- to best-case scenarios. The resulting survey included 23 multiple choice questions and was posted on the Internet (http://as01.ucis.dal.ca/ramweb/surveys/fishery_assessment/) in five different languages (i.e., English, Spanish, French, Portuguese, and German). We searched for contacts (email addresses and phone numbers) of fishery experts for all coastal territories in the world. Our sources of information were reports on scientific and administrative meetings relevant to fisheries, Web pages of nongovernmental organizations, Web pages of fishery management organizations in each territory, and proceedings of international conferences on fisheries. The final directory included contact information for 13,892 people. We sent personalized emails using recommendations of email marketing companies to prevent filtering of emails by local servers and promote participation. The survey started in April 2007 and was completed in April 2008. For zones where we did not receive an email response, we carried out phone interviews with local experts, and both email and phone queries were done until at least one full set of responses was available for each zone. We received 1,188 positive responses including at least one from each country with ocean access. Multiple responses for the same zone were averaged.
Expert opinion surveys have been very popular in social, medical, political, and economic sciences [40], and some examples exist in fisheries studies (e.g., [41]). In fisheries research, expert opinions have been categorized as a “highly reliable” method given that overall, it works as a form of “peer review approach” and, for some crucial issues, is the only knowledge available (see [42]). The approach is also cost-efficient and relatively fast. The collection of empirical data for an analysis of this scale could prove ineffective because country-scale data are patchy, in most cases inaccessible through traditional searching engines, and because old data may not describe current conditions. For these reasons, we chose the survey of local experts to acquire data.
The quality of expert opinion surveys relies on the consistency of responders and their understanding of the issues. These problems are defined as reliability and validity [40], which in statistical terms are analogous to precision and accuracy. The former basically considers the extent to which responders agree in their responses and the latter the extent to which the responses approach the truth. Evaluation of data reliability and validity also allows assessment of the extent of expert biases, which may arise for different reasons (e.g., cultural differences, patriotism, opposition to governmental institutions, etc.). Our assessment of reliability and validity was as follows:
The metric we used to quantify fisheries sustainability has been recently published in two independent publications [30],[31], but not applied to the landings of any country. Here, we provide a brief description of its rationale and calculation, but extended details are provided by Libralato et al. [31] and Coll et al. [30].
Fisheries catches represent a net export of mass and energy that can no longer be used within an ecosystem; failure of the ecosystem to compensate for that energy loss implies overexploitation. This notion of overexploitation will require establishing a contrast between the loss of energy in the ecosystem due to a particular catch, the energy at the base of the food web in the area where the catch was taken, and reference points indicating whether the ratio between the energy that is taken (by fishing) and produced (through primary production) is sustainable or not. This concept has been recently incorporated into a metric that aims to quantify the probability that an ecosystem is being sustainably fished (Psust: after [31]). The metric first calculates the amount of Primary Productivity Required (PPR after [43]) to sustain a catch as , where s is the total number of caught species, Wi is catch weight of each species i, TE is transfer efficiency specific for the ecosystem, and TLi is the trophic level of species i. The metric assumes a conservative 9:1 ratio for the conversion of total weight to carbon [43]. The loss of energy in the ecosystem (i.e., Lindex, after [31]) is calculated by comparing PPR to the primary production at the base of the food web (PP) as , where TLc is the mean trophic level of the catch as calculated from the TL and weight of each species in the catch. PP is parameterized from chlorophyll pigment concentrations and photosynthetically active radiation [30]. The probability that such energy loss is sustainable (i.e., Psust) is calculated by comparing Lindex to reference Lindexes in which overfishing or sustainability have previously been identified. Reference Lindexes were quantified for different regions worldwide using a set of well-documented mass balance models representative of exploited ecosystems and constructed with independent information for each ecosystem [31]. Each of these models is classified as overfished if it meets one or more of the following criteria: (1) biomass of any species falls below minimum biologically acceptable limit, (2) diversity decreases, (3) year-to-year variation in populations or catches increases, (4) resilience to perturbations decreases, (5) economic and social benefits decrease, and (5) nontargeted species get impaired (see [30],[31] and references therein for justification of these criterion). Models were defined as sustainable when the impacts of exploitation did not result in any of the above symptoms. The frequency of sustainable or overfished Lindexes allowed us to calculate the probability of sustainability (Psust) for any particular Lindex value as , where N is the number of models in which Lindexes lead to sustainable or overfishing conditions. Probabilities of fisheries sustainability were calculated for each EEZ in the world using catch data as from the Sea Around Us fisheries database, which contains harmonized data from a variety of sources including the Food and Agriculture Organization (i.e., statistics on fisheries catches from 1950 to 2004; [44]). That database adjusted landings data to account for the fishing of long-distance fishing fleets (i.e., landings that are reported by one country, but fished in a different one). Landings data were also adjusted to include discards [45] and a global estimate of illegal, unreported, or unregulated catches [46],[47].
Data on fisheries sustainability was quantified for the year 2004 and linked to the effectiveness of fisheries management using a classification/regression tree. A classification tree tests for significant differences in fisheries sustainability among the quarters of each attribute (note that the first and fourth quarters are the extremes of a scale from worst- to best-case scenarios for each attribute; see Figure 2). The attribute that maximizes differences among quarters (i.e., smallest p-value) is placed at the root of the tree and the EEZs in each of those quarters separated in different branches. Subsequently, the EEZs in each branch are tested for significant differences among quarters of the remaining attributes. The attribute that maximizes differences among quarters is placed at the base of the branch and the EEZs in each of those quarters separated in upper branches. The process is repeated until no differences are found within each branch in any remaining attribute. This analysis included all attributes considered in this study: scientific robustness, policymaking transparency, implementation capability, fishing capacity, subsidies, access to foreign fishing, and country wealth (i.e., 2007 per capita Gross Domestic Product larger or smaller than US$10,000, respectively). Given the inflation of Type I errors due to multiple comparisons, significance was set at p<0.01.
|
10.1371/journal.ppat.1004967 | A Novel AT-Rich DNA Recognition Mechanism for Bacterial Xenogeneic Silencer MvaT | Bacterial xenogeneic silencing proteins selectively bind to and silence expression from many AT rich regions of the chromosome. They serve as master regulators of horizontally acquired DNA, including a large number of virulence genes. To date, three distinct families of xenogeneic silencers have been identified: H-NS of Proteobacteria, Lsr2 of the Actinomycetes, and MvaT of Pseudomonas sp. Although H-NS and Lsr2 family proteins are structurally different, they all recognize the AT-rich DNA minor groove through a common AT-hook-like motif, which is absent in the MvaT family. Thus, the DNA binding mechanism of MvaT has not been determined. Here, we report the characteristics of DNA sequences targeted by MvaT with protein binding microarrays, which indicates that MvaT prefers binding flexible DNA sequences with multiple TpA steps. We demonstrate that there are clear differences in sequence preferences between MvaT and the other two xenogeneic silencer families. We also determined the structure of the DNA-binding domain of MvaT in complex with a high affinity DNA dodecamer using solution NMR. This is the first experimental structure of a xenogeneic silencer in complex with DNA, which reveals that MvaT recognizes the AT-rich DNA both through base readout by an “AT-pincer” motif inserted into the minor groove and through shape readout by multiple lysine side chains interacting with the DNA sugar-phosphate backbone. Mutations of key MvaT residues for DNA binding confirm their importance with both in vitro and in vivo assays. This novel DNA binding mode enables MvaT to better tolerate GC-base pair interruptions in the binding site and less prefer A tract DNA when compared to H-NS and Lsr2. Comparison of MvaT with other bacterial xenogeneic silencers provides a clear picture that nature has evolved unique solutions for different bacterial genera to distinguish foreign from self DNA.
| During evolution, the bacteria frequently acquire new genes through horizontal transfer, in order to adapt new environments. However, foreign DNA sequences acquired are more likely to decrease rather than increase the fitness of the recipient bacteria. Therefore, many bacterial genera have evolved unique proteins to selectively repress the transcription of foreign genes. The opportunistic pathogen Pseudomonas aeruginosa is the principal cause of the morbidity and mortality in cystic fibrosis patients and is among the major causes of nosocomial infections. As a xenogeneic silencer, the MvaT protein of P. aeruginosa is a master regulator of horizontally acquired genes including many critical for drug resistance and virulence. Here, we have characterized the DNA sequences preferentially targeted by MvaT, and identified differences in sequence preferences between MvaT and other xenogeneic silencers. The high resolution structure of the DNA-binding domain of MvaT in complex with a high affinity DNA target reveals a novel AT-rich DNA minor groove recognition mechanism, which perfectly explains the characteristic of MvaT’s DNA sequence preferences. Comparison between MvaT and other bacterial xenogeneic silencers demonstrates how unique solutions have been employed by different bacterial genera in distinguishing foreign DNA from DNA of their own genome.
| Horizontal gene transfer, or lateral gene transfer, refers to the acquisition of foreign genes not from a direct ancestor by an organism. It is a major evolutionary force in bacteria and unicellular eukaryotes [1], and a recent study even suggests that horizontal gene transfer may also contribute to animal evolution [2]. Although the bacteria frequently acquire foreign genes in order to adapt the environment, the expression of foreign genes in a new host is more likely to reduce the bacterial fitness [3]. Bacterial xenogeneic silencing proteins selectively bind to and repress the transcription from regions of DNA that are significantly more AT-rich than the overall genome, which are likely to be acquired through horizontal gene transfer [4–6]. By dampening the expression of many AT-rich genes, these silencers improve the probability that newly-acquired sequences will be tolerated by the recipient organism. A recent study also shows that one of the major functions of xenogeneic silencers is to suppress the toxic intragenic transcription initiation of horizontally acquired AT-rich DNA [7]. Silencing of foreign DNA can potentiate bacterial evolution by allowing a pool of potentially useful genes to exist cryptically in the population. Such genes occasionally find use when an individual cell in the population evolves the necessary regulatory circuitry to control the gene’s expression under the appropriate environmental and temporal contexts [8–13]. As a result of their activity, xenogeneic silencers are the master regulators of horizontally acquired sequences, including many critical for drug resistance and virulence, in a large number of important bacterial pathogens including Mycobacteria, Vibrio, Salmonella, Escherichia, Yersinia, Bordetella, and Pseudomonas [14–28].
Xenogeneic silencing proteins can be divided into three distinct groups based on sequence similarity: the H-NS-like proteins that are found in many genera of alpha, beta, and gamma-Proteobacteria, the Lsr2 proteins found in the actinobacteria, and the MvaT-like proteins of the pseudomonads [5]. They all share the same domain arrangement, with an N-terminal oligomerization domain and a C-terminal DNA binding domain [29, 30]. These proteins can cooperatively bind to large sections of DNA, and recent studies suggest that the formation of nucleoprotein filaments is essential for their gene-silencing function [31–34]. The DNA-binding domains of Lsr2 from Mycobacterium tuberculosis and H-NS from Escherichia coli and Salmonella typhimurium have been solved and their interactions with DNA have been modeled using titration data from NMR studies [16, 29, 35–37]. Interestingly, Lsr2 and H-NS, despite being structurally distinct, employ a common mechanism to selectively target AT-rich DNA by inserting an AT-hook-like “Q/RGR” motif into the minor groove of DNA. Another H-NS family protein Ler, with the AT-hook-like motif replaced by a “VGR” sequence, only inserts the arginine side chain into the minor groove of DNA and can alleviate H-NS mediated silencing of virulence operons in enteropathogeic E. coli [38].
MvaT was originally identified as a global regulator of virulence gene expression in Pseudomonas aeruginosa and subsequently shown to be a repressor of the cupA fimbrial cluster necessary for biofilm formation [23, 39, 40]. MvaT was characterized as an H-NS functional analog due to its ability to complement various phenotypes of the E. coli Δhns mutant [41]. Transcriptome and chromatin immunoprecipitation coupled with DNA microarrays (ChIP-on-chip) analysis have shown that MvaT, and its paralog MvaU, both preferentially bind AT-rich regions of the P. aeruginosa genome to regulate ~350 genes and that there is nearly complete overlap in the sets of genes under the control of each protein, although why apparently redundant paralogs exist is unclear [22, 42]. In addition to the cupA operon, a lot of virulence genes from P. aeruginosa are repressed by MvaT/MvaU, such as the genes required for synthesis of the redox-active pigment and toxin pyocyanin, and the genes related to the type III and type VI secretion systems [22]. Loss of both MvaT and MvaU is lethal to P. aeruginosa strain PAO1, a phenomenon that has been traced to the fact that these proteins inhibit activation of the Pf4 prophage [43]. Structural predictions suggest that MvaT/MvaU likely evolved from the H-NS family, however both proteins share very low sequence identity with H-NS and lack the canonical H-NS motif that contains the AT-hook-like structure [5, 41]. This leaves open the question of how the MvaT-like proteins target AT-rich DNA.
In this work we performed structural and functional analyses of the DNA binding properties of MvaT. High-throughput DNA binding assays indicate that MvaT prefers highly flexible sequences that contain multiple TpA steps and has considerable tolerance to GC-base pair interruptions in the binding sites. We solved the solution structure of the DNA-binding domain of MvaT in its free form and in complex with a high affinity DNA dodecamer and find that, like H-NS and Lsr2, MvaT recognizes structural features in the minor groove unique to AT-rich DNA. With an “AT-pincer” motif inserted into the minor groove and several lysine residues interacting with DNA sugar-phosphate backbone, the double helix was distorted by significantly opening of the minor groove, reminiscent of several other proteins that target flexible TpA-rich sequences. The roles of residues predicted to be critical for binding were assessed both in vitro and in vivo which reveal that, while single amino acid substitutions did not completely abolish DNA binding activity due to extensive contacts between MvaT and DNA, those substitutions with significantly reduced DNA binding activity render the protein non-functional in vivo. Finally, our structural data indicates that the MvaT-like proteins likely evolved from an H-NS paralog but acquired a mode of binding that is significantly distinct from other xenogeneic silencers.
Previous ChIP-on-chip analysis indicated that MvaT preferentially binds to AT-rich regions of the chromosome [22]. Electrophoretic mobility shift assays (EMSA) were performed with purified full-length MvaT against a radiolabeled 340 bp fragment of the cupA1 promoter (%GC = 54), previously shown to bind MvaT in vivo and in vitro. For a negative control we employed a 204 bp fragment of the P. aeruginosa PAO1 gene PA3900 (%GC = 74), which lies within a ~180 kb GC-rich region of the genome that does not contain any MvaT binding sites in vivo. MvaT was able to shift both labeled fragments of DNA when incubated with either fragment alone, with apparent dissociation constants of 57 ± 22 nM and 154 ± 53 nM towards the cupA1 promoter and PA3900 fragments, respectively. The binding affinity of MvaT for the GC-rich PA3900 fragment is very close to that (~148 nM) determined previously with EMSA [32].
Although the affinity of MvaT towards the cupA1 promoter DNA is only ~3 times higher than that of the PA3900 fragment, significant differences were observed when competition assays were employed (Fig 1). MvaT, preincubated with the radiolabeled GC-rich PA3900 fragment, was displaced by unlabeled cupA1 promoter DNA at less than a 5-fold molar excess. In contrast, a 200-fold excess of unlabeled PA3900 DNA was unable to displace MvaT from a pre-formed complex with the radiolabeled cupA1 fragment. These data indicate that MvaT oligomers likely interact with GC-rich DNA through a series of coupled low affinity interactions. Subunits within MvaT oligomers dissociate from the GC-rich target frequently, enabling the protein oligomer to be removed from the DNA rapidly. On the cupA1 promoter fragment, which is comparatively AT-rich, the MvaT protein forms an oligomeric complex sufficiently stable that a 25-fold excess of unlabeled cupA1 DNA is required to displace it. This indicates that once bound to an appropriate target fragment, the MvaT oligomers make highly stable complexes that dissociate with very slow off rates. These results are consistent with the finding that MvaT can cooperatively bind to large sections of DNA and form nucleoprotein filaments [32]. In such nucleoprotein filaments, the dissociation of MvaT primarily occurs at the two ends where the energy cost is lower than inside the filament, resulting in slow off rates for the binding.
The DNA binding specificity of MvaT was further interrogated using protein binding microarrays (PBM). GST-tagged MvaT was applied to microarrays containing 41,944 double-stranded 60-mer oligonucleotide target sequences each comprising a constant 25-mer primer sequence followed by a variable 35-mer sequence. Two different arrays (designated ME and HK) were designed such that all possible non-palindromic 8-mers are represented 32 times (16 times for palindromic 8-mers) on each array [44, 45]. The combined data (average) from the two independent array experiments were used to analyze the relative binding preferences, providing an unbiased estimate of relative preference to each 8-mer that is robust to variations in position, location and flanking sequence [29, 44].
The relative binding preference of MvaT for each 8-mer sequence was calculated using a rank-based, non-parametric statistical measure (E-score) that is largely invariant to protein concentrations. This facilitates a comparison of different experiments on the same scale, which is useful when assessing differences in binding targets between different proteins like MvaT and H-NS [44, 46]. The E-score ranges from 0.5 (highest) to -0.5 (lowest). Random permutations of the array data indicate there should be no random 8-mer sequence that achieves an E-score above 0.45 [47]. The PBM experiments identified multiple 8-mers with E-scores above 0.45 for MvaT and the highest score was around 0.49, indicating that some sequences were clearly preferential targets for binding by MvaT than others (S1 Dataset).
As shown in Fig 2, the target preferences of MvaT were compared to those previously determined in similar experiments for both H-NS and Lsr2 [29]. Consistent with earlier studies, the sequences most tightly bound by all three proteins were overwhelmingly AT-rich, as the PBM E-score of MvaT is also positively correlated with the AT-content of 8-mers (Fig 2A). This provides the basis for their functional similarity and that MvaT can complement various phenotypes of the E. coli Δhns mutant [41].
However, a more detailed comparison of binding preferences between MvaT, H-NS, and Lsr2 indicates that subtle but important differences exist in their modes of binding. Both H-NS and MvaT are biased toward sequences containing multiple TpA dinucleotide steps (also called TpA steps), while TpA steps do not affect Lsr2 binding (Fig 2B). Indeed sequences with the highest affinities for both H-NS and MvaT were composed of multiple adjacent TpA steps and one of the highest scoring 8-mers for each was TATATATA. The absence of TpA steps from a given sequence, however, imparts a higher penalty on binding of MvaT than it does on either H-NS or Lsr2.
More than any other dinucleotide step including those containing a G or C, TpA steps confer the greatest amount of flexibility on DNA [48]. In stark contrast, A-tracts (sequences composed of several consecutive A or T bases without a TpA-step), are the most rigid and inflexible of all DNA sequences [49]. Base stacking within A-tract sequences also compresses the minor groove more than any other type of sequence. While the presence of A-tracts in a sequence generally improved binding by both H-NS and Lsr2, A-tracts imparted a significant penalty on binding by MvaT (Fig 2C).
The effect that G or C bases within the target site have on binding by the various proteins was also assessed (Fig 2D). Notably all proteins had lower affinity for AT-rich sequences when a single G or C nucleotide was placed in the center but the effect on binding by MvaT was the lowest. Binding of the proteins to 8-mer sequences containing six A or T bases and two G or C bases at various positions was assessed. 8-mer sequences with two G or C bases can have as many as 6 adjacent A or T nucleotides (e.g. AATATAGG), to as few as two (e.g. AAGTTGAA). These data indicate that the penalty for both H-NS and Lsr2 increases as the number of contiguous A or T nucleotides decreases. MvaT, on the other hand, seemed to be more tolerant to interruptions by G or C nucleotides within an AT-rich binding site and many high-scoring 8-mer sequences contained only sets of two adjacent T or A nucleotides. These differences in sequence preferences imply that the DNA recognition mechanism of MvaT may be different from that of H-NS and Lsr2.
Significant advances have recently been made in the computational prediction of specific DNA structure parameters. Structures of MvaT-bound sequences predicted by DNAshape [50] reveal that the top 20 scoring 100% AT 8-mer sequences are markedly different in structure from the 20 100% AT 8-mers with the lowest score (S1 Fig). Sequences with high MvaT binding E-scores all contain significant stretches where the minor groove exceeds 5.5 Å in width while the sequences that scored the lowest were predicted to contain much narrower minor grooves, typically below 4 Å. Top scoring sequences also had higher values for roll and propeller twist, and lower values for helical twist than the lower scoring sequences.
We determined the solution structure of the C-terminal DNA-binding domain of MvaT (MvaTctd, residues 77–124) from P. aeruginosa strain PAO1 using nuclear magnetic resonance (NMR), with nearly complete resonance assignments achieved (S2 Fig). Restraints and structural statistics are summarized in Table 1. MvaTctd consists of a three-stranded antiparallel β–sheet (β1, residues 83–86; β2, residues 93–96; β3, residues 120–122), and two α-helices (α1, residues 102–111; α2, residues 113–119). MvaTctd adopts the topology of β1-β2-α1-α2-β3, and a loop (loop2, residues 97–101) links the β1-β2 hairpin and the helices α1-α2. The N- and C-terminal regions are highly flexible (Fig 3A). The two consecutive helices are packed on one side of the 3-stranded β–sheet and form a hydrophobic core composed of residues Tyr85, Ile94, Thr96, Thr103, Leu104, Trp107, Trp111, Val116, Trp119 and Ala120 (Fig 3B). The overall fold and secondary structure composition of MvaTctd are similar to those of the C-terminal domain of H-NS (H-NSctd), except that H-NSctd does not have a third strand β3 and it is a 310 helix instead of the helix α2 in MvaTctd [29].
The interaction surfaces of MvaTctd and DNA were mapped by NMR titration experiments. A DNA duplex d(CGCATATATGCG)2, which we refer to as “3AT”, was chosen as it contains a sequence among those with highest scores from our PBM study. Comparison of 2D 1H-15N HSQC spectra of MvaTctd in free and DNA-bound form reveals that the residues with significant combined NH chemical shift differences (Δδcomb > 0.25 ppm) are Arg80, Lys81, Tyr85, and Lys97-Lys105 (Fig 3C and 3D), mainly located on the N-terminal region, loop2 and the beginning of helix α1. These residues, except Tyr85, are clustered on the structure of MvaTctd, constituting the DNA binding site. Tyr85 is probably affected indirectly through its aromatic side chain, which protrudes straight towards loop2. The stoichiometry of binding is one molecule of MvaTctd to one molecule of 3AT duplex as assessed by fitting of the chemical shift changes with various DNA concentrations, which was also confirmed by the results of isothermal titration calorimetry (ITC, see below). By comparing 2D 1H NOESY spectra of the DNA, free or in complex with protein, we also mapped the regions of 3AT that interact with MvaTctd. Significant intra-residual H1’-H6/H8 NOE peak shifts (H1’ or H6/H8 Δδ > 0.025 ppm) occur at the central ATATATG residues (Fig 3E and 3F).
To confirm that MvaTctd binds the minor groove we performed a competition experiment using netropsin, a natural oligopeptide that binds the minor groove of AT-rich DNA. With the addition of increasing concentration of netropsin into a sample of 15N-MvaTctd containing a two-fold excess of 3AT, the NMR signals of MvaTctd shifted from the DNA-bound form back to the free form gradually, indicating that the MvaTctd/DNA complex was disrupted by netropsin in a concentration-dependent manner. At a netropsin/DNA ratio of 2.5:1, the 2D 1H-15N HSQC spectrum is nearly identical to that of free MvaTctd, indicating that netropsin almost completely releases MvaTctd from 3AT.
Unlike Lsr2 and H-NS [16, 29], the DNA-bound MvaTctd has a single set of NH signals, and none of the NH signals disappear upon DNA binding (S2 Fig), enabling us to determine the solution structure of MvaTctd in complex with the 3AT dodecamer (Fig 4A). The MvaTctd/3AT complex structure was determined using 3,541 experimental distance restraints, including 57 intermolecular distance restraints, and the backbone heavy atom RMSD of the 20 final structures with lowest AMBER energies is 0.49 Å (Table 1).
The structure of DNA-bound MvaTctd is almost the same as that of the free MvaTctd, as their backbone heavy atom RMSD between the mean structures is only 0.90 Å. The major difference is that the N-terminal region becomes more rigid, consistent with the gradual emergence of the ε-NH signal of the Arg80 side chain when bound to 3AT (S2B Fig). The interaction surface of MvaTctd is positively charged (Fig 4B) and buries a surface area of 1,707±14 Å2. Loop2 of MvaTctd is partially inserted into the DNA minor groove with the backbone of residues Gly99 and Asn100 making contacts to the AT base pairs at the bottom, while residues Lys97 and Gly98 are tilted toward the top of the minor groove. The backbone amides of Gly99 and Asn100 form hydrogen bonds with the O2 atom of T19 and the N3 atom of A18, respectively. The side chain of Asn100 extends along the minor groove, and its δ-NH2 group forms hydrogen bonds with the O2 atoms of T9 and T17 (Fig 4C and 4D). In addition, the side chain of Arg80 is also inserted into the DNA minor groove with its guanidino group hydrogen bonded to O2 atoms of T5 and T21 (Fig 4C and 4D) and thus stabilizes the N-terminal region of MvaTctd. The side chains of Arg80 and Asn100 are pointed away from each other, and occupy a region covering all six AT base pairs together with loop2. This DNA binding motif was thus given the name “AT-pincer” as the minor groove-intercalating residues are from two different loops of MvaTctd. Besides interacting with the DNA minor groove, MvaTctd also contains six lysine residues (Lys81, Lys83, Lys97, Lys102, Lys105 and Lys108) that are well-positioned to make hydrophobic or electrostatic contacts with the DNA sugar-phosphate backbone through side chain methylene groups or the ε-amino groups, respectively (Fig 4C and 4D). The Lys81 side chain is stabilized upon binding 3AT, while the side chains of Lys83, Lys97, Lys102, Lys105 and Lys108 are already well-defined in the structure of free MvaTctd, as these lysine side chains are stabilized by hydrophobic interactions with nearby residues. This extensive network of lysine residues significantly increases the DNA contact surface and is a distinguishing feature of MvaTctd. This AT-rich DNA binding mode is clearly different from that of H-NS and Lsr2, which share a common DNA binding mechanism through the AT-hook-like motif (detailed comparison in the Discussion section).
Upon binding of MvaTctd, the minor groove of 3AT is expanded and the base-stacking geometry is significantly rearranged. MvaTctd-bound 3AT shows increased roll and inclination angles compared with a free DNA decamer d(GGATATATCC)2 with the same central ATATAT sequence (PDB 2LWG [51]), ~9.6° and ~15.5° per step in average respectively (Fig 5A), indicating that the base pairs locally bend toward the major groove and thus open the minor groove [49]. The minor groove width of MvaTctd-bound 3AT reaches a minimum at the middle A6pT7 base step and progressively widens toward each end of the ATATAT sequence. The side chains of Arg80 and Asn100 are located at the A4pT5 and A8pT9 base steps respectively, where the minor groove widths are significantly widened compared with those in the structure of 2LWG. The above mentioned lysine side chains are fit well to the backbone of 3AT in the MvaTctd/DNA complex, while their conformations remain similar to those in free protein. It is thus likely that the distortion of the DNA duplex is to accommodate the conformation of this “lysine network”. In contrast, A-tract DNA was reported to possess a very narrow minor groove, and the roll and inclination angles are consecutively small or negative [49, 52]. A-tract DNA also exhibits an overall bend of ~15° towards the minor groove [53], whereas the free and MvaTctd-bound 3AT are nearly straight (Fig 5B). These may explain why A-tract DNA is not preferred by MvaT, as the rigid A-tract DNA does not have a favorable conformation for MvaT binding.
To study the importance of the residues implicated by the MvaTctd/DNA complex structure, we generated a series of MvaTctd mutants containing substitutions R80A, K81A, K83A, K97A, N100A, K102A, K105A and K108A. All these mutations except K83A do not have a significant effect on the overall structure of MvaTctd as the 2D 1H-15N HSQC spectra of these mutants are similar to that of wild-type (WT) MvaTctd (S3 Fig). Since the NH signals of K83A are completely different from that of WT MvaTctd, the mutation K83A likely resulted in an overall fold change. This indicates Lys83 side chain plays an important role in stabilizing the structure, as it forms a salt bridge with Glu117 and its aliphatic groups make hydrophobic contact with Tyr85, Ala120, and Leu122.
Each mutant, excluding K83A, was subsequently titrated with the 3AT dodecamer and assessed by NMR (S4 Fig). For mutants K81A and K97A, the NH signals shift patterns during DNA titration are quite similar to those of WT MvaTctd in terms of scale and direction, suggesting that the two mutations should not have a significant effect on the DNA binding of MvaTctd. Mutants R80A, K102A show reduced chemical shift changes, but with similar NH signals shift directions to those of WT MvaTctd. Mutations N100A, K105A and K108A have the most profound effects on NH signal perturbation patterns induced by 3AT, with much smaller NH chemical shift changes and some NH signals shift directions are quite different from those of WT MvaTctd, indicating that these mutations may significantly perturb the DNA binding mode of MvaTctd.
ITC experiments were performed to characterize the binding affinities between DNA and MvaTctd or its mutants (Figs 6A and S5). As expected, WT MvaTctd exhibits the highest affinity for 3AT, with a Kd of about 10.7 μM. The binding affinities for mutants R80A and K108A are over 10 times weaker than WT MvaTctd, while Kd values of mutants N100A and K105A are increased by over 3 times. On the contrary, mutations K81A, K97A and K102A only result in slightly lower DNA binding affinity (Kd increased by ~70%).
MvaTctd binds to a GC-rich DNA, d(CGCGCGCG)2 duplex, with much weaker affinity (Kd ~360 μM) (S5I Fig). It is noteworthy that WT MvaTctd and its mutants bind to AT-rich DNA in an endothermic manner, indicating that the binding is an entropy-driven process. It has been suggested that the large positive enthalpy change is due to the desolvation of the DNA minor groove and distortion of DNA duplex as we observed in the MvaTctd/DNA complex [54]. The favorable entropy change can be attributed to the displacement of highly ordered water molecules in the DNA minor groove [55]. In contrast, the binding of MvaTctd to GC-rich DNA is driven by both enthalpy and entropy, suggesting that MvaT binds GC-rich DNA in quite different mode.
We next tested whether those residues of MvaT that are implicated in DNA binding are important for the function of MvaT in cells of P. aeruginosa. It has been shown previously that MvaT in P. aeruginosa represses phase-variable expression of the cupA fimbrial genes [23]. In the strain PAO1 ΔmvaT cupA1 lacZ, where the lacZ gene is positioned downstream of the chromosomal cupA1 gene, the cupA genes are expressed in a phase-variable manner, manifested by the appearance of blue and white colonies on LB agar plates that contain the chromogenic substrate 5-bromo-4-chloro-3-indolyl-β-D-galactopyranoside (X-Gal). This reversible ON–OFF switching of cupA gene expression observed in cells of the ΔmvaT mutant strain can be repressed by providing the mvaT gene in trans, giving rise to white colonies on LB agar plates containing X-Gal [23].
To test whether the MvaT mutants containing substitutions R80A, K81A, K83A, K97A, N100A, K102A, K105A and K108A could repress phase-variable expression of the cupA fimbrial genes as efficiently as WT MvaT, we introduced into cells of the reporter strain PAO1 ΔmvaT cupA lacZ plasmids directing the synthesis of either WT MvaT or an MvaT mutant, each containing a vesicular stomatitis virus-glycoprotein (VSV-G) epitope tag at its C-terminus. Cells were then grown on LB agar plates containing X-Gal. Western blotting using an antibody against the VSV-G epitope tag revealed that all proteins were made at comparable amounts except for K105A, which was slightly less abundant than the others (Fig 6B).
Unlike WT MvaT containing a VSV-G epitope tag (MvaT-V), the MvaT-V mutants R80A, K83A, N100A, K105A and K108A failed to repress phase-variable expression of the cupA lacZ reporter (Fig 6B). Conversely, MvaT-V mutants K81A, K97A, and K102A could repress cupA activity, similarly to WT MvaT-V. To further elucidate the effect of the MvaT-V variants on cupA repression, lacZ expression was quantified in cells grown in liquid culture by β-galactosidase assay. This confirmed that the MvaT-V variants K81A, K97A, and K102A could repress cupA activity just as well as WT MvaT-V (Fig 6C). Of the MvaT-V mutants that were unable to repress cupA, the R80A variant was most similar to the empty vector control and thus the most defective, whereas the K105A mutant was least defective. Because substitution K105A appears to reduce the abundance of MvaT-V, it is possible that the reduced ability of the K105A mutant to repress phase-variable expression of the cupA genes can in part be explained by the effects of this substitution on protein abundance. The impaired function for the K83A variant likely resulted from the change of protein fold as discussed above. Both mutants N100A and K108A were significantly impaired in their ability to repress cupA expression to similar extents (Fig 6C). These findings demonstrate that mutants of MvaT that exhibit notably reduced DNA binding affinity in vitro are therefore functionally impaired in vivo.
In this study, we have performed a comprehensive analysis of the DNA binding properties and mechanism of the xenogeneic silencer MvaT from the opportunistic pathogen P. aeruginosa. We have demonstrated that MvaT has higher binding affinity towards the AT-rich genomic target DNA, compared with the GC-rich non-target DNA. In addition, MvaT binds its target DNA with much slower off rate, which should also contribute to its selectivity for target DNA. The PBM data reveal that MvaT in general has similar DNA sequence preferences as H-NS and Lsr2. However, it is also found that MvaT prefers AT-rich sequences with multiple TpA steps and has considerable tolerance to GC-base pair interruptions, which is unique to MvaT. Our structural and functional evidence indicates that MvaT employs a novel AT-rich DNA recognition mechanism, distinct from that of H-NS and Lsr2, to carry out the xenogeneic silencing function.
Although the DNA binding affinity of a single DNA binding domain is relatively low (~10 μM), the full-length MvaT binds its genomic AT-rich target DNA with much higher affinity due to the multivalent effect, since the full-length MvaT is oligomerized in solution through its N-terminal domains [30], and thus it binds the target DNA containing multiple binding sites as a oligomer with multiple DNA binding domains. As a result, the binding energy of each full-length MvaT molecule in a nucleoprotein complex on DNA is not only contributed by its interaction with DNA, but also by its interaction with neighboring molecules of DNA-bound MvaT (mediated through their N-terminal domains), which significantly increase the overall stability of the complex. The multivalent interaction is widespread in chromatin biology [56], and it is also employed by other xenogeneic silencers. The Kd values of H-NSctd (~9 μM) and Lsr2ctd (~4 μM) for binding 3AT DNA are similar to that of MvaTctd [29], while their full-length proteins also bind DNA with much higher affinities [33, 57]. Multivalent binding consisted of individual weak interaction is more susceptible to competition than a monovalent tight binding could be [58], which make it easier for counter silencers, such as Ler, to alleviate the repression of xenogeneic silencers in the regulation of gene expression [59].
The structure of MvaTctd in complex with 3AT DNA reveals that the DNA binding domain of MvaT recognizes AT-rich DNA through both the “AT-pincer” motif inserted into the DNA minor groove and the “lysine network” with multiple positive charges interacting with DNA backbone phosphates. MvaT lacks the AT-hook-like “Q/RGR” motif that is critical for both H-NS and Lsr2 to bind DNA, although they all specifically target the minor groove of AT-rich DNA. Homology predictions indicate that MvaT could share structural similarity to the H-NS family proteins found in the Xanthomonadaceae including the plant pathogen Xylella fastidiosa. An alignment of the MvaT DNA-binding domain with a diverse set of H-NS family members indicates that MvaT shares similarity to some H-NS homologs in regions immediately N-terminal to the canonical H-NS motif (TW(S/T)G(Q/R)GRTP). The canonical motif, however, is replaced by a different sequence (VIETKGGNH) that is highly conserved among the MvaT-like proteins and absent in members of the H-NS family (Fig 7A).
The overall fold of MvaTctd is similar to that of H-NSctd, although the relative orientations of the helices and the β sheet for these two proteins are different (Fig 7B). Loop2 of MvaTctd, comprising the sequence (KGGNH), is two residues shorter than the corresponding loop of H-NSctd, where the AT-hook-like “Q/RGR” motif is embedded (Fig 7A and 7B). The first missing residue in MvaT corresponds to the minor groove binding “Q/R” of the “Q/RGR” motif. The AT-hook-like motifs of H-NS and Lsr2 are inserted into the DNA minor groove adopting a flat conformation with the side chains of the first “Q/R” and last “R” residues extended in opposite direction parallel to the minor groove floor. However, upon MvaT binding DNA, only Gly99 and Asn100 of loop2 insert into the minor groove and form hydrogen bonds with the bases. The side chain of Asn100 has a similar conformation as the last arginine of the “Q/RGR” motif, and thus the MvaT Gly99 and Asn100 residues appear to function as half of the AT-hook structure. Interestingly, the role of the missing first “Q/R” residue of the AT-hook-like motif is compensated by the side chain of Arg80 from the N-terminal region of MvaTctd, which thus enables MvaT to recognize the AT-rich DNA minor groove with an “AT-pincer” motif composed of residues Arg80 and Gly99-Asn100. As a result, there is a cavity in the protein/DNA interface with a minimum radius of 1.2 Å above the A6-T19 base pair (Fig 7B and 7C) as calculated using CAVERS [60], which could theoretically accommodate an exocyclic NH2 group from a GC-base pair. This should explain the higher tolerance of MvaT for GC-base pair interruptions as revealed by our PBM data, since the DNA minor groove is almost fully occupied by the AT-hook-like motifs of H-NS and Lsr2.
It was recently proposed that both direct (base readout) and indirect (shape readout) DNA readouts are important in DNA recognitions for most DNA binding proteins [61, 62]. While the “AT-pincer” motif of MvaT inserts into the minor groove and achieves the base readout, the side chains of several lysine residues (residues 81, 83, 97, 102, 105 and 108) of MvaT interact with the DNA sugar-phosphate backbone and constitute the minor groove shape readout. Mutagenesis studies indicate that both the “AT-pincer” motif and the “lysine network” are important for the DNA binding affinity of MvaTctd. Our functional assay also reveals that point mutation of residue Arg80 or Asn100 of the “AT pincer” motif, as well as residue Lys105 or Lys108 of the “lysine network”, impairs the function of MvaT to repress phase-variable expression of the cupA lacZ reporter. Structural comparison of free and DNA-bound MvaTctd shows that most of the key lysine side chains do not change their conformations upon binding DNA, while 3AT adopts a nearly straight conformation with its minor groove widened to accommodate MvaT binding. This can explain why MvaT favors flexible DNA sequences with multiple TpA steps over A-tract DNA. It is well known that A-tract DNA is more rigid and intrinsically bent ~15° towards the minor groove [49, 53], and therefore it would be energetically costly for the A-tract DNA to open up its minor groove and adopt a favorable conformation for MvaT binding.
The differences in DNA distortion upon binding by MvaT and H-NS/Lsr2 are notable. Minor groove distortion is most dramatically illustrated by the eukaryotic TATA binding protein (TBP), which opens the minor groove to unwind the double helix by approximately 120°, and compresses the major groove to bend the DNA by almost 80°. The critical parameter for TBP binding is not the sequence per se, but rather the inherent flexibility of the TpA step, which allows the protein to dramatically open the minor groove and make extensive contacts with both the phosphate backbone and bases that make up the floor of the groove. The distortions triggered in DNA when bound by MvaT are significant, although relatively small in comparison to TBP. At the other extreme are the eukaryotic HMG-I(Y) proteins that contain narrow and flexible AT-hook motifs, each composed of an R-G-R-P motif. The AT-hook forms a narrow crescent shaped structure that intercalates into the minor groove without significantly distorting the helix trajectory [63], which is probably the case of H-NS and Lsr2 when binding DNA with their AT-hook-like motifs. The MvaT DNA binding mode is least similar to Lsr2, which prefers A-tract DNA with a narrow minor groove and is largely insensitive to TpA steps.
Among the three families of bacterial xenogeneic silencers, H-NS and Lsr2 share a common AT-hook-like DNA recognition mechanism even though they are structurally dissimilar, while the sequence and structural similarities between MvaT and H-NS suggest they may share a common evolutionary origin. It is curious that MvaT adopts a distinct binding mode to recognize similar, but not identical, DNA sequence targets. Genomic analysis suggests that xenogeneic sequences frequently display higher AT-content compared to the host genome [64]. However, the mean genome-wide AT-contents for different bacterial genera could be quite different, such as ~48% for E. coli and S. typhimurium, and ~34% for M. tuberculosis and P. aeruginosa. We have previously reported that the fraction of bound sequence from ChIP-on-chip data begins to increase when the AT-content reaches ∼50% for H-NS from S. typhimurium and ∼38% for Lsr2 from M. tuberculosis, which correspond to the mean AT-content of the corresponding genome, respectively [29]. More interestingly, H-NS from S. typhimurium and Lsr2 from M. tuberculosis exhibit nearly identical binding patterns when the AT-content of the bound sequence is normalized against the mean AT-content of the respective genome. We have suggested that the “RGR” AT-hook-like motif employed primarily in xenogeneic silencers from M. tuberculosis (Lsr2) and B. vietnamiensis (Bv3F), both with low AT-content genomes, enables tighter binding to sequences of relatively lower AT-content [29]. In comparison, the “QGR” AT-hook-like motif in H-NS from S. typhimurium and E. coli, both with high AT-content genomes, has lower affinity towards mildly AT-rich DNA. Unexpectedly, the DNA binding affinity of the DNA binding domain of MvaT from P. aeruginosa is lower than that of Lsr2 from M. tuberculosis, both with similar low AT-content genomes, while it is similar to that of H-NS from S. typhimurium with high AT-content genome. Our biochemical and structural studies revealed that MvaT prefer flexible DNA sequences with multiple TpA steps and can better tolerate GC insertion in AT-rich sequences. It may follow that Pseudomonas has evolved MvaT with considerable GC-tolerance to cope with its low AT-content genome. It may also be possible that specific features like TpA steps are underrepresented in the Pseudomonas “core” genome compared to the mobile genome. It is still not clear why MvaT from Pseudomonas employ a different solution to distinguish foreign from self DNA, and the determinants may lie in detailed sequence characteristics of its genome.
While previous studies have been mainly focused on revealing functional similarities of xenogeneic silencers from different bacteria, their differences are largely overlooked. It is apparent that the abilities of xenogeneic silencers from different bacteria to distinguish foreign from self DNAs should be dependent on the mean AT-content of their corresponding genomes, and it is even possible that these xenogeneic silencers have to fine-tune their DNA binding properties to cope with different sequence characteristics of their own genomes and foreign DNAs. More studies should be inspired to further explore the correlation between the molecular mechanisms of xenogeneic silencers from different bacteria and the characteristics of their genomes.
P. aeruginosa strain PAO1 ΔmvaT cupA lacZ has been described previously [23]. E. coli DH5αF’IQ (Invitrogen) was used as the recipient strain for all plasmid constructions. E. coli strain BL21 (DE3) was used for protein purification.
When growing P. aeruginosa, gentamicin was used at 25 μg/ml for liquid cultures and 30 μg/ml for solid media. Phase-ON and phase-OFF colonies of the reporter strain were visualized following growth on LB agar containing 75 μg/ml X-Gal.
DNA sequence encoding the C-terminal domain of MvaT (residues 77–124) from P. aeruginosa was subcloned into the pET21b vector, directly upstream of the His6-tag coding sequence. Point mutations were generated using the site-directed mutagenesis kit (SBS Genetech). E. coli BL21 (DE3) strain harboring the plasmid was cultured in LB medium at 35°C, and protein was over-expressed by induction with 100 μM IPTG until OD600 > 1.0. For 15N and 13C isotopically labeling, the bacteria were first grown in LB medium till OD600 > 0.9, then collected and resuspended in 15N, 13C-labeled M9 minimal medium for continuing growth, and 100 μM IPTG was added to induce protein expression after 40 min. The Cells were harvested by centrifugation 9 h after induction and resuspended in lysis buffer (50 mM Tris-HCl, 1 M NaCl, 20 mM imidazole, pH 9.0), then lysed by freezing and thawing, followed by sonification. After centrifugation, the supernatant containing the His-tagged fusion protein was applied to Ni-NTA affinity column (Qiagen) and eluted by elution buffer (50 mM Tris-HCl, 1 M NaCl, 250 mM imidazole, pH 9.0). Protein was further purified with size exclusion chromatography on a superdex 75 column (Amersham) in 50 mM sodium phosphate with 50 mM NaCl (pH 6.0). DNA samples were prepared by hybridization of self-complementary oligonucleotides, first heated to 94°C for 5 min and then annealed by slowly cooling down to room temperature. Protein/DNA complex samples were prepared by mixing the protein and DNA duplex at a 1:1 ratio.
The NMR sample of free MvaTctd contained ~1 mM uniformly 15N, 13C—labeled protein in 50 mM sodium phosphate, 50 mM NaCl (pH 6.0) with 90% H2O/10% D2O, along with 0.01% NaN3 and 0.01% DSS. The NMR sample of MvaTctd/DNA complex contained ~1 mM 1:1 complex of uniformly 15N, 13C—labeled protein and unlabeled DNA in the same buffer. All NMR spectra were collected at 298 K on 600, 700 or 800 MHz Bruker Avance spectrometers equipped with triple-resonance cryoprobes. Proton chemical shifts were referenced directly to DSS. 15N and 13C chemical shifts were referenced indirectly to DSS. Backbone and aliphatic side chain resonances were assigned by 2D 1H-15N HSQC, 2D 1H-13C HSQC, 3D HNCACB, 3D CBCA(CO)NH, 3D HNCO, 3D HBHA(CBCA)(CO)NH, 3D (H)CCH-COSY, 3D (H)CCH-TOCSY and 3D H(C)CH-TOCSY. Aromatic resonances were assigned using 3D 1H-13C-edited-NOESY optimized for aromatic resonances. Resonances of DNA complexed with MvaTctd were assigned using 2D F1, F2-15N/13C-filtered NOESY spectrum. There is a single set of resonances for the two strands of the palindromic 3AT duplex for the MvaTctd/DNA complex sample, indicating that the binding process is fast on NMR time scale. Free DNA resonances were assigned with 2D 1H TOCSY and 2D 1H NOESY spectra. Data were processed using NMRPipe [65] and analyzed by NMR View [66].
The NMR samples used for DNA titration experiment all contained 0.1 mM uniformly 15N labeled protein in 50 mM sodium phosphate, 50 mM NaCl (pH 6.0) with 90% H2O/10% D2O. A series of 2D 1H-15N HSQC spectra with gradually increased DNA concentration (0.02 mM, 0.04 mM, 0.08 mM, 0.12 mM, 0.16 mM and 0.2 mM) were collected at 298 K on a Bruker Avance 500 MHz spectrometer with a triple-resonance cryoprobe and the chemical shifts changes were analyzed.
2D 1H NOESY experiments were performed to investigate the chemical shift perturbations for the DNA duplex d(CGCATATATGCG)2 samples with or without MvaTctd on a Bruker Avance 800 MHz spectrometer equipped with cryoprobe at 298 K. The NMR sample contained 0.4 mM DNA in 50 mM sodium phosphate, 50 mM NaCl (pH 6.0) with 100% D2O, and lyophilized protein powder of MvaTctd was added to final concentrations of 0.25 mM and 0.5 mM. The fingerprint region of intraresidual H1’-H6/H8 NOE cross peaks was analyzed.
For netropsin competition experiment, 0.1 mM uniformly 15N labeled MvaTctd was mixed with 0.2 mM DNA duplex in 50 mM sodium phosphate, 50 mM NaCl (pH 6.0) with 90% H2O/10% D2O. 2D 1H-15N HSQC spectra were collected with gradual addition of netropsin at concentrations of 0.05 mM, 0.1 mM, 0.2 mM, 0.3 mM and 0.5 mM, on a Bruker Avance 500 MHz spectrometer with a triple-resonance cryoprobe.
Distance restraints were derived from 3D 1H-15N-edited NOESY-HSQC and 3D 1H-13C-edited NOESY-HSQC experimental data. Intermolecular NOEs were obtained from 3D F1-15N/13C-filtered, F2-13C-edited NOESY and 2D F1-15N/13C-filtered, F2-15N-edited NOESY spectra. Additional intermolecular NOEs were obtained by analyzing the 3D 1H-15N-edited NOESY-HSQC and 3D 1H-13C-edited NOESY-HSQC spectra of MvaTctd/DNA complex. Protein dihedral angle restraints were obtained using TALOS+ [67]. Restraints on side chain χ1 angles were derived based on the intra-residual NOEs patterns.
For the structure calculation of free MvaTctd, initial structures were generated by CYANA with restraints from the CANDID module [68]. The initial structures were then used as filter models to refine the NOE assignments and distance restraints using SANE [69]. These refined distance restraints were then used to calculate the refined structures with DYANA module of CYANA. This procedure was carried out iteratively as the refined structures can be used as the filter models for next round of SANE-DYANA calculation. When there was no distance violation larger than 0.5 Å, 100 structures with the lowest target function values from the 200 structures calculated with DYANA were selected for further refinement using AMBER 12 [70] using the generalized Born (GB) solvation model [71]. SANE was also used for the refinement process until there was basically no distance violation bigger than 0.2 Å and no angle violation was bigger than 5°. The top 20 structures with the lowest AMBER energies were selected for representation, and a mean structure was generated using SUPPOSE and energy minimized by AMBER 12. The quality of structures was analyzed using PROCHECK-NMR [72].
For the complex structure calculation, structures of MvaTctd and DNA were first calculated and refined separately, as described above. In addition to the NOE distance restraints, theoretical restraints for B-form DNA were also used in defining the structure of DNA. The α, β, γ, δ, ε and ζ backbone torsion angles of DNA were restrained to ranges -60° ± 30°, 180° ± 30°, 60° ± 35°, 130° ± 50°, 225° ± 75° and -95° ± 35° (or 180° ± 30°), respectively. Sugar pucker was fixed to C2’-endo by setting the pseudorotation phase angle P to a range of 135° ± 45° as implemented in AMBER program. Glycosidic torsion angle χ were set to anti conformation with a value of -120° ± 40°. Hydrogen bonds were used to maintain the typical Watson-Crick base-pairing. The protein/DNA complex structure was obtained and further refined in AMBER 12 using the GB solvation model by combining the MvaTctd and DNA structures with intermolecular NOE restraints. Briefly, coordinates of protein and DNA were arbitrarily combined, and the complex structure was calculated with the addition of intermolecular distance restraints all set to 50 Å, and gradually decreased to 20 Å, and 8 Å, then to the actual restraint distance, while the weight for intermolecular distance restraints was increased from 0 to 2 to 20 to 25 kcal/mol · Å2. The top 20 structures with lowest AMBER energies from the calculated 100 structures were selected and analyzed using PROCHECK-NMR. The mean structure was generated using SUPPOSE and energy minimized by AMBER 12. DNA helical and groove parameters were analyzed using the program CURVES+ [73].
The cupA1 340 bp fragment was PCR amplified from P. aeruginosa PAO1 genomic DNA using primers GT044 (5’ GCGAAGCCGTGGTTCGAGTTGTT) and GT045 (5’ ATCCCGGCCTCTCTTGCTTGTCTT). A 204 bp fragment of the PA3900 gene was PCR amplified from the same genomic DNA using primers GT049 (CCGCAGGTGGCTGAACA) and GT050 (CGAATGCGGTGCGTTGATGG). PCR products were 5' end-radiolabeled with γ-32P ATP using T4 polynucleotide kinase (New England Biolabs). 400 nM DNA fragment, 4 μL γ-32P ATP (3,000 Ci/mmol, 10 mCi/mL, Perkin Elmer), 1x polynucleotide kinase buffer and T4 polynucleotide kinase enzyme (New England Biolabs) were incubated in a total of 40 μL at 37°C for 30 min. The reaction was stopped by the addition of 1 μL of 0.5 M EDTA and excess radioisotopes were removed using a G-25 spin column (GE Healthcare Life Sciences). Spin column resin was resuspended by vortexing the column upside down for 30 sec. The cap was loosened one-quarter turn and the bottom closure was snapped off. The column was then placed into a 1.5 mL microcentrifuge tube and spun at 2,800 rpm (735xg) in a tabletop centrifuge. The column was then placed into a fresh 1.5 mL microcentrifuge tube and the radiolabeling reaction was applied. DNA was eluted by spinning at 2,800 rpm (735xg) for 2 min in a tabletop centrifuge. 760 μL H2O was added to a final volume of 800 μL to dilute the DNA to a working stock of 20 nM. DNA was aliquoted and stored at -20°C. 1 μL of this stock will yield a final concentration of 1 nM in a 20 μL EMSA binding reaction. 1 nM radiolabeled DNA was incubated with varying concentrations of protein in binding buffer (15 mM HEPES pH 7.9, 40 mM KCl, 1 mM EDTA, 0.5 mM DTT, 5% glycerol). Addition of varying amounts of protein altered the concentrations of solutes from tube to tube. Buffer conditions were therefore normalized such that each reaction contained 8 mM Tris pH 8.0, 15 mM HEPES pH 7.9, 40 mM NaCl, 40 mM KCl, 1.4 mM EDTA, 0.5 mM DTT, 9.95% glycerol. Binding reactions were incubated at room temperature for 15 min before the addition of excess cold (unlabelled) DNA, where appropriate. The samples were further incubated another 15 min at room temperature. 2.5 μL of 10x DNA loading dye (10 mM Tris-HCl pH 7.5, 10 mM EDTA, 65% sucrose, 0.3% bromophenol blue) was added to each reaction and samples were loaded onto a 6% native polyacrylamide gel that had been pre-run for 1 h at 100 volts at 4°C. Samples were run at 70 volts for 165 min at 4°C, dried in a Gel Dryer (Labnet International) for 1 h at 80°C and exposed overnight on a storage phosphor screen (GE Healthcare Life Sciences/Molecular Dynamics). Gels were visualized the following morning using a Typhoon 9400 imager with an image resolution of 50 μm. The apparent dissociation constant, which corresponds to the protein concentration when the DNA is half bound, was determined by curve fitting based on quantified intensities of unbound DNA bands. The average of four replicates and their standard errors is reported.
PBM experiments were performed as previously described [29]. Sequence encoding full length MvaT from P. aeruginosa was cloned into vector pTH6838 [74] using the isothermal assembly method [75] to generate a vector producing a chimeric protein with glutathione S-transferase (GST) attached to the N-terminus of MvaT. Data was parsed with a custom Python script (written by and available from W.W.N) and plotted using the ggplot2 data visualization package [76].
ITC experiment was performed using MicroCal iTC200 system (GE Healthcare) at 283 K. 0.15 mM DNA was placed in the cell and titrated with MvaTctd or its mutants (2.3 mM–4.2 mM) injected in 2.4 μL aliquots. Corresponding “protein to buffer” controls were performed for background correction. ITC titration data were analyzed using Origin 7.0 (OriginLab) provided with the instrument. Standard deviation was calculated by according to the fit by Origin.
Cells of P. aeruginosa were permeabilized with sodium dodecyl sulphate and CHCl3 and assayed for β-galactosidase activity as described previously [77]. Assays were performed twice in triplicate on separate occasions. A representative data set is shown.
Plasmid pPSV-MvaT-V, that allows expression of the WT MvaT with a C-terminal vesicular stomatitis virus-glycoprotein epitope-tag (MvaT-V), has been described previously [30]. Site-directed mutagenesis of mvaT was carried out by the PCR to introduce mutations specifying amino acid substitutions R80A, K81A, K83A, K97A, N100A, K102A, K105A and K108A. Mutant coding sequences were digested with BamHI and XhoI and cloned into pPSV-MvaT-V cut with the same enzymes, to generate plasmids pPSV-MvaT(R80A)-V, pPSV-MvaT(K81A)-V, pPSV-MvaT(K83A)-V, pPSV-MvaT(K97A)-V, pPSV-MvaT(N100A)-V, pPSV-MvaT(K102A)-V, pPSV-MvaT(K105A)-V and pPSV-MvaT(K108A)-V.
Production of the MvaT-V proteins was confirmed by Western blotting using an anti-VSV-G antibody (Sigma-Aldrich). An antibody against the α-subunit of RNA polymerase (Neoclone) was used to control for differences in protein loading.
|
10.1371/journal.pntd.0004698 | An In Vitro/In Vivo Model to Analyze the Effects of Flubendazole Exposure on Adult Female Brugia malayi | Current control strategies for onchocerciasis and lymphatic filariasis (LF) rely on prolonged yearly or twice-yearly mass administration of microfilaricidal drugs. Prospects for near-term elimination or eradication of these diseases would be improved by availability of a macrofilaricide that is highly effective in a short regimen. Flubendazole (FLBZ), a benzimidazole anthelmintic registered for control of human gastrointestinal nematode infections, is a potential candidate for this role. FLBZ has profound and potent macrofilaricidal effects in many experimental animal models of filariases and in one human trial for onchocerciasis after parental administration. Unfortunately, the marketed formulation of FLBZ provides very limited oral bioavailability and parenteral administration is required for macrofilaricidal efficacy. A new formulation that provided sufficient oral bioavailability could advance FLBZ as an effective treatment for onchocerciasis and LF. Short-term in vitro culture experiments in adult filariae have shown that FLBZ damages tissues required for reproduction and survival at pharmacologically relevant concentrations. The current study characterized the long-term effects of FLBZ on adult Brugia malayi by maintaining parasites in jirds for up to eight weeks following brief drug exposure (6–24 hr) to pharmacologically relevant concentrations (100 nM—10 μM) in culture. Morphological damage following exposure to FLBZ was observed prominently in developing embryos and was accompanied by a decrease in microfilarial output at 4 weeks post-exposure. Although FLBZ exposure clearly damaged the parasites, exposed worms recovered and were viable 8 weeks after treatment.
| Onchocerciasis and lymphatic filariasis are debilitating diseases caused by infections with filarial nematodes. The World Health Organization aims to eliminate these infections as public health problems. Despite prolonged control efforts, including chemotherapy through mass drug administration (MDA), transmission and infections persist. Addition of a microfilaricide that is efficacious in a short regimen would enhance prospects for achieving elimination goals. We investigated the long-term effects of the macrofilaricidal drug, flubendazole (FLBZ), on Brugia malayi. Adult parasites were exposed in culture to FLBZ at pharmacologically relevant concentrations (100 nM—10 μM) for up to 24 hr prior to implantation into the abdominal cavity of a jird for long-term maintenance. The greatest drug effect was on embryogenesis; morphological damage was most evident in early developmental stages. There was also a decrease in the release of microfilaria (mf) from the adult. Interestingly, no damage was observed to fully formed mf. Although further studies are required to determine to what extent these findings can be extrapolated to a field setting, an exposure profile which may produce similar effects in vivo has been defined.
| Infections with filarial parasites that cause lymphatics filariasis (LF) and onchocerciasis can lead to debilitating symptoms and cause great economic losses in endemic countries [1,2]. Control measures have relied on mass drug administration (MDA) of either ivermectin or diethylcarbamazine with albendazole since the Global Programme to Eliminate Lymphatic Filariasis (GPELF), the Onchocerciasis Control Programme (OCP) and the African Programme for Onchocerciasis Control (APOC) were created with the aim of eliminating LF and onchocerciasis as public health problems [3,4]. These drugs appear to act mainly as microfilaricides in an MDA setting that provides yearly dosing for an extended period of time to achieve elimination or local eradication [5]. With the recent decline in individuals reported to be infected with LF and onchocerciasis [6–9], the goal of elimination/control set by the World Health Organization [10] is closer to being achieved. Yet there remain a large number of individuals infected with these parasites. Additionally, MDA programmes for onchocerciasis within Africa are geographically limited due to severe adverse events associated with acute killing of Loa loa microfilaria (mf) in individuals bearing high parasitemia following treatment with ivermectin [11]. The introduction of a safe macrofilaricidal drug into control programs would is predicted to greatly enhance the ability to eliminate these infections in a timely manner.
Flubendazole (FLBZ), a benzimidazole (BZ) anthelmintic, is a candidate macrofilaricide for use in onchocerciasis and LF control programs. Initially introduced for treatment of infections of livestock animals with gastrointestinal (GI) parasitic nematode infections [12], FLBZ was subsequently approved for the same indication in humans [13], for which it is highly efficacious [14,15]. FLBZ has exhibited very high macrofilaricidal efficacy when administered parenterally in experimental filarial models [16–18] and in a human trial in onchocerciasis [19]. Unfortunately, available formulations of the drug afford very limited oral bioavailability. Additionally, the formulation used for parenteral dosing in the human onchocerciasis study (19) led to severe injection site reactions, and its development was not pursued.
Recent efforts have been made to re-formulate FLBZ to enable oral dosing [18,20,21]. Definition of the pharmacokinetic profiles needed for efficacy with an orally-bioavailable formulation would be facilitated by knowledge of the time-concentration exposure profiles at which FLBZ is detrimental to the survival of adult filariae. Previous data show that exposure to pharmacologically relevant concentrations of FLBZ or its bioactive reduced metabolite (R-FLBZ) in culture elicits damage to the hypodermis, developing embryos, and intestine of adult female B. malayi, but this damage is not accompanied by apparent changes in motility or viability [22]. Developing an exposure-efficacy profile in vitro can assist in the definition of target pharmacokinetic profiles for dose selection in advanced development. We adapted a transplant model for B. malayi in an effort to define a concentration of FLBZ that would be lethal after short-term exposures (≤ 1 day). The present study examined long-term concentration-dependent effects of exposure to FLBZ in vitro in B. malayi survival and viability after recovery from naïve jirds following transplantation.
The transplant surgery was carried out under AUP 15–07 (2) and was approved by the TRS Labs Inc., Institutional Animal Care and Use Committee (IACUC).
Male jirds (Meriones unguiculatus) approximately 24–30 weeks of age (55–75 grams) were used as the source for and recipients of parasites in this study. The jirds were multiple housed (3-5/cage) in solid bottom, clear/translucent cages with bedding and wire mesh lids. The study room was maintained on a 12 hour light/dark cycle within a temperature range of 18–26°C and a relative humidity range of 30 to 70%. Jirds were fed ad libitum with an appropriate certified rodent diet and water was provided ad libitum by an automatic watering system and/or water bottles.
Adult male and female B. malayi were isolated from the peritoneal cavity of jirds >120 days post-infection as described [23,24]. Briefly, recovered adult worms were washed three times with warm (37°C) RPMI-1640 medium supplemented with 100 U/mL penicillin, 100 μg/mL streptomycin, and 0.4% gentamycin (Sigma-Aldrich Corp., St. Louis, MO, USA; hereafter referred to as RPMI).
Adult females were exposed to varying concentrations of FLBZ (10 μM, 1 μM, or 100 nM; Epichem Pty Ltd, Murdoch, WA, Australia) in vitro for 6, 12, or 24 hr. FLBZ solutions were prepared by dissolving the respective drug in 100% DMSO, with addition to RMPI to a final DMSO concentration of 0.1%. Control RPMI contained an equivalent concentration of DMSO. Following exposure, male and female worms (10–15 each) were rinsed with RPMI and then transplanted into the peritoneal cavity of naïve jirds as described [25]. The recipient jird was anesthetized with a 1:1:5 cocktail of xylazine:saline:ketamine and the fur was removed from the right ventral abdomen with electric clippers. The skin was wiped with 70% ethanol prior to making a 1 cm incision in the skin and body wall to expose the peritoneal cavity. Worms were aspirated into a Pasteur pipette which was inserted into the incision and the worms expressed into the peritoneal cavity. The incision was closed with Autoclip staples until necropsy 5 days, 4 weeks or 8 weeks later.
B. malayi were fixed in glutaraldehyde (5% glutaraldehyde in 0.1 M sodium cacodylate buffer, pH 7.2) for a minimum of 48 hr in preparation for histological processing. Worms from each treatment were combined into groups and coiled prior to embedding in Histogel (FisherScientific; Pittsburgh, Pennsylvania, USA), which allowed visualization of various anatomical regions in multiple worms on a single slide. Dehydration, clearing, and vacuum infiltration with paraffin were completed using a Sakura VIP tissue processor. Parasites were then embedded in paraffin with the ThermoFisher HistoCentre III embedding station. A Reichert Jung 2030 rotary microtome was used to cut 4–5 micron sections, which were dried at 56°C for 2–24 hr. Slides were stained with haematoxylin and eosin prior to examination under light microscopy.
Sections were assessed independently by three parasitologists familiar with filarial nematode morphology, including a board-certified pathologist/parasitologist (CDM), as described in [22]. Briefly, worms from two independent experiments were examined for damage to the body wall, including cuticle, hypodermis and longitudinal muscle; intestine; and reproductive tract, including the uterine wall and embryonic stages (classified as early [oocytes, early morulae, late morulae] or late [sausage, pretzel, microfilariae]); and the pseudo-coelomic cavity. For comparative analysis of drug-induced effects, tissues were classified into four categories: no damage (0), minor (1), moderate (2), or severe (3). The damage score was determined by assessing tissues for nuclear and cytoplasmic distortions, cellular size and shape, membrane integrity, accumulation of debris, and distortion of overall anatomical integrity.
Two methods of analysis were performed as previously described [22]. The first adhered to classical techniques used by histopathologists to determine tissue damage, in which all sections on a slide were surveyed, interpreted and translated into a single damage score for each tissue type. The second method involved scoring damage in each tissue type for each worm section on a slide. These scores were averaged for all sections on the slide to obtain the damage score.
Statistical analyses were performed using the GraphPad Prism 6 package. Percent recovery, microfilarial abundance, embryogram, and histology results were analysed using a two-way ANOVA between treatment groups and time points. All statistical tests were interpreted at the 5% level of significance.
Embryograms of treated worms differed substantially from those of controls (Fig 3A). Late developing stages (sausage, pretzel and stretched mf) were markedly reduced, while early developing morula were significantly increased in treated worms relative to the controls. Treated worms not only contained large numbers of degenerating embryos (Fig 3B) they were also found to contain fewer embryos overall (S2 Fig).
Tissue damage observed in worms transplanted into naïve jirds differed from that observed in vitro [22]. Damage to the intestine observed in vitro was not observed following transplantation (Table 1). Minor damage to the hypodermis was observed five days following transplantation in both the treated and control groups; however, this damage appeared to resolve by four weeks. While the damage score in treated groups at four weeks returned a statistically significant p value, the rather low score indicates this effect may not be biologically relevant. Damage to the developing embryos was the most prominent effect. While extensive damage to embryos was observed in all treatments at five days, this damage appeared to resolve in the control group by four weeks. Damage to embryos of treated worms, however, was pronounced and did not resolve at four weeks. At this time point, damage to treated embryos was significantly higher than in the controls. Morula stage embryos were the most extensively damaged (Fig 4D and 4E) while mf were unaffected (Fig 4F).
An important step in the development of an anthelmintic is identifying the exposure profile (concentration and duration of exposure) which leads to death or irreversible damage in culture. These data can be used to predict efficacious pharmacokinetic (PK) patterns, leading to more efficient selection of doses for clinical trials in diseases, such as the human filariases, that have long end points, based on PK data rather than efficacy. This is particularly true for FLBZ, which has little apparent acute toxicity for adult filariids in culture [18,22], but is highly efficacious when administered to infected animals in parenteral regimens that afford prolonged exposure to low blood levels [18]. Replicating this exposure pattern necessitates long-term parasite maintenance in culture, which has been difficult to achieve without the presence of feeder cells that may compromise the integrity of added drugs.
Efforts to reformulate FLBZ to provide an orally bioavailable macrofilaricide necessitate replication of the efficacy achieved in a “long, low” exposure profile in a “high, short” paradigm. To determine if short exposures (~1 day) to high but pharmacologically relevant concentrations of FLBZ can cause lethal damage to adult filariids, we implemented an in vivo protocol in which the long-term effects of short-term exposure to FLBZ on B. malayi can be determined by transplanting treated worms into naïve jirds following drug exposure. The choice of FLBZ concentration and duration of exposure in this study was based on pharmacokinetic studies in rats, mice and pigs[20,27–29]. In rats, a cyclodextran oral formulation resulted in a maximum plasma level of approximately 7 μM and remained at a concentration of >1 μM for 6 hours [20]. Dosing pigs with a cyclodextran oral formulation saw FLBZ in the plasma for longer durations, similar to that of the 12 hour time point in this study, albeit at lower concentrations [27]. The concentrations chosen have also been shown to elicit detrimental effects on adult females over short-term in vitro incubations [22].
Several important conclusions can be drawn from this study. First, adult filariids have the capacity to recover from damage. This is demonstrated by the observation that control worms recovered 5 days after transplantation showed clear signs of tissue damage; however, damage was not evident in control worms recovered 4 or 8 weeks post-transplantation, suggesting that the process of removal from the initial jird host and maintenance in culture for 24 hr is traumatic, but that the organisms can recover.
Second, exposure to FLBZ for 6–24 hr is deleterious to adult filariids; female B. malayi exposed to the drug in culture and recovered 4 weeks after transplant exhibited considerable tissue damage, especially to reproductive tissues. These worms were unable to produce mf.
Third, adult B. malayi are resilient to drug-induced damage; worms recovered 8 weeks after transplantation had generally resumed production of mf and had resolved the damage observed at 4 weeks after transplantation. This degree of recovery was unanticipated and suggests the possibility of a more sophisticated and robust healing response to injury in these organisms than we had anticipated.
Worm recovery from control groups was comparable to recovery rates obtained in earlier studies [30,31]. FLBZ exposure had no effect on recovery of adult worms. This result suggests that 24 hr incubation in up to 10 uM FLBZ does not cause irreversible damage that leads to worm death within eight weeks following exposure.
How FLBZ eliminates adult filariae following parenteral dosing in vivo remains unknown. Damage from prolonged, low-concentration exposure to FLBZ following parenteral administration results in slow killing in weeks to months [16–19]; efficacy may require an immune response as is thought to be the case for microfilaricidal agents [23,32,33]. It remains a goal of reformulation efforts to recapitulate the high efficacy of parenteral FLBZ with an oral regimen. The current results suggest that adult female B. malayi can recover from a short exposure to FLBZ, but leave unanswered the question of whether oral regimens compatible with field use of a macrofilaricide (up to 7 consecutive days) can cause lethality.
It is evident that the process of transplantation is stressful to the worms, as we observed damage to control worms recovered 5 days post-transplantation; however, they are able to recover from this injury (Table 1). While control worms recovered from the transplantation process, FLBZ-exposed worms were less able to do so (Table 1), especially in reproductive tissues, suggesting that they are indeed compromised by the drug.
FLBZ damages the hypodermis, intestine and developing embryos in adult female B. malayi exposed to the drug in culture [22]. In the present study, drug damage to the intestine or hypodermis was resolved after longer residence times in the host. In an early study which injected FLBZ parenterally, no alterations in intestinal cells were reported in recovered worms beyond a decrease in microtubules until the experiment ended at day six [34]. It was suggested this was due to the limited role of the intestine in nutrient acquisition by filariid parasites; unfortunately, damage to the hypodermis, which also plays a role in nutrient acquisition, was not reported. That this earlier work confirmed the lack of damage observed in the intestine and hypodermis in the present study does not mean that effects on these tissues can play no role in the macrofilaricidal activity of FLBZ. While hypodermal damage was not observed, this previous study reported a loss of intestinal microtubules, which likely plays a role in the extensive damage to the hypodermis observed in vitro [22]. However, it cannot be ignored that exclusion of the host at the time of anthelmintic exposure is a limitation to in vitro culture systems as it overlooks an important component: the host response. Since we exposed parasites to FLBZ prior to transplantation into a naïve jird, an important factor may be missing in the development of drug-induced tissue damage.
Consistent among FLBZ studies is the damage caused to developing embryos [22,34,35]. In this study, the embryonic stage which displayed the greatest damage was the morula (Table 1, Fig 4). Disruption to the integrity of morulae is prominent in treated worms which exhibited an apparent loss of cellular adhesion and dispersion of these cells (Fig 4D and 4E). Similar results were observed in O. gibsoni infected cattle administered five daily doses of mebendazole (MBZ); two weeks following treatment, degenerating morula was the most notable effect [36]. They also report the presence of various developmental stages in the same uterine section, consistent with our results where early morula are found alongside degenerating morula (Fig 4D). Four weeks following treatment degenerate morula were mixed with normal mf. Mf exhibited normal appearance for as long as eight weeks post exposure and were found mixed with oocytes and embryonic debris [36]. This degeneration of morula correlated with an increase in released egg antigen that is not observed with exposure to microfilaricidal drugs IVM and DEC [36].
Embryograms indicate that morulae were also the most abundant stage, which increased with treatment and coincided with a decrease in the proportion of later developmental stages (sausage, pretzel, mf; Fig 3). Exposure of filariae to anti-Wolbachia antibiotics resulted in a similar phenomenon, whereby the proportion of later developmental stages decreased with increasing drug exposure [37]. No previous studies have documented changes in filarial embryogram profiles associated with benzimidazole exposure. However, histological analysis of nodules from human onchocerciasis patients treated intramuscularly with FLBZ revealed that females contained only oocytes and small numbers of mf two month following treatment. The Forsyth [36] and Dominguez-Vazquez [19] studies both administered drug over a period of five days. While we see similar trends in exposed embryos, this suggests that multiple doses, rather than single exposures as in this study, may be required for high efficacy.
That there was no damage to stretched intrauterine mf at any time point is not surprising, as this observation is consistent with other studies [16,22,30,38]. In view of the fact that mf are in an arrested developmental state, there is likely to be a reduced requirement for microtubule-dependant processes with which FLBZ may interfere. However, there was a concurrent impairment of mf release, measured as reduced abundance of mf recovered from the peritoneal cavity (Fig 2). At eight weeks, there was a drastic increase in mf abundance in the control group compared to the treatment groups. There was a slight rebound in mf numbers in treated groups between 4 and 8 weeks after transplantation. Diminution in released mf could result from uterine blockage due to degenerating embryos or an inability of females to release mf. Alternatively, if B. malayi embryos follow the same developmental cycle as Onchocerca volvulus [26], it may be that embryos present at treatment were damaged and newly developed oocytes were unable to mature to mf. The potential for embryos to recover, or for the females to resume normal embryo production, is yet to be determined. Nevertheless, the limited effect observed on mf is encouraging as it supports the potential safety of FLBZ for use in L. loa endemic regions. MDA campaigns for onchocerciasis in Africa are limited due to the activity of ivermectin against L. loa mf and association with severe adverse events [11]. An ideal drug for onchocerciasis would have little effect on mf in macrofilaricidal regimens.
The present findings demonstrate that FLBZ elicits detrimental effects on developing embryos following long-term maintenance in the peritoneal cavity of jirds. Damage was most evident in early developmental stages and resulted in a decreased output of stretched mf, which is presumed to limit transmission. Conversely, FLBZ did not have direct microfilaricidal effects, which had implications for the utility of FLBZ in areas co-endemic for L. loa. This is an important observation, as current MDA programmes are restricted by the SAEs resulting from rapid mf killing by existing drugs. If FLBZ is shown to have a similar lack of microfilaricidal effect on L. loa, it could further substantiate its utility as a macrofilaricide in Loa endemic regions. It will, therefore, be critical to determine the effects of FLBZ on L. loa mf.
|
10.1371/journal.pbio.1000097 | Identification of Human S100A9 as a Novel Target for Treatment of Autoimmune Disease via Binding to Quinoline-3-Carboxamides | Despite more than 25 years of research, the molecular targets of quinoline-3-carboxamides have been elusive although these compounds are currently in Phase II and III development for treatment of autoimmune/inflammatory diseases in humans. Using photoaffinity cross-linking of a radioactively labelled quinoline-3-carboxamide compound, we could determine a direct association between human S100A9 and quinoline-3-carboxamides. This interaction was strictly dependent on both Zn++ and Ca++. We also show that S100A9 in the presence of Zn++ and Ca++ is an efficient ligand of receptor for advanced glycation end products (RAGE) and also an endogenous Toll ligand in that it shows a highly specific interaction with TLR4/MD2. Both these interactions are inhibited by quinoline-3-carboxamides. A clear structure-activity relationship (SAR) emerged with regard to the binding of quinoline-3-carboxamides to S100A9, as well as these compounds potency to inhibit interactions with RAGE or TLR4/MD2. The same SAR was observed when the compound's ability to inhibit acute experimental autoimmune encephalomyelitis in mice in vivo was analysed. Quinoline-3-carboxamides would also inhibit TNFα release in a S100A9-dependent model in vivo, as would antibodies raised against the quinoline-3-carboxamide–binding domain of S100A9. Thus, S100A9 appears to be a focal molecule in the control of autoimmune disease via its interactions with proinflammatory mediators. The specific binding of quinoline-3-carboxamides to S100A9 explains the immunomodulatory activity of this class of compounds and defines S100A9 as a novel target for treatment of human autoimmune diseases.
| What molecules and mechanisms underlie the development of autoimmune diseases such as multiple sclerosis, rheumatoid arthritis, and systemic lupus erythematosus are largely unknown. To gain some insight into the process, we use a class of chemical compounds, quinoline-3-carboxamides (Q compounds), which modify disease in both experimental animal models and in clinical trials, but whose target(s) have been elusive. We show that these Q compounds bind to a molecule called S100A9 that is expressed on the surface of various monocyte populations in the peripheral blood. Furthermore, we show that Q compounds inhibit the interaction of S100A9 with two well-known proinflammatory receptors (the Toll-like receptor 4 [TLR4] and receptor of advanced glycation end products [RAGE]). We provide a missing piece to the puzzle in that we identify S100A9 as a target of Q compound drugs and identify a new mechanism where S100A9 promotes inflammation at early stages of immune activation and thereby a role in the development of autoimmune disease.
| The medical need for novel treatments of human autoimmune/inflammatory disease is high. Quinoline-3-carboxamides (Q compounds) have been explored as treatments for autoimmune/inflammatory diseases in humans. They have shown proof-of-concept in clinical trials for the treatment of multiple sclerosis (MS) [1–4] and Type I diabetes [5], and are currently in Phase III clinical development for the treatment of MS [6] and are about to enter Phase II for the treatment of systemic lupus erythematosus (SLE). The target molecule and the mode of action of this class of compounds have remained unknown for over 25 years. Q compounds are unique in that they have a potent effect on disease development in several animal models of autoimmune/inflammatory disease without inducing suppression of adaptive immunity [7–10]. From these studies, it was obvious that the molecular target for Q compounds was novel since no known signalling pathway could explain the experimental data obtained. Furthermore, it appeared likely that the mode of action of Q compounds would be targeting early stages of immune stimulation that could be common for many autoimmune disorders while keeping the immune effector stage intact.
S100A9 [11–13] belongs to the family of calcium-binding S100 proteins and has been extensively studied [13–17]. It is expressed in granulocytes and at early stages of monocyte differentiation [14]. Complexes of S100A8 and S100A9 (S100A8/A9) are expressed and released at inflammatory sites [15,17]. A correlation between serum levels of S100A8/A9 and disease activity has been observed in many inflammatory disorders [18]. Direct inflammatory activities of the S100A8/A9 proteins include the description of mouse S100A8 as an endogenous ligand of TLR4 [17], activation of monocytes [17], and activation of endothelial cells [16,19,20]. S100A9 has also been detected on the cell surface of murine macrophages at sites of inflammation [21], but the role of surface-bound S100A9 in immunity and inflammation is still unclear. We present here data that point to a central role for S100A9 in the control of immune responses leading to inflammatory disease.
In order to identify the target molecule of Q compounds, we synthesised analogs of these compounds containing linkers that would facilitate detection of the interaction between these molecules and protein targets (Figure 1A). The molecule was modified as indicated in the R1 and R2 position to create a compound suitable for photoaffinity labelling of proteins (ABR-216893; the asterisk [*] indicates a 14C atom in this compound), or in the R1 position to create a compound (ABR-225356) labelled with FITC. Since no reliable in vitro system has been established for assaying the biological effect of Q compounds, we verified that these linker-containing Q compounds still had biological effect using the in vivo model acute experimental autoimmune encephalomyelitis (aEAE) (unpublished data). We also used the FITC-labelled compound (ABR-225356) to investigate binding to human peripheral blood mononuclear cells (PBMC). We observed that only the monocyte (CD14+) fraction was surface stained with ABR-225356 (unpublished data). On the basis of this, we decided to use human peripheral blood monocytes as a source of protein in our effort to isolate quinoline-binding molecules.
Human PBMC were separated into CD14+ and CD14− fractions, incubated with 14C labelled ABR-216893, and photoaffinity labelled. The membrane fraction of both cell populations was subsequently prepared, and the proteins separated on two-dimensional gels followed by autoradiography. Labelled proteins found exclusively on gels from the CD14+ cell fraction were extracted from the gels and identified using MALDI-TOF (matrix assisted laser desorption/ionization time-of-flight). The most prominent binding protein was identified as S100A9 and was selected for further analysis (Figure 1B). In the next step, recombinant human S100A9 was analysed for binding to a defined Q compound (ABR-215757; currently in clinical development for treatment of SLE) using surface plasmon resonance (SPR). As shown in Figure 1C, it was evident that recombinant S100A9 bound strongly to ABR-215757 coupled to solid phase. As S100A9, in most cases, is found colocalised with S100A8 at inflammatory sites, we decided to analyze homo- and heterodimeric complexes of S100A8 and S100A9 for their interaction with Q compounds. Figure 1D shows that binding was more or less exclusively restricted to homodimeric S100A9, whereas only weak binding was observed for the S100A8/A9 complex, and close to baseline levels for S100A8. Last, we determined that the Q compound/S100A9 interaction could be competed in a dose-dependent way by free compound (Figure 1E). Additional proteins were identified using photoaffinity labelling but could not be verified using follow-up SPR analysis. We concluded from these studies that human S100A9 was a potential pharmacological target for Q compounds.
S100A9 belongs to the S100 family of proteins that are known to be Ca++ binding proteins and are involved in inflammatory processes [18]. However, S100A9 has also been shown to bind Zn++, and this interaction might have conformational consequences for the protein [22]. We therefore investigated the dependence of the interaction between ABR-215757 and S100A9 for both these divalent ions. When Ca++ or Zn++ were titrated in the presence of a fixed concentration of either Zn++ or Ca++, we found that both ions were required for S100A9 binding to Q compounds (Figure 1F). If titration was carried out in the absence of either Ca++ or Zn++, the binding of S100A9 to ABR-215757 was reduced to baseline values (Figure S5). It should also be noted that the levels of Ca++ or Zn++ required for optimal S100A9/quinoline interaction are within the concentration range found for these ions in human serum [23]. We are aware of the fact that S100A9 is prone to form dimers [16,18] and complexes of even higher oligomeric states at high concentrations, and therefore expect the interaction between Q compounds and human S100A9 to be occurring primarily with at least bivalent structures of the molecule. This assumption was supported by the complex kinetics and sigmoidal-shaped dose-response curve for S100A9 binding to immobilised Q compound.
The next question was to gain some understanding concerning the mechanism whereby S100A9 binding by Q compounds could inhibit autoimmune disease. Members of the S100 protein family have been shown to interact with the proinflammatory molecule RAGE (receptor for advanced glycation end products) [18,24], but to our knowledge, there are no data in the literature showing a direct, physical interaction between RAGE and S100A9. Furthermore, it has been shown that soluble RAGE may alleviate aEAE [25]. We therefore decided to investigate whether human S100A9 was a human RAGE ligand using SPR. In this study, RAGE was covalently coupled to the sensor chip to allow exposure of the extracellular domain of RAGE to S100A9, thus reconstituting a biological model in which anchored membrane receptor interacts with soluble ligands. As shown in Figure 2A, S100A9 interacted strongly with immobilised RAGE when injected at the concentration yielding half-maximal binding to ABR-224649, almost a 6-fold higher response compared to the S100A8/A9 heterodimer. Moreover, the binding of S100A8 to RAGE was negligible. The interaction between S100A9 and RAGE was also strictly dependent on the presence of physiological concentrations of both Ca++ and Zn++ (Figure 2B). Given the similarities between the binding conditions for S100A9 interaction with RAGE and Q compounds, we proceeded to test whether a Q compound could compete for RAGE binding to S100A9. Indeed, ABR-215757 in increasing concentrations competed for RAGE-S100A9 binding in the presence of Ca++ and Zn++ (Figure 2C). Furthermore, direct binding of ABR-215757 to RAGE was not observed, indicating that the inhibition of the interaction was mediated by binding of ABR-215757 to S100A9 (unpublished data). In a separate experiment in which human RAGE/Fc and Fc alone were allowed to interact with S100A9 immobilised on the chip, we observed no interaction with Fc alone (Figure S4A). Thus, under our standard conditions, homodimeric S100A9 is the primary RAGE ligand (Figure 2D), as well as target for Q compound binding.
Having defined S100A9 as a RAGE ligand, we wanted to investigate whether other proinflammatory signalling molecules would also interact specifically with human S100A9. We had noted that one Q compound had been shown to inhibit lipopolysaccharide (LPS)-induced toxic shock [26]. We therefore decided to investigate whether TLR4 could be a possible S100A9 ligand and whether Q compounds could interfere with such interactions. Since TLR4 is inactive in the absence of the coreceptor MD2, we here used a complex of human TLR4 and MD2 for amine coupling to a biosensor chip to be used in a SPR assay. As shown in the left panel of Figure 3A, S100A9 showed strong binding when injected over immobilised TLR4/MD2 using our standard conditions, and produced a more than 5-fold higher signal than the S100A8/A9 heterodimer. Furthermore, the signal obtained after injection of S100A9 was proportional to the amount of TLR4/MD2 coupled to the solid phase (Figure S7). We could also demonstrate that the binding of S100A9 to TLR4/MD2 is TLR4 specific since the TLR4/MD2 complex interacted with immobilised S100A9 with high affinity whereas MD2 alone did not (Figure S4B). Hence, we can show here for the first time that human S100A9 is an endogenous TLR4 ligand. The interaction between human S100A9 and TLR4/MD2 was strictly dependent on the presence of both Zn++ (Figure 3B) and Ca++ (unpublished data), which could explain why this interaction has not been previously described. We then proceeded to investigate whether the Q compound ABR-215757 could interfere with human S100A9 binding to TLR4/MD2. As shown in Figure 3C, a dose-dependent inhibition of the interaction was observed, whereas only very weak inhibition was seen with a control substance (Figure S6). We also wanted to test whether soluble TLR4/MD2 could displace binding of S100A9 to immobilised TLR4/MD2.
Interestingly, TLR4/MD2, injected together with S100A9, was not only able to efficiently block the interaction of S100A9 with immobilised TLR4, but also inhibit the interaction between S100A9 and immobilised Q compound and RAGE, respectively (Figure 3D). This observation indicates that TLR4/MD2, RAGE, and Q compound compete for the same binding region in human S100A9. The TLR4/MD2 complex is known to bind LPS, and therefore we investigated whether LPS could interfere with the binding of human S100A9 to the immobilised human TLR4/MD2 receptor complex. In contrast to the dose-dependent displacement of S100A9 binding by soluble TLR4/MD2, LPS had no effect on this interaction even at 200 ng/ml (Figure S1A). Thus, human S100A9 in the presence of Ca++ and Zn++ can interact specifically with two distinct proinflammatory receptors.
With the results above, we had a foundation on which to understand the effect of Q compounds on inflammatory disease in humans. However, these compounds have also shown a broad activity in several disease models in mice [1,8–10]. Thus, we needed to validate our findings using mouse proteins. Figure 4A illustrates that very similar results to those obtained with the human proteins were obtained both with regard to mouse S100A9 binding to mouse RAGE, mouse S100A9 binding to Q compound, and mouse S100A9 binding to the mouse TLR4/MD2 fusion protein (mLPS-Trap [27]). Also, all interactions showed similar requirements for Ca++ and Zn++ (unpublished data). Furthermore, the interaction of mouse S100A9 with solid-phase mouse RAGE and mouse TLR4/MD2 could both be competed by the Q compound ABR-215757 (Figure 4B). Analogous to the human S100A9-TLR4/MD2 interaction, soluble TLR4/MD2 displaced mS100A9 binding to immobilised TLR4/MD2 in a manner independent of both LPS and MD2 (Figure S1B). Moreover, as was shown for human S100A9, homodimeric mouse S100A9 bound much stronger to immobilised Q compound, RAGE and TLR4 than as a heterodimer with S100A8 (Figure 4C). We conclude from this series of experiments that neither the interactions between S100A9 with RAGE and TLR4/MD2, nor the competition of this interaction by Q compound, are species specific.
Having determined that S100A9 interacted specifically with Q compounds, we next wanted to determine whether S100A9 would qualify as a bona fide pharmacological target for the Q compounds. To this end, we selected six compounds (see Table S1) from our chemical libraries of Q compounds [28] and tested these for their binding strength to human and mouse S100A9 and to human and mouse S100A8/A9 heterodimers, their potency in inhibiting the interaction between human and mouse S100A9 and RAGE, and their potency in inhibiting the interaction between human and mouse S100A9 and TLR4/MD2.
Multivariate analytical tools (principal component analysis [PCA] and partial least squares projections to latent structures [PLS]) were used to derive the structure-activity relationship (SAR) for the binding activity of a series of quinoline compounds to the S100A9 homodimers and the S100A8/A9 heterodimers (Table S1). When the potency of these compounds in inhibiting aEAE in vivo was directly correlated to their potency in inhibiting the interaction between human S100A9 and human RAGE, an excellent correlation was observed (R2 = 0.98) (Figure 5A).
We then proceeded to apply PCA modelling to the dataset, i.e., the five S100A9 and the two S100A8/A9 assays and the aEAE model, a two-component model with R2X = 0.97 and Q2 = 0.87, was obtained. The first model dimension reflected as much as 68% of the total variation. The principal component (PC) scores revealed differences between the homodimer and the heterodimer. An overall inspection of the loading plot (Figure 5B) reveals that the aEAE inhibition (point 6) and the S100A9 homodimer binding (points 3–5, 7, and 8) are situated close to each other on the first principal component (p[1]), indicating that strong positive correlations exist among them. On the other hand, the S100A8/A9 heterodimer binding (points 1 and 2) are more distant from the aEAE point, meaning that S100A8/A9 heterodimer binding and aEAE are not strongly correlated.
A set of five quinoline derivatives incorporating different substitution patterns at position 5 with relative binding affinities measured to the S100A9 homodimer in the mS100A9–RAGE interaction assay was used to derive the SAR of the binding activity of the quinoline derivatives towards the S100A9 homodimers. The results confirmed that the structural modifications carried out on the 5-position have a profound effect on binding affinity. The PLS evaluation resulted in a three-component model, obtained with cross validation, giving a SAR model with R2Y = 0.99 (85% + 12% + 2%) and Q2 = 0.81, which indicates that mS100A9 homodimer binds the quinoline compounds with a high structural selectivity. The observed and predicted half-maximal inhibitory concentration (IC50) values for these compounds for inhibition of mS100A9/RAGE interactions are shown graphically in Figure S2 and very similar results were obtained for all other S100A9 interactions investigated. The analysis pointed out the major importance of steric and hydrophobic factors (L, B1, π of the 5-substituent, and the acid strength of the 4-hydroxy group. Furthermore, local electrostatics at positions 4 and 5 were important for the biological activity. In an ensuing step, the SAR model was further tested using an additional quinoline derivative, i.e., ABR-212662 (Table S1). This compound was selected based upon its substitution and variation within the activity range, i.e., being unsubstituted in the 5-position and displaying low binding activity. The observed and predicted binding activities for this compound showed high correspondence and were 1,026 and 1,235 μM, respectively. Hence, this SAR model is robust and valid for prediction as used. We conclude from the data shown that S100A9 by its SAR to disease inhibition qualify as a pharmacological target molecule for Q compounds.
Having shown that S100A9 binding by Q compounds showed a SAR with their activity in inhibiting autoimmune disease, the next step in our investigation was to validate S100A9 as a drug target in vivo. We first considered the obvious experiment of using S100A9 null mice [29]. To this end, we back-crossed the S100A9−/− animals against C57BL/6 mice and induced experimental autoimmune encephalomyelitis (EAE) using MOG peptide (Figure S8). We observed that the S100A9−/− animals had a more severe disease than C57BL/6 controls, but still responded to treatment with Q compounds. This was an unexpected result given the very strong SAR between the binding strength of Q compounds to S100A9 and their potency in inhibiting aEAE (Figure 5). However, the absence of an obvious functional phenotype with a specific gene deletion does not necessarily prove that the protein it codes for has an insignificant function in an intact host. The S100 family of proteins is large and complex. For example, whereas S100A12 has been shown to be a RAGE ligand in humans [30], its gene is absent in the mouse genome [31]. S100A8 is expressed almost exclusively as a S100A8/A9 heterodimer, but whereas S100A9−/− mice are viable, the S100A8−/− genotype is embryonically lethal [32]. In addition, S100A9−/− mice show spontaneous alterations of their inflammatory response also in other experimental models [17]. Given that S100A9 convey important biological functions, it can be suspected that biological redundancy may occur in the S100A9−/− animal, in which another molecule(s), maybe from the S100 family, would serve as a ligand for RAGE, TLR4, and Q compounds. Such a molecule could have very limited function in a genetically intact animal.
To be able to perform the in vivo validation of S100A9 as a pharmacological target for Q compounds in wild-type animals, we therefore turned to an alternative approach. Hence, we decided to generate a set of monoclonal antibodies to S100A9 that could compete for S100A9 binding to RAGE and TLR4/MD2. S100 proteins are rather conserved during evolution [33,34]. Assuming that their biological function also has been conserved, one may speculate that it would be difficult to obtain antibodies to key regulatory epitopes using xenoimmunisation. We therefore elected to immunize S100A9−/− mice with recombinant human S100A9 in order to obtain antibodies to novel, potentially functional, epitopes on the S100A9 protein. Approximately 50 S100A9-specific hybridomas were obtained in this experiment, and one of these, 43/8, was used for further validation.
Figure 6 shows the basic features of the 43/8 antibody. It binds both human and mouse S100A9 (Figure S3A). The antibody will also surface stain human monocytes in fluorescence-activated cell sorting (FACS) analysis but not as brightly as the S100A8/A9-specific antibody 27E10 (Figure 6A). Fab fragments of the 43/8 antibody (Figure S3B) will also inhibit the interaction of S100A9 and RAGE, as well as S100A9 and TLR4/MD2, showing almost complete inhibition at a concentration of 200 nM (Figure 6B). Very similar results were obtained with intact 43/8 antibody, but not with an isotype control (unpublished data). We could also demonstrate that the epitope recognized by the 43/8 antibody is exposed in an optimal way only in the presence of Ca++ and Zn++. As is shown in Figure 6C, human S100A9 binds to immobilised intact 43/8 antibody with a more than 10-fold higher signal when injected with Ca++ and Zn++.
Vogl et al. [17] has demonstrated that the induction of systemic TNFα production by LPS is perturbed in S100A9−/− animals. We therefore selected this model for our in vivo validation. C57BL/6 mice were treated with the Q compound ABR-215757 2 h before being challenged intraperitoneally with 3 or 6 μg of LPS. Ninety minutes later, the animals were sacrificed, and the serum TNFα levels were determined. As shown in Figure 7, Q compound significantly inhibited TNFα production at both levels of LPS challenge, with the effect being most pronounced using the 6-μg challenge. We then proceeded to use the 43/8 Fab in the same assay. As shown in Figure 7, after challenge with 3 μg of LPS, the TNFα production was significantly inhibited also using 43/8 Fab. We conclude from this experiment that Q compounds, or an antibody Fab fragment that mimics Q compounds in the sense that it inhibits S100A9 interaction with TLR4 and RAGE, can inhibit a biological activity shown to be compromised in S100A9−/− animals [17]. Hence, we consider these data as an in vivo validation of S100A9 as a pharmacological target for Q compounds.
Although the prognosis and clinical management of patients with chronic inflammatory diseases has improved during the last decades, there is still a great medical need for new treatments. Treatments especially that do not compromise immune function and that are suitable for chronic dosing are urgently needed. A group of compounds that fulfils these criteria are the Q compounds that have been in clinical development for over two decades, but whose molecular target and mode of action are unknown. The present investigation defines S100A9 as one molecular target for Q compounds, and their detailed effects on autoimmune disease in mice and humans [1–5,7–10] can now be studied in a more rational fashion. Interestingly, the effect of Q compounds resembles the phenotype recently described for the S100A9 knockout model, in that a diminished TNFα response after LPS challenge was observed [17]. That the target molecule for these compounds, S100A9, interacts with signalling pathways that are early and potent mediators of proinflammatory responses (RAGE and TLR4) could shed some light on the ability of Q compounds to mediate this effect without causing overt suppression of adaptive immunity. It can be speculated that the interference with proinflammatory signalling at the level of antigen-presenting cells may suppress the reactivity of autoimmune T cells. In addition, human S100A9 can now be regarded as a novel therapeutic target for the treatment of autoimmune diseases.
The interactions between S100A9 and Q compound, RAGE, or TLR4/MD2 were all strictly dependent on physiological levels of Ca++ and Zn++. The interaction between S100A9 and Zn++ especially appears to induce a dramatic structural change in the protein, which also was shown to significantly affect the binding of the 43/8 monoclonal antibody. It is interesting to note that Zn++ has been shown to have a profound impact on the structure of other S100 proteins [35]. Also, elevated levels of Zn++ are seen at inflammatory sites, and many extracellular proteins contain Zn++ binding sites [23]. Thus, it can be speculated that the elevation of Zn++ is a feedforward signal for inflammation and acts by inducing conformational changes in proteins and thereby facilitating novel interactions. That both RAGE and TLR4/MD2 interact with the same surface on human S100A9 and are competed by Q compounds is intriguing. We have investigated several of the mouse and human TLRs for binding to the same interphase on human S100A9 without finding additional targets (unpublished data). However, we expect that other proinflammatory molecules will eventually be shown to interact with the same molecular surface on human S100A9, and are also open for the idea that other forms of S100 protein combinations may bind to proinflammatory mediators. The common theme remains that a molecular form of S100 proteins can interact with several proinflammatory mediators as a mechanism to modulate the quality of the immune response and inflammatory reactions.
At first glance, the data presented here are in conflict with previously published data [17]. In this study, biosensor experiments were conducted with recombinant murine S100A8 immobilized on the chip using amine coupling. Binding to murine TLR4/MD2 fusion protein (mLPS-Trap) was demonstrated, as was the ability of murine S100A8, but not S100A9, to stimulate TNFα production of bone marrow cells from wild-type mice. In the present study, however, S100 proteins were injected over a surface with immobilized TLR4 to preserve the Ca/Zn conformation of S100A9 that is a prerequisite for binding activity. Under these conditions, S100A8 only showed weak interaction with TLR4. The biological reason for this discrepancy may be explained by the fact that the biological function of S100A8 and S100A9 is regulated in a complex manner, which additionally may differ between mice and men. For example, human S100A9 activates integrin affinity of CD11b/CD18 on monocytes, whereas human S100A8 has no effect [36]. Vice versa, murine S100A8 activated murine macrophages whereas murine S100A9 was inactive. Regulatory effects of human S100A9 on tubulin metabolism are completely abrogated by phosphorylation of S100A9 at threonine 113 [37]. This MAPK p38-dependent phosphorylation site is not conserved in murine S100A9. It seems therefore likely that murine S100A9 may mediate so far undefined regulatory mechanisms in vivo, which may be responsible for the discrepancy between different experimental findings between mice and men in vivo and in vitro. Activation of the innate immune system is crucial for initiation and amplification of many inflammatory responses and autoimmune diseases. During this process, endogenous danger signals called alarmins or damage-associated molecular patterns (DAMPs) play a pivotal role via interaction with specific pattern-recognition receptors [38]. S100A8 and A9 have been identified as important endogenous DAMPs due to their activation of TLR4 [14,17]. Thus, specific blocking of S100 proteins, as presented here, represents the first report of targeted intervention with a DAMP-mediated inflammatory process, which has already shown pharmacological activity in mice and men [4,10].
S100A8 and S100A9 are two members of the S100 protein family. Multivariate analytical tools were used to derive the SAR for the binding activity of a series of Q compounds towards the S100A9 homodimer and the S100A8/9 heterodimer, with the assumption that similar analogs bind to the same binding site in a similar binding mode. The results indicate that the Q compounds bind the S100A9 homodimers with high structural selectivity and that this binding showed a strong correlation to their ability to inhibit autoimmune disease. On the other hand, the correlation between Q compound binding to S100A8/A9 heterodimers and inhibition of autoimmune disease was poor. The bulk of S100A8 and S100A9 protein is expressed as S100A8/A9 heterodimers and most of this protein is found as soluble protein in serum. S100A8 and S100A9 are also expressed on the cell surface of monocytes [12,15]. Whether the pharmacological activity of Q compounds is primarily mediated by blocking soluble or membrane-bound S100A9 will be a subject for future studies.
Murine and human S100A8, S100A9, S100A8/A9, and human S100A12 were either produced recombinantly in Escherichia coli or purified from granulocytes; essentially as described [39]. Mouse TLR4/MD2 fusion protein (mLPS-Trap) was obtained through collaboration [27]. Carrier-free recombinant human RAGE/Fc (hRAGE), human IgG1Fc (hFc) mouse RAGE/Fc (mRAGE), human TLR4/MD-2 complex (hTLR4/MD-2), human MD-2, mouse anti-hRAGE (clone #176902), mouse anti-hTLR4 (clone #285219), and rat anti-mTLR4 (clone #267518) were purchased from R&D Systems. The mouse anti-human S100A9 monoclonal antibody (clone 43/8) was produced in-house by immunisation of S100A9−/− mice. LPS from E. coli was obtained from Sigma. Protein concentration was determined using the microtiter plate BCA assay from Pierce with bovine serum albumin as standard, or by absorbance measurement at 280 nm using molar absorption coefficient. Biotinylation of the 43/8 monoclonal antibody was made using the NHS-LC-biotin reagent from Pierce. The antigen binding fragment (Fab) of mouse anti-human S100A9 monoclonal antibody 43/8 was prepared by enzymatic digestion on immobilized ficin using the mouse IgG1 Fab preparation kit from Pierce. Gel electrophoresis under denaturing conditions was run on 4%–12% Bis-Tris NuPAGE gels with MES-SDS as running buffer (Invitrogen). For details on the synthesis and features of quinoline compounds, see Jönsson et al. [28] and references therein.
The following antibodies and second steps reagents were used for surface stain of human PBMC, CD14-APC, mouse IgG1 (BD Biosciences Pharmingen), 27E10-FITC (BMA Biomedicals), Streptavidin-Alexa Fluor 488 (Invitrogen), and biotinylated 43/8 monoclonal antibody. Stained cells were analyzed by flow cytometry on a FACSCalibur (Becton Dickinson) using CellQuest software. For protein isolation, PBMC were divided into CD14+ and CD14− populations using positive selection of CD14+ cells with magnetic beads (Miltenyi Biotec). Both cell types were incubated with ABR-216893 in the dark on ice after which the cells were exposed for a light source for 30 min, lysed, and protein extracts prepared as described [40]. The proteins were subsequently subjected to conventional two-dimensional gel analysis and autoradiography. Radioactive spots that were present selectively in extracts from CD14+ cells were isolated, the proteins eluted, trypsin digested, and prepared for analysis in a Bruker Reflex III instrument (Bruker Daltonik) using protocols and software supplied by the manufacturer.
TNFα induction after intraperitoneal challenge with LPS was performed essentially as described [17]. In brief, mice were pretreated for 2 h with 10 mg/kg ABR-215757 or PBS, after which 3 or 6 μg of LPS was injected intraperitoneally. After an additional 90 min, the animals were sacrificed, and the level of TNFα in blood was determined using commercial TNFα antibodies (eBioscience; http://www.ebioscience.com).
S100A9−/− mice (10 wk of age) were injected intraperitoneally with 100 μg of recombinant human S100A9 precipitated in alum. Six weeks later, the mice were boosted with the same dose of antigen and the spleen cells fused to SP2/0 5 d later. S100A9 reactive clones were selected using ELISA, and positive clones were subcloned by limiting dilution.
The SPR analysis was carried out with the Biacore 3000 system (GE Healthcare). Sensor chips, amine coupling kit, immobilization and running buffers, and regeneration solutions were obtained from GE Healthcare. Working solutions of all reagents used for Biacore analysis were prepared in 0.01 M Hepes, 0.15 M NaCl (pH 7.4) containing 0.005% v/v Surfactant P-20 (HBS-P; GE Healthcare) by buffer exchange on Fast Protein Desalting Micro-Spin Columns from Pierce. ABR-215757 was immobilised onto a CM5 chip through an amino-linker (ABR-224649). Other reagents (i.e., RAGE, TLR4/MD2, and various antibodies) were immobilised to the aimed density using random amine coupling chemistry. Activity of ligands after immobilisation was tested by injecting specific antibodies (unpublished data). In some experiments, S100A9 was immobilized either by random amine coupling or with a known orientation (i.e., by sulhydryl group conjugation to the only cysteine in S100A9 at position 3).
Binding to the various surfaces was performed by injecting the analyte at a flow rate of 30 μl/min in a physiological buffer (HBS-P) containing 1 mM Ca2+ and 10 μM Zn2+ as proposed for S100A8/9 by Robinson et al. [16]. A typical analysis cycle consists of: (1) pumping running buffer for 1 min to obtain a stable baseline; (2) injection of sample for an appropriate period of time (association); (3) pumping running buffer for 2.5 min (dissociation); (4) injection of a short pulse (30 s) of 15 μl 10 mM glycine-HCl (pH 2.0) (regeneration); and (5) pumping running buffer for 2 min (stabilisation after regeneration) at a flow rate of 30 μl/min. As S100A9 is a calcium-binding protein and shown to require low concentrations of Zn2+ to adapt a biologically active conformation [16], titration of Zn2+ and Ca2+ for optimal binding of S100A9 to immobilised ABR-224649, RAGE, and hTLR4/MD2 was performed. In one experiment, S100A9 was injected into a buffer with a fixed Ca2+ concentration (1 mM) and Zn2+ in the range 0–50 μM. In a second experiment, Ca2+ was varied from 0–2 mM at a fixed Zn2+ concentration (10 μM). In subsequent analyses, regeneration was carried out under more mild conditions, i.e., by injecting 30 μl of HBS-P containing 3 mM EDTA (HBS-EP; GE Healthcare) for 60 s, to prolong the lifespan of the chip.
In order to study displacement of S100A9 binding to immobilised Q compound, RAGE, and TLR4/MD2, S100A9 at a concentration yielding approximately half-maximal binding was incubated in the absence or presence of serially diluted Q compounds. Compounds were also injected over the immobilised surfaces in the absence of S100A9 to exclude, or make possible correction for, any direct binding of compound to the surface.
Evaluation was carried out using BIAevaluation Software version 3.2 (GE Healthcare). The response at steady-state was obtained by fit of sensorgrams to standard binding models, where appropriate, or calculated as responses at late association or early dissociation phase using the Steady state affinity function in BIAevaluation. Affinity was determined from kinetic analysis (on- and off-rates) or as the apparent affinity after plotting responses versus concentration of analyte in a saturation curve. In the inhibition assay format, the competitor concentration yielding 50% inhibition (IC50) was calculated by fitting data to a one-site competition model in GraphPad Prism.
Multivariate analytical tools (PCA and PLS) were used to derive the SAR for the binding activity of a series of quinoline compounds to the S100A9 homodimer and the S100A8/A9 heterodimer (Table S1). The software SIMCA-P+ 11 (Umetrics AB; http://www.umetrics.com) was used to conduct the multivariate data analysis. A number of physicochemical descriptors for size, lipophilicity, and electronic characteristics were used to correlate structural or property descriptors of the compounds with their biological activities. The 5-substituents of the quinoline compounds were described by two dimensionally–based structure descriptors, i.e., STERIMOL parameters (L, B1, and B5) as steric parameters, and the substituent constant π as a hydrophobic parameter. The experimentally determined and assigned carbon 13C-NMR chemical shifts of atoms from positions 3 to 10 on the quinoline template were used to reflect local electrostatics. There were only minor 13C-NMR shift differences between carbons in positions 2, 11, and 1′-4′, and the 13C shifts of these latter atoms were not used when establishing the SAR models. The acidity constants (pKa) in water of the corresponding ortho-substituted benzoic acid derivatives were used to correlate structure and acid strength of the 4-hydroxy group. The steric and hydrophobic parameters used were 2D-based structure descriptors known from the literature [41]. 13C NMR spectra were recorded with an operating frequency of 125.8 MHz. Spectra were recorded in D2O with a small addition of NaOD at ambient temperature. The shift scale was referenced to 3-(trimethylsilyl)-propane sulfonic acid Na-salt (TSPSA) defined as 0.00 ppm. Signals from two rotameric forms in equilibrium (E/Z isomerism) were obtained from the anion form of the compounds, and only the major form was used.
A training set of five quinoline derivatives with structural diversification performed at position 5 of the quinoline ring system was used for the SAR. PCA was used to uncover any relationship between the binding activities of the quinoline derivatives at the S100 proteins and the inhibitory effect of these derivatives in the aEAE model. PLS was then used to model the relationship between the physiochemical descriptors used to characterize the compounds and their biological responses. The PCA included the binding affinities towards murine and human S100A9 homodimers, murine S100A8/A9 heterodimer, and 50% effective dose (ED50) values from an aEAE mice model. The PLS analysis included binding affinity from the murine S100A9 homodimer and a total of 13 physicochemical variables used to describe the same set of five compounds. All variables were mean centred and scaled to unit variance.
|
10.1371/journal.ppat.1000846 | SREB, a GATA Transcription Factor That Directs Disparate Fates in Blastomyces dermatitidis Including Morphogenesis and Siderophore Biosynthesis | Blastomyces dermatitidis belongs to a group of human pathogenic fungi that exhibit thermal dimorphism. At 22°C, these fungi grow as mold that produce conidia or infectious particles, whereas at 37°C they convert to budding yeast. The ability to switch between these forms is essential for virulence in mammals and may enable these organisms to survive in the soil. To identify genes that regulate this phase transition, we used Agrobacterium tumefaciens to mutagenize B. dermatitidis conidia and screened transformants for defects in morphogenesis. We found that the GATA transcription factor SREB governs multiple fates in B. dermatitidis: phase transition from yeast to mold, cell growth at 22°C, and biosynthesis of siderophores under iron-replete conditions. Insertional and null mutants fail to convert to mold, do not accumulate significant biomass at 22°C, and are unable to suppress siderophore biosynthesis under iron-replete conditions. The defect in morphogenesis in the SREB mutant was independent of exogenous iron concentration, suggesting that SREB promotes the phase transition by altering the expression of genes that are unrelated to siderophore biosynthesis. Using bioinformatic and gene expression analyses, we identified candidate genes with upstream GATA sites whose expression is altered in the null mutant that may be direct or indirect targets of SREB and promote the phase transition. We conclude that SREB functions as a transcription factor that promotes morphogenesis and regulates siderophore biosynthesis. To our knowledge, this is the first gene identified that promotes the conversion from yeast to mold in the dimorphic fungi, and may shed light on environmental persistence of these pathogens.
| The dimorphic fungi are the most common cause of invasive fungal disease worldwide. In the soil, these fungi grow as mold that produce infectious spores; when inhaled into the warmer lungs of a mammalian host, the spores convert into yeast, which cause infection. The change in shape between mold and yeast is a crucial event in the lifecycle of these fungi. The molecular regulation of this morphologic switch, or phase transition, is poorly understood. The goal of our research was to identify and characterize novel gene(s) that govern the phase transition in dimorphic fungi using Blastomyces dermatitidis as a model organism. Using insertional mutagenesis, we identified a gene, SREB, which encodes a transcription factor that affects phase transition and regulates the production of iron-gathering molecules or siderophores. When SREB is deleted, B. dermatitidis fails to complete the conversion from yeast to mold, grows poorly at environmental temperature, has yellow-orange colony pigmentation, and cannot properly repress the biosynthesis of siderophores. We also identified two types of siderophores produced by B. dermatitidis. To our knowledge, SREB is the first gene identified that promotes the conversion from yeast to mold, a process important for survival in the environment and generation of infectious spores.
| The endemic dimorphic fungi are comprised of seven ascomycetes that include Blastomyces dermatitidis, Histoplasma capsulatum, Coccidioides immitis, Coccidioides posadasii, Paracoccidioides brasiliensis, Sporothrix schenckii, and Penicillium marneffei. These fungi possess the unique ability to switch between two different morphologies, yeast and mold, in response to external stimuli [1]. In nature, they grow as mycelia that produce conidia, which are the infectious particles; when aerosolized spores are inhaled into the warmer lungs of a mammalian host, they convert into pathogenic yeast and cause necrotizing infection [1]. The dimorphic fungi collectively are the most common cause of invasive fungal disease worldwide and account for several million infections each year [2]. Unlike opportunistic fungi, such as Cryptococcus or Aspergillus, the dimorphic fungi can infect both immunocompetent and immunocompromised hosts [3]–[5]. The size of the inhaled inoculum and the integrity of the cell-mediated immune system influence the extent and severity of infection [1],[3]. Clinical manifestations range from asymptomatic infection to symptomatic disease and include pneumonia, acute respiratory distress syndrome, and disseminated disease involving multiple organ systems [1],[3].
The ability of the dimorphic fungi to switch between the two different morphologies is crucial for pathogenesis. Although temperature is postulated to be the major stimulus that induces phase transition, other stimuli, including carbon dioxide tension, steroid hormones, and oxidative stress influence this morphologic switch [1], [6]–[9]. Phase transition is a complex process that involves the coordinated expression and repression of many genes in response to external stimuli, which alters cell wall composition, metabolism, intracellular signaling, and morphology [10]–[13]. The identification of DRK1 (dimorphism-regulating kinase-1) in B. dermatitidis and H. capsulatum offered strong genetic evidence that phase transition is required for pathogenicity [10]. DRK1 functions as a global regulator and has pleotropic effects on the cell, controlling morphogenesis, cell wall composition, sporulation, expression of yeast-phase specific genes, and virulence. DRK1 null mutants remain locked in the mycelial phase, fail to sporulate or express the essential virulence factors BAD1 (Blastomyces adhesin-1 in B. dermatitidis) and CBP1 (Calcium binding protein-1 in H. capsulatum), and are avirulent in a murine model of infection [10]. Three additional genes, RYP1, RYP2, and RYP3, have been described that regulate morphogenesis in H. capsulatum. Silencing the expression of RYP1, 2 or 3 results in hyphal growth at 37°C and inappropriate sporulation [12],[13].
The goal of this study was to identify and characterize additional genes that regulate the phase transition in dimorphic fungi, using B. dermatitidis as a model system. While progress has been made in identifying genes that regulate the morphological transition from mold to yeast, to our knowledge, no genes have been identified that regulate the switch in the other direction in the dimorphic fungi – that is, from the yeast to mold form. The mold form is believed to be required for the growth and survival of the dimorphic fungi in the environment by enabling propagation in soil and transmission to humans through the generation of conidia. Herein, we describe a gene, SREB, identified through insertional mutagenesis, which impacts multiple disparate fates in B. dermatitidis, including the phase transition of yeast to mold, cell growth at 22°C, and the biosynthesis of siderophores.
Agrobacterium tumefaciens-mediated DNA transfer was used to mutagenize haploid, uninucleate conidia of B. dermatitidis strain T53-19. Following selection with hygromycin, 22,000 transformants were visually screened by light microscopy for morphologic alterations including growth as hyphae or pseudohyphae at 37°C or as yeast at 22°C. In this study, one of the mutants identified by the screen, 3-15-1, was characterized in detail. This mutant, unlike the parent strain, was pigmented yellow and failed to complete the conversion from yeast to mold (Figure 1A, 1B). Southern blot hybridization demonstrated a single site of insertion (Figure S1).
The genomic DNA flanking the insert in 3-15-1 was amplified using adapter PCR, sequenced, and analyzed using a BLASTn search against the genome sequence of B. dermatitidis strain 26199. No rearrangements or deletions were identified in the DNA flanking the insert. Additional BLAST analyses indicated that the insert interrupted a region 692 base-pairs (bp) upstream of a predicted open reading frame with nucleotide homology to Pencillium chrysogenum SREP, which encodes a GATA transcription factor that regulates the biosynthesis of siderophores [14]. We named this homolog SREB (siderophore biosynthesis repressor in Blastomyces) in B. dermatitidis.
FGENESH analysis of the nucleotide sequence predicted that SREB contained a 1909 nucleotide (nt) coding region interrupted by two short introns (88 and 74 nt). Each intron was located in a zinc-finger coding region and contained the expected 5′-splice donor (GTNNGT) and 3′-splice acceptor (pyrimidine-AG) sequences [15]. The length, location, and number of introns interrupting the open reading frame were conserved among B. dermatitidis SREB, H. capsulatum SRE1, A. nidulans SREA, and N. crassa SRE [16]–[18]. The SREB coding region was predicted to encode a 636 amino acid protein.
The predicted amino acid sequence of SREB had homology to siderophore biosynthesis repressors in other fungi including Aspergillus nidulans SREA, Penicillum chrysogenum SREP, Neurospora crassa SRE, Ustilago maydis URBS1, Schizosaccharomyces pombe FEP1, Candida albicans SFU1, Cryptococcus neoformans CIR1, and Histoplasma capsulatum SRE1 (Figure 1C) [14], [16]–[22]. SREB contained several conserved domains characteristic of GATA transcription factors that regulate iron assimilation, including two zinc finger motifs separated by a cysteine-rich region (CRR) and a C-terminus predicted to have a coiled-coil domain (Figure 1C) [17],[23]. With the exception of C. neoformans CIR1, fungal GATA transcription factors that regulate the acquisition of iron contain two zinc fingers [22]. This zinc finger arrangement is unique because most GATA transcription factors in fungi contain only one zinc finger [17]. The cysteine residues in each zinc finger of SREB were arranged in a conserved class IV motif, Cys-X2-Cys-X17-Cys-X2-Cys [24]. The cysteine-rich region contained four conserved cysteine residues, which have been demonstrated to coordinate the binding of iron in H. capsulatum [16].
Mutant 3-15-1 failed to convert from yeast to mycelia or produce conidia following a shift in incubation temperature from 37°C to 22°C (Figure 1A). In contrast, the parent strain T53-19 converted to mycelia when grown at 22°C and produced conidia. Mutant 3-15-1 accumulated little biomass at 22°C, but remained viable (as measured by the exclusion of 0.2% eosin stain), and converted to normal yeast morphology when the incubation temperature was shifted back to 37°C (data not shown).
The yellow-orange pigmentation of mutant 3-15-1 and the predicted amino acid sequence suggested that SREB functioned as a repressor of siderophore biosynthesis. Deletions of SREB homologs in P. chrysogenum (SREP), A. nidulans (SREA), and N. crassa (SRE) produce similar discoloration [14],[17],[18]. To assess for the dysregulation of siderophore biosynthesis in the insertion mutant, we used a colorimetric assay to detect the production of hydroxymate-type sideophores in culture supernatants [25]. Under iron-poor conditions, both T53-19 and 3-15-1 produced an abundance of siderophores as measured by this assay (data not shown). Under iron-replete conditions, mutant 3-15-1 continued to produce siderophores, whereas parent strain T53-19 repressed siderophore biosynthesis (Figure 2A).
To determine if the mutant phenotype was from altered expression of SREB, and not due to another mutation incurred during insertional mutagenesis, we set out to complement the mutant phenotype. Insertional mutant 3-15-1 was re-transformed via A. tumefaciens to provide an intact gene copy of SREB and its endogenous promoter. Complemented strains A5 and D5 grew as white colonies that did not discolor the medium, suppressed siderophore production under iron-replete conditions (10 µM FeSO4), and converted fully to mycelia when grown at a temperature of 22°C (Figure 2A-C). Retransformation of 3-15-1 with a vector lacking SREB did not complement the mutant phenotype (empty vector strain) (Figure 2A-C). Whereas Northern analysis demonstrated a reduction in the abundance of SREB transcript in mutant 3-15-1 compared to the parental strain, message levels were overexpressed in both complemented strains (Figure 2D). Thus, complementation reversed the mutant's phenotypic defects, supporting the idea that the insert was responsible for the dysregulation of siderophore biosynthesis and the alteration in morphogenesis.
To confirm that SREB represses the biosynthesis of siderophores and affects morphogenesis in B. dermatitidis, we disrupted this gene in wild-type isolate 26199 using homologous recombination. To minimize the probability that the phenotype observed in mutant 3-15-1 was unique to strain T53-19, we used a different B. dermatitidis strain, 26199, to generate a null mutant. The rate of allelic replacement was 0.04% (1/2670). The null mutant, SREBΔ, grew as yellow-pigmented colonies that discolored the surrounding medium and failed to properly repress siderophore biosynthesis when iron was abundant (Figure 3A, 5B). The intensity of pigmentation was dependant on exogenous iron and independent of temperature (37°C vs. 22°C) (data not shown). In contrast, the parent strain grew as white-colored yeast and repressed the production of siderophores under iron-replete conditions as measured by the ferric perchlorate assay (Figure 3A, 5B). SREBΔ failed to complete the yeast-to-mold phase transition following a shift in temperature from 37°C to 22°C, did not exhibit radial growth, and accumulated little biomass at 22°C (Figure 3A, 3B). The defect in phase transition persisted during prolonged incubation (>14 days) at 22°C; however, a few hyphal strands would develop and could only be observed by light microscopy. Similar to insertional mutant 3-15-1, SREBΔ remained viable at 22°C (as measured by 0.2% eosin exclusion) and converted back to yeast following a shift in temperature from 22°C to 37°C (data not shown). In the yeast form, the SREBΔ mutant grew at the same rate as the parent strain (Figure 3C). The morphologic defect at 22°C was independent of exogenous iron concentrations (data not shown).
Analysis of the null mutant by PCR indicated disruption of SREB and the absence of any deletion or rearrangment of the genomic DNA flanking the transgene (data not shown). Southern blot analyses demonstrated replacement of SREB with a hygromycin resistance cassette and the absence of additional deletions in the genomic DNA flanking the transgene in SREBΔ (Figure 4A-E). Northern analysis demonstrated the loss of SREB transcript in SREBΔ (Figure 4F).
To confirm the phenotype in SREBΔ was due to disruption of the siderophore biosynthesis repressor gene, we re-transformed the null mutant using A. tumefaciens to insert a copy of SREB. Complemented strains grew as white-colored colonies and properly suppressed the biosynthesis of siderophores when iron was abundant (Figure 5A, 5B). Following a temperature shift from 37°C to 22°C, complemented yeast strains converted to mold (Figure 5C). This conversion was slower in the complemented strains (14–17 days) when compared to the wild-type isolate (<7 days) (data not shown). The complemented strains underwent radial growth at 22°C; however, colony expansion was less than the wild-type isolate (data not shown). Prolonged incubation did not result in catch-up growth. Analysis of transcript abundance demonstrated restoration of message levels in C#25 and overexpression in C#6 when compared to wild-type and SREBΔ strains (Figure 5D).
To test if the expression of SREB was influenced by the concentration of exogeneous iron, we grew wild-type B. dermatitidis strain 26199 under iron-poor and –replete conditions. Northern blot analysis demonstrated that the expression of SREB was increased during conditions of iron abundance and repressed when iron was limited (Figure 4F).
In fungi, the expression of genes that encode proteins involved with iron assimilation are often co-expressed or -repressed when iron is limited or abundant, respectively. To investigate whether this was also true in B. dermatitidis, we analyzed the expression of several genes in response to exogenous iron. Under iron-poor conditions, B. dermatitidis wild-type strain 26199 induced the expression of genes involved in the biosynthesis of siderophores (SIDA), transport of ornithine from the mitochondria into the cytosol (AMCA), uptake of siderophores (MIRB, MIRC), and a bZIP transcription factor (HAPX) (Figure 6). Conversely, these genes were repressed when iron was abundant (Figure 6). The disruption of SREB de-repressed the expression of each of these genes. Thus, SREB regulates genes involved in siderophore biosynthesis and uptake in B. dermatitidis (Figure 6).
To further characterize the regulatory role of SREB on siderophore biosynthesis, we used LC/MS and reverse-phase HPLC to identify the specific type(s) of siderophores secreted by B. dermatitidis wild-type and null mutant yeast cells. Starting with wild-type cells grown under iron-limited conditions, siderophores from culture supernatant were isolated using column chromatography. Mass spectroscopy of the eluate showed two large peaks at 4.16 and 7.26 minutes with molecular weights of 538.2 and 822.2 that correspond to dimerum acid and coprogen, respectively (Figure 7A-C). Reverse-phase HPLC of the eluate and comparison of retention times to siderophore standards confirmed the identities of these siderophores (Figure 7D). Under iron-replete conditions, wild-type B. dermatitidis repressed the biosynthesis of dimerum acid and coprogen (Figure 7D). In contrast, the null mutant continued to produce both siderophores (Figure 7D).
To identify candidate genes regulated by SREB that may promote the phase transition, we first used MAST analysis to search the Blastomyces genome for GATA transcription factor-binding motifs in intergenic regions located ≤2000 bp upstream of predicted genes. Our initial search for the classic GATA transcription factor-binding motif, HGATAR, revealed the presence of this motif upstream of nearly all B. dermatitidis genes. This finding is similar to Schrettl et al., who found widespread distribution of this motif in Aspergillus fumigatus [26].
An extended version of the HGATAR motif, ATC-w-gAta-a, has been recently described and was demonstrated to occur at a 5.4-fold higher frequency in the promoter of genes regulated by A. fumigatus SREA, an SREB homolog, when compared to the entire A. fumigatus genome [26]. We revised our strategy and searched for this extended motif in the promoter of genes in the B. dermatitidis genome. We identified a total of 1,213 genes with at least one of the following motifs located ≤2 kb upstream of the start codon: ATC-(A/T)-GATA-(A/G), ATC-(A/T)-GATA-(T/C), ATC-(A/T)-GATT-A, ATC-(A/T)-GATC-A, ATC-A-GATG-A, ATC-C-GATA-A, and ATC-A-AATA-A. This gene-set included genes involved in siderophore biosynthesis and uptake (i.e. SIDA, MIRB, AMCA). Two or more upstream GATA motifs were present in 232 (19.1%) in the gene-set. Hwang and colleagues identified the motif (G/A)-ATC-(A/T)-GATA-A upstream of siderophore biosynthesis and transport genes regulated by SID1 in H. capsulatum [27]. We found this longer motif upstream of 271 (22.3%) of our 1,213 MAST-identified genes; however, MIRB and MIRC, both involved in siderophore uptake, lacked the motif.
To classify the 1,213 candidate genes into functional categories and facilitate further analysis, we annotated the predicted protein products of these genes as well as the complete B. dermatitidis predicted proteome against the eukaryotic orthologous groups (KOG) database. The results, shown in Table 1, indicate that the KOG-annotated GATA-containing genes fall into many categories of gene function (i.e. transcription, RNA metabolism, signal transduction, cell remodeling and metabolism). The frequency of KOG-annotated genes with upstream GATA motifs within a particular KOG category was compared to the frequency of genes in the same KOG category within all KOG-annotated genes in the B. dermatitidis genome. Three KOG categories were significantly over-represented in the candidate gene-set harboring GATA sites: amino acid transport and metabolism (KOG code E), secondary metabolites biosynthesis, transport and catabolism (KOG code Q), and lipid transport and metabolism (KOG code I) (Table 1 and Table S1). This suggests that these cellular process pathways may be important for SREB regulation, although it does not exclude a role for the GATA-containing genes in other KOG groupings.
In a complimentary approach to identify genes that may be regulated by SREB, we performed a preliminary microarray analysis. Using an expression array with 70-mer oligonucleotides representing the 10,567 open reading frames of B. dermatitidis strain 26199, we used two-color spotted analysis to compare isogenic wild-type vs. SREBΔ at 37°C and at 22°C 48 hours after the temperature shift downward (data not shown). At least 38 of the genes identified by MAST analysis were differentially expressed (increased or decreased by ≥2-fold), including seven genes classified by KOG to be involved in lipid transport and metabolism. To validate the microarray results, we performed quantitative RT-PCR on a subset of four genes found to be altered in expression; three from the lipid transport and metabolism KOG category, and one from the carbohydrate metabolism category. At 22°C, the null mutant strain failed to upregulate the expression of a lipid transfer protein and acetoacetyl-CoA synthase (Figure 8). Conversely, the expression of a peroxisomal dehydratase was over-expressed at 37°C and 22°C, when compared to the wild-type isolate (Figure 8). We also confirmed the altered expression of a glycosyl hydrolase postulated to be involved in cell-wall remodeling. In the null mutant, this gene is over-expressed at 37°C and 22°C, when compared to the wild-type isolate (Figure 8). Thus, we have begun to identify candidate genes and processes that may be direct or indirect targets of SREB and contribute to the phase transition from yeast to mold.
The use of A. tumefaciens-mediated DNA transfer for insertional mutagenesis has advanced our understanding of the endemic dimorphic fungi at the molecular level [10],[12],[13],[28]. We used this technology to mutate B. dermatitidis conidia and screen for transformants with altered morphology during growth at 22°C and 37°C. Analysis of mutant 3-15-1 uncovered a GATA transcription factor, SREB, which regulates siderophore biosynthesis and affects morphology in B. dermatitidis. GATA transcription factors are zinc-finger proteins that bind conserved motifs to induce or repress gene expression [16],[22],[26],[27],[29]. These genes are found widely in eukaryotes, but they function differently in fungi, plants, and animals [30],[31]. In fungi, GATA transcription factors regulate diverse functions including the response to blue light, switching of mating-type, uptake of nitrogen, pseudohyphal growth during nitrogen starvation, biosynthesis of siderophores, and iron assimilation [17],[22],[29],[32],[33].
Our analysis indicates that SREB has pleotropic effects in B. dermatitidis - it promotes the transition from yeast to mold at environmental temperature and represses the biosynthesis of siderophores. Following a shift of incubation temperature from 37°C to 22°C, the insertional and null mutants were unable to complete the phase transition or accumulate significant biomass when compared to the parent strain. To our knowledge, B. dermatitidis SREB is the first gene identified in the dimorphic fungi that promotes the conversion of yeast to mold. Much of the field's attention has been focused on genes that regulate the phase transition from mold to yeast; only a few genes have been identified that regulate growth or morphology in the dimorphic fungi at environmental temperature (i.e. 22–25°C). In H. capsulatum, the mold-specific gene MS8 regulates mycelial morphology and growth, but not the phase transition [34]. In P. marneffei TupA is required for maintenance of mycelial morphology at 25°C; null mutants convert to mycelia following a temperature shift from 37°C to 25°C, but revert to yeast morphology with prolonged incubation [35].
We hypothesize that B. dermatitidis SREB binds DNA to regulate many genes that, in turn, control such disparate functions as phase transition and the response to abiotic stress, including iron availability. Using MAST analysis we identified a large number of genes with putative GATA transcription factor binding sites. When compared to the entire B. dermatitidis genome, candidate genes involved with the biosynthesis of secondary metabolites as well as amino acid and lipid metabolism were found to be over-represented. Some of these candidate genes were indeed altered in expression in SREBΔ, as detected in preliminary microarray analysis and validated by RT-PCR. The enrichment of genes involved in secondary metabolism and amino acid metabolism were not unexpected, in part, because SREB regulates siderophore biosynthesis, a process that requires the transport and metabolism of amino acids. The abundance of genes containing GATA binding sites involved in lipid transport and metabolism was surprising. To our knowledge, regulation of lipid metabolism and transport in fungi by GATA transcription factors has not been described.
Changes in fatty acid metabolism in the dimorphic fungi are associated with the phase transition and are postulated to impact morphogenesis [36]–[42]. Exposure of H. capsulatum mycelia to unsaturated fatty acids prolongs the mold-to-yeast conversion following a shift in temperature from 25 to 37°C [36]. In contrast, treatment with saturated fatty acids accelerates the phase transition [36]. In C. immitis, exposure to exogenous fatty acids alters the conversion of spherules to mycelia [37]. Reduced expression of the Δ9-desaturase gene, OLE1, in C. albicans, impairs hyphal formation [38]. Differences in the concentration of unsaturated fatty acids (oleic and linoleic acids) and unsaturated sphingolipids (N-2′-hydroxy-(E)-Δ3-octadecenoate) have been described in the yeast and mold forms of H. capsulatum and P. brasiliensis [39]–[42]. In P. brasiliensis, several genes involved in lipid metabolism, have been demonstrated to be phase-regulated [43]. Thus, further investigation of genes involved in fatty acid metabolism may clarify the mechanism by which SREB promotes the phase transition from yeast to mold.
B. dermatitidis insertional and null mutants have multiple alterations in the regulation of iron assimilation, as indicated by their yellow-orange appearance, constitutive production of siderophores, and derepression of iron-regulated genes during conditions of iron abundance.
Iron acquisition must be tightly regulated for proper cellular function and to avoid toxicity due to iron overload [17],[44]. Under iron-replete conditions, SREB represses genes involved in the production (SIDA, AMCA) and uptake (MIRB, MIRC) of siderophores. AMCA encodes a transferase that shuttles ornithine from the mitochondria to the cytosol [44]. The first step in siderophore biosynthesis involves the conversion of ornithine into N5-hydroxy-L-ornithine, which is catalyzed by an L-ornithine-N5-monooxygenase encoded by SIDA [45]. Siderophores secreted into the environment bind iron and then can be taken up by the cell through permeases such as MIRB and MIRC [46]. Analysis of the B. dermatitidis genome did not reveal an ortholog to A. nidulans MIRA, which facilitates the uptake of xenosiderophores, specifically enterobactin [46]. Deletion of SREB resulted in de-repression of SIDA, AMCA, MIRB, and MIRC expression under iron-replete conditions. Similar to P. chrysogenum SREP, N. crassa SRE, and A. nidulans SREA null mutants, disruption of SREB in B. dermatitidis resulted in discoloration of the fungus [14],[18],[45]. In addition, we identified two extracellular siderophores, dimerum acid and coprogen, produced by B. dermatitidis when grown under iron-poor conditions. When iron is abundant, SREB represses the biosynthesis of both these siderophores.
Similar to A. nidulans and H. capsulatum, the expression of B. dermatitidis SREB is upregulated when iron is abundant, and repressed when iron is limited [16],[17]. Repressors of siderophore biosynthesis are not uniformly regulated at the transcriptional level in other fungi, as orthologs of SREB including SRE, URBS1, FEP1, and SFU1 are constitutively expressed regardless of exogenous iron concentrations [18]–[21]. SREB is expressed as a single transcript, similar to SRE and URBS1 [18],[19]. In contrast, SREP, SREA, and FEP1 are expressed as two separate transcripts due to the presence of two transcriptional start sites [14],[17],[20].
B. dermatitidis SREB may participate in a regulatory circuit with the bZIP (basic leucine zipper) transcription factor, HAPX. Computational analysis of the promoter region of HAPX in B. dermatitidis revealed putative GATA binding sites. Moreover, iron-poor conditions induced HAPX expression in wild-type B. dermatitidis, whereas iron abundance reduced its expression. In A. nidulans, HAPX represses SREA as well as genes that encode iron-dependent proteins such as CYCA (cytochrome C), ACOA (aconitase), LYSF (homoaconitase) when iron availability is limited [44]. We found that deletion of SREB resulted in the expression of HAPX under iron-poor and iron-replete conditions.
Our findings support the idea that B. dermatitidis SREB functions as a transcription factor that regulates the biosynthesis of siderophores and promotes the conversion from yeast to mold. We propose that SREB inhibits genes involved with the biosynthesis and uptake of siderophores under conditions of iron abundance. Our findings also suggest that SREB affects phase transition independently of iron assimilation, perhaps, by altering the expression of genes involved with lipid metabolism or cell wall remodeling. The iron-related defects do not explain the failure to convert from yeast to mold since growth under iron-poor conditions had no effect on the defect in morphogenesis. GATA transcription factors in other fungi have been demonstrated to regulate morphogenesis as well as the response to temperature. S. cerevisiae ASH1 encodes a GATA transcription factor that inhibits mating-type switching and induces filamentous growth under conditions of nitrogen limitation [29]. C. neoformans CIR1, an ortholog of B. dermatitidis SREB, regulates genes involved in reductive iron assimilation and siderophore transport, but also genes critical for virulence including those required for thermotolerance, capsule production, and melanin biosynthesis [22].
In summary, we identified and characterized a GATA transcription factor that represses the biosynthesis of siderophores and promotes the phase transition from yeast to mold. To our knowledge, B. dermatitidis SREB is the first gene identified in dimorphic fungi that promotes the conversion of yeast to mycelia. By using bioinformatic and expression analyses we identified several genes whose expression may be directly or indirectly regulated by SREB. We investigated a sample of these genes, including ones in KOG categories for lipid and carbohydrate metabolism, and found that their expression is affected by the deletion of SREB. Future work will strive for a more complete description of how SREB promotes the yeast to mold phase transition. Because growth in the mold form is thought to be essential for the survival of dimorphic fungi in nature and the generation of infectious particles, SREB may be needed for the evolutionary maintenance of this species. The generation of an SREB null mutant provides a unique opportunity to elucidate the SREB regulon and identify genes that govern growth in the mold form, as well as other traits in this human fungal pathogen.
Blastomyces dermatitidis strains used in this study included T53-19 and American Type Culture Collection (ATCC) 26199. T53-19 sporulates, but is weakly virulent in a murine model of infection, and ATCC strain 26199 is highly virulent, but does not sporulate [10],[47]. The genome of strain 26199 has been sequenced by the Genome Sequencing Center at Washington University (http://genome.wustl.edu). B. dermatitidis yeast and mold were grown on Histoplasma macrophage medium (HMM), 3M medium (3M), Potato dextrose agar (PDA), or Middlebrook 7H10 agar medium containing oleic acid-albumin complex (7H10; Becton Dickinson and Company, Franklin Lakes, NJ) [48]–[50]. Agrobacterium tumefaciens strain LBA1100 harboring the Ti helper plasmid pAL1100 (gift from C. van den Hondel; Leiden University, The Netherlands) was maintained on Luria-Bertani (LB) medium supplemented with 0.1% glucose, spectinomycin 100 µg/ml, and kanamycin 100 µg/ml once transformed with a binary vector [28].
Conidia from B. dermatitidis strain T53-19 were mutagenized using A. tumefaciens containing pBTS165 [10],[28],[51]. This binary vector contains a resistance cassette, hygromycin phosphotransferase (hph), integrated into the T-DNA that is driven by a glyceraldehyde-3-phosphate dehydrogenase (gpdA) promoter derived from Aspergillus nidulans [10]. Conidia harvested from mycelial cultures by manual disruption were counted using a hemocytometer, suspended in phosphate buffered saline (PBS) to a final concentration of 2×107/ml, and co-cultivated with A. tumefaciens (6×108 cells/ml) on a Biodyne A nylon membrane (Pall Gelman, Ann Arbor, MI) on induction medium containing 200 µM acetosyringone (IMAS medium) [28]. After 72 hours of incubation at 22°C, the biodyne membranes were transferred to 3M medium supplemented with hygromycin 100 µg/ml (AG Scientific Inc., San Diego, CA) and cefotaxime 200 µM (Sigma-Aldrich), and incubated at 37°C or 22°C. Individual transformants were visually screened by light microscopy for altered morphology: growth as hyphae or pseudohyphae at 37°C or yeast at 22°C. Replica plates were used to identify transformants that lost viability upon shifting the incubation temperature from 22°C to 37°C.
Adaptor PCR was used to amplify DNA flanking the pBTS165 insert from insertional mutant 3-15-1 [52]. Following the digestion of genomic DNA by restriction enzymes StuI, HpaI, and XmnI, which do not cut in pBTS165, adaptors were ligated to the restriction fragments using T4 DNA ligase (New England Biolabs, Ipswich, MA). PCR was performed using primers specific for the adaptors and pBTS165. The PCR products were separated by agarose gel electrophoresis and purified using the QIAquick gel extraction kit (Qiagen, Valencia, CA) and sequenced by the DNA Sequencing Laboratory at the University of Wisconsin Biotechnology Center. Sequence flanking the insert was analyzed using GSC (Genome Sequencing Center) BLAST (http://genome.wustl.edu/tools/blast) and National Center for Biotechnology Information (NCBI) tBLASTx (http://blast.ncbi.nlm.nih.gov/Blast.cgi). FGENESH was used to identify predicted exons and introns in the SREB gene (www.softberry.com).
Two vectors, pBTS4-KO1 and pBTS4-KO2, were used to delete SREB in B. dermatitidis strain 26199 by homologous recombination and resulted in two null mutants, T1#23 and T12#16, respectively. Although both null mutants had similar phenotypes, T1#23 contained an additional 2,214 bp deletion in the 5′ untranslated region that was upstream of the disrupted SREB gene. Herein, T12#16, which has no additional deletions, is referred to as SREBΔ. Plasmid pBTS4-KO2 contained 1611 bp of 5′ upstream sequence and 1747 bp of coding and 3′ downstream sequence flanking hph. The 1611 bp and 1747 bp products were amplified from B. dermatitidis 26199 genomic DNA using F and R primers containing SacI, BbsI, SbfI, or ClaI restriction sites (F-1611-SacI 5′-TTTGAGCTCACTTTACTCTTCGGACGGGTTTT; R-1611-BbsI 5′-TTTTCGATTGTCTTCAGCCAAAAGCCCCGTCATTCCTGT; F-1747-SbfI 5′-TT-TCCTGCAGGTTGCAGCGTGAGGCGGAAGA; R-1747-ClaI 5′-TTTATCGATTGACAGGGCAG-GCTACATA). PCR products were separated by agarose gel electrophoresis, purified using QIAquick PCR purification kit (Qiagen, Valencia, CA), sequenced, and ligated into pBTS4 in sequential fashion following restriction digest to flank the hph-resistance cassette [53]. After sequence and restriction digest analyses confirmed integration of the ligated PCR fragments, pBTS4-KO2 was electroporated into A. tumefaciens strain LBA1100 [28]. B. dermatitidis strain 26199 (2×107 yeast/ml) was transformed with A. tumefaciens containing pBTS4-KO2 (6×108 bacteria/ml) on Biodyne A membranes on IMAS medium. After 72 hours of incubation at 22°C, the Biodyne membranes were transferred to HMM medium supplemented with 10–20 µM FeSO4, hygromycin 25 µg/ml, cefotaxime 200 µM, and incubated at 37°C. Transformants were visually screened for yellow pigmentation. The null mutant was cloned to obtain individual colonies and establish a line of cells. SREB gene deletion was confirmed by PCR, and Southern and Northern blot analyses (see below).
Insertional mutant 3-15-1 was re-transformed with pBTS47-11+13 using A. tumefaciens-mediated DNA transfer. This plasmid contained the SREB coding region, 1990 bp of 5′ sequence upstream of the start codon, 603 bp of 3′ sequence downstream of the stop codon, and a nourseothricin resistant cassette. Genomic DNA was amplified using primers ggp11-XbaI (5′-TTTCTAGAACAACTACCTCTACATGACACT-GC) and ggp13-SbfI (5′-TTTCCTGCAGGGAGCCTTTTCTTTCTGTCAA). The PCR products were separated by agarose gel electrophoresis, purified using QIAquick PCR gel extraction kit (Qiagen, Valencia, CA), sequenced, and ligated into pBTS47 to generate pBTS47-11+13. The null mutant, SREBΔ, was re-transformed by A. tumefaciens with pBTS47-5331, which contains the SREB coding region, 2655 bp of 5′ sequence upstream of the start codon, 603 bp of 3′ sequence downstream of the stop codon, and a nourseothricin resistant cassette. The protocol for A. tumefaciens-mediated DNA transfer was similar to that described in the previous section. Transformants were screened for white colony pigmentation on HMM medium supplemented with 20 µM FeSO4, nourseothricin 25 ug/ml (Werner Bioagents, Germany), and cefotaxime 200 µM at 37°C incubation.
B. dermatitidis was grown to late log phase in liquid HMM at 37°C incubation. Genomic DNA was extracted using the method described by Hogan and Klein [54]. Southern blot hybridization was performed as described [28],[55]. The fate of the transforming DNA in the insertional mutant was determined using probes specific for T-DNA and non-T-DNA sequences. An 822 bp amplicon constructed using primers 5′-CGATG-TAGGAGGGCGTGGATA and 5′-GCTTCTGCGGGCGATTTGTGT was used to probe hph within the T-DNA. An 8 kb BglII restriction fragment generated from pBTS4 was used to probe the non-T-DNA sequence. Deletion of SREB in the null mutant was analyzed using PCR-generated probes specific for the SREB coding region (1303 bp; 5′-CCCGCTCTTTGCTTAACC-CGTATG and 5′-CTGGTGATAAAGAAGGGCTGAA), hph (822 bp; 5′-CGATGTAGGAGGGCG-TGGATA and 5′-GCTTCTGCGGGCGATTTGTGT), 5′ region flanking SREB (1663 bp; 5′-ACTT-TACTCTTCGGACGGGTTTTC and TATCTGCGCTTTTGGTAGTAGGAG), and the 3′ region flanking SREB (1747 bp; 5′-TTGCAGCGTGAGGCGGAAGA and 5′-ACAAATCGTAGCACCAG-TC). All probes were radiolabeled with α-32P dCTP using a Prime-a-Gene labeling system (Promega, Madison, WI). Unincorporated radionucleotides were removed using ProbeQuant G50-micro columns (GE Healthcare, Buckinghamshire, UK). Following hybridization, the blot was washed sequentially with low stringency (0.25 M NaPO4, 2% SDS, 1 mM EDTA) and high stringency (0.04 M NaPO4, 1% SDS, 1 mM EDTA) solutions, exposed to a storage phosphor screen (Molecular Dynamics, Sunnyvale, CA) and scanned using a Storm 660 imaging system (Molecular Dynamics, Sunnyvale, CA).
Ferric perchlorate was used to measure siderophore production semi-quantitatively [25]. B. dermatitidis was grown at 37°C in liquid 3M or HMM under iron-poor or replete (10 µM FeSO4) conditions. Iron-poor media consisted of HMM or 3M prepared with F-12 Ham's nutrient mixture lacking FeSO4, or trace elements lacking FeSO4, respectively. In addition, exogenous iron was not added to these media. As the yeast entered stationary growth (A600 = 3.5−4.0), culture supernatants were collected, filtered (0.2 µM), and added to a ferric perchlorate solution (5 mM Fe(ClO4)3 in 0.1 N HCl). Absorbance was measured at 425 or 495 nm. Plasticware was used whenever possible. Glassware was treated with 2N HCl to remove residual traces of iron [56]. Analysis of variance (ANOVA) was used to analyze the results from the ferric perchlorate assay. Tukey's Honest Significant Difference method was used to adjust the p-values for multiple comparisons.
B. dermatitidis was grown to mid-log phase at 37°C in liquid HMM with no added iron (iron-poor medium), 10 µM FeSO4, or 50 µM FeSO4. Total RNA was extracted using the phenol-guanidinium thiocyanate-1-bromo-3-chloropropane extraction method [55]. In brief, yeast were washed with PBS, beaten with beads, and treated with TRI Reagent followed by 1-bromo-3-chloropropane (Molecular Research Center Inc., Cincinnati, OH). RNA was precipitated using a 1∶1 concentration of isopropanol and a high salt solution (Molecular Research Center Inc., Cincinnati, OH), washed with 75% ethanol, and resuspended in water that was pre-treated with diethyl pyrocarbonate (DEPC; Calbiochem, San Diego, CA). Total RNA was further purified using RNeasy kit (Qiagen, Valencia, CA) and enriched for mRNA using oligo(dT)-polystrene chromatography (Sigma-Aldrich). Northern hybridization was performed as described using 2.0-2.3 µg poly(A)+-enriched mRNA per sample [55]. Gene expression was analyzed using probes constructed by PCR against SREB (SreF 5′-CCCGCTCTTTGCTTAACCCGTATG; SreR 5′-CTGGTGATAAAGAAGGGCTGAA) SIDA (SidA-F1 5′-AGACAGTACTCAAGAACGACAA; SidA-R1 5′-GCTGTCATCGCTGGGCTTTAGTGC), MIRB (MirB-F 5′-CTCCTCCTCGTCGCTTTCGCACTA; MirB-R 5′-CCCTGAGGTCCCCGT-AGATGAG), MIRC (MirC-F 5′-TGATGGCATTCTCAACCTCCC; MirC-R 5′-AACCTGCGGTGAT-GAAACCAC), AMCA (AmcA-F 5′-GTCCGCATTACTCATCTG; AmcA-R 5′-CGCCTCATAAATC-GTAA), HAPX (HapX-F 5′-CCGGTACCCCTCAAGCCCACAACT; HapX-R 5′-AAATACTTCAAC-ACGCCCATAACG), and actin (Actin-F 5′-TCGGCCGTCCTCGCCATC and Actin-R 5′-TCCAG-ACTCGTCGTAGTCCTGC).
Total RNA was extracted from B. dermatitidis wild-type and SREB null mutant strains grown in HMM at 37°C and 22°C in a similar fashion as described above; modifications included grinding cells frozen in liquid nitrogen in a mortar and pestle. Wild-type and SREB null mutant cells were grown for 48 hours at 22°C prior to RNA extraction. RNA, at 10 ug/sample, was treated with Turbo DNase (Applied Biosystems/Ambion, Austin, Tx) and further purified using RNeasy kit (Qiagen, Valencia, CA). cDNA was generated from 1 ug of DNase-treated RNA using iScript cDNA synthesis kit (Bio-Rad, Inc., Hercules, CA). Real-time PCR reactions were comprised of 1x SSoFast EvagGreen supermix (Bio-Rad), 0.5 mM of each primer, and 1 ul of 10-fold diluted cDNA template in a total volume of 10 ul. All reactions were performed in triplicate for two biological replicates. Real-time PCR was performed using a Bio-Rad iCycler MyiQ. Cycling conditions were 1 cycle at 95°C for 30 seconds followed by 40 cycles of 95°C for 5 seconds and 60°C for 10 seconds. Melting curve analysis was performed following the completion of the PCR. Gene expression was normalized relative to the expression of alpha-tubulin based on R (relative expression) = 2−ΔCt, ΔCt = Cttarget gene–Cttubulin [57]. Primers used to amplify transcripts from the following genes were: Lipid transfer protein (BDBG_03618-1F 5′- CCATCAATGCTGCCATCAAC; BDBG_03618-1R 5′-GGTCTCACCCTTGTCGTTTG), glycosyl hydrolase (BDBG 03183-1F 5′-GCTCTCCCAAGACATACATCAG, BDBG_03183-1R 5′-CCAT-AGCAAACTTCCCAAAAG), peroxisomal dehydratase (BDBG_00052-1F 5′-CCCATTGTGCTA-ACCTTCAAG, BDBG_00052-1R 5′-AACTCCATCCGTCGCCTC), acetoacetyl-CoA synthase (BDBG_09522-1F 5′-GCTCTCGGCACGCTCATAC, BDBG_09522-1R 5′-GGTGGTGACGG-GAGAAATG) and alpha-tubulin (BDBG_00020-2F 5′-GGTCACTACACCATCGGAAAG-3′, BDBG_00020 2R 5′-CTGGAGGGACGAACAGTTG).
The annotated genome and predicted proteome of B. dermatitidis strain SLH14081 was used for MAST analysis and KOG annotation. The genome of this strain (75.35 Mb; 9,555 genes) has been sequenced and annotated by the Broad Institute (www.broadinstitute.org/annotation/genome/blastomyces_dermatitidis/MultiHome.html and ACBT01000000). The absence of annotation in the sequenced genome of 26199 precluded its use for computational analysis.
MAST/MEME (multiple em for motif elicitation) software in unix (version 4.2.0) was used to identify GATA transcription factor binding motifs in the genome of B. dermatitidis SLH14081 [58]. A fifth-order Markov background model was built for SLH14081 using the MEME utility fasta-get-markov. To find the location of previously identified motifs, MAST was run with a given motif frequency table, the Markov background model (-bfile) and options to produce text output as a ‘hit list’ (–text –hit_list). For a search with the ATCwgAtaa motif [26], a p-value of 0.0005 was used (–mt 0.0005). MAST output and Broad gene coordinates (http://www.broadinstitute.org/annotation/genome/blastomyces_dermatitidis/MultiHome.html) were parsed using a custom perl script to find intergenic motifs <2kb upstream of predicted genes. A total of 84,965 motifs were found in the genome assembly, of which 79,458 were in intergenic regions. Of these, 3,372 copies were found <2 kb upstream of 2,468 genes. Genes with the following motifs were retained: ATC-(A/T)-GATA-(A/G), ATC-(A/T)-GATA-(T/C), ATC-(A/T)-GATT-A, ATC-(A/T)-GATC-A, ATC-A-GATG-A, ATC-C-GATA-A, and ATC-A-AATA-A. These motifs are found upstream of genes regulated by A. fumigatus SREA, an SREB homolog [26].
To discover new motifs using MEME, we identified orthologs in SLH14081 of the iron-upregulated genes from A. fumigatus (BDBG_00046, BDBG_00047, BDBG_00048, BDBG_00050, BDBG_00053, BDBG_00054, BDBG_00055, BDBG_01314, BDBG_02226, BDBG_06775, BDBG_06965, BDBG_08034, BDBG_08208, BDBG_09322) and searched the 1 kb upstream for common motifs; MEME options were set for any number of motifs per region (-mod anr), the above described Markov background model, and a minimum width of 6 (-minw 6). This identified a motif of vATCwGATAA, which is similar to the motif described by Hwang and colleagues [27].
For KOG annotation and analysis, the predicted proteome from B. dermatitidis strain SLH14081 was retrieved from the Broad institute (http://www.broadinstitute.org/annotation/genome/blastomyces_dermatitidis/MultiDownloads.html, accessed: 11/09/2009) and compared against the NCBI KOG database (ftp://ftp.ncbi.nih.gov/pub/mmdb/cdd/, accessed: 11/09/2009)) using RPSBLAST (e-value 1e-05) [59],[60]. Two data sets were generated with the first containing all B. dermatitidis genes encoding proteins that registered a KOG annotation. The second set included B. dermatitidis proteins encoded by the candidate genes with upstream GATA sites. The KOGs for both sets were correlated to their associated categories, and the total number of proteins within each category was tabulated. A two-tailed Fisher's exact test was used to determine if the number of proteins in each category were over- or under-represented when compared to all KOG-annotated proteins in the B. dermatitidis proteome. Categories were considered over-represented if the p-value of the right of the Fisher's exact test was less than 0.05 and over-represented if the left tail was less than 0.05.
To isolate and identify siderophores produced by B. dermatitidis 26199 wild-type and null mutants, we used column chromatography, liquid chromatography/mass spectroscopy (LC/MS), and reverse-phase high-pressure liquid chromatography (HPLC). Supernatants were harvested from B. dermatitidis grown in liquid HMM at 37°C under iron-poor (no added iron) and iron-replete (10 µM FeSO4) conditions when the cultures entered stationary growth (A600 = 3.5−4.0). Culture supernatants were filtered (0.2 µM), treated with 2% ferric chloride and applied to a column (K 9/30, GE Healthcare) packed with Amberlite XAD-2 resin (Supelco, Bellefonte, PA). The resin and column were prepared according to the manufacturer's recommendations. Following a water wash (7 bed volumes; flow rate of 0.2 ml/min), siderophores were eluted from the resin using methanol (1.7 bed volume; flow rate of 0.1 ml/min), reduced to dryness, and re-suspended in water (100 µl). Colorless supernatants that contained siderophores developed an orange color when treated with ferric chloride. This allowed for visual assessment of binding and elution of siderophores from the resin [61],[62].
The Mass Spectroscopy Facility at the University of Wisconsin Biotechnology Center performed LC/MS analysis of concentrated eluate collected from wild-type B. dermatitidis grown under iron-poor conditions following XAD-2 column chromatography. For HPLC, siderophores were separated on a C18 column (Agilent Eclipse XDB-C18 column; 4.6×150 mm) using a water-acetonitrile gradient containing 0.1% trifluoroacetic acid (Sigma-Aldrich). The gradient of acetonitrile was increased from 5% to 15% over 15 minutes, and 15% to 25% over 35 minutes. The flow rate was 0.5 ml/min and the absorbance was measured at 465 nm. Retention times were compared to siderophore standards (HPLC calibration kit – coprogens and fusarinines; EMC microcollections, Tubingen, Germany).
The nucleotide sequences for SREB, SIDA, AMCA, MIRB, MIRC, and HAPX from B. dermatitidis strain 26199 were obtained from the Genome Sequencing Center, Washington University, Saint Louis, MO (http://genome.wustl.edu/tools/blast). Although this genome is publically available, it is not annotated. Allelic sequences can be found at the Broad Institute (http://www.broadinstitute.org/annotation/genome/blastomyces_dermatitidis/MultiHome.html) and have the following gene locus identification numbers: SREB (BDBG_01059), SIDA (BDBG_00053), AMCA (BDBG_00128), MIRB (BDBG_05798), MIRC (BDBG_08034), HAPX (BDBG_01314). Additional gene locus numbers include: lipid transfer protein (BDBG_03618), glycosyl hydrolase (BDBG_03183), peroxisomal dehydratase (BDBG_00052), acetoacetyl-CoA synthase (BDBG_09522), and alpha-tubulin (BDBG_00020).
|
10.1371/journal.pcbi.1005862 | A free-boundary model of a motile cell explains turning behavior | To understand shapes and movements of cells undergoing lamellipodial motility, we systematically explore minimal free-boundary models of actin-myosin contractility consisting of the force-balance and myosin transport equations. The models account for isotropic contraction proportional to myosin density, viscous stresses in the actin network, and constant-strength viscous-like adhesion. The contraction generates a spatially graded centripetal actin flow, which in turn reinforces the contraction via myosin redistribution and causes retraction of the lamellipodial boundary. Actin protrusion at the boundary counters the retraction, and the balance of the protrusion and retraction shapes the lamellipodium. The model analysis shows that initiation of motility critically depends on three dimensionless parameter combinations, which represent myosin-dependent contractility, a characteristic viscosity-adhesion length, and a rate of actin protrusion. When the contractility is sufficiently strong, cells break symmetry and move steadily along either straight or circular trajectories, and the motile behavior is sensitive to conditions at the cell boundary. Scanning of a model parameter space shows that the contractile mechanism of motility supports robust cell turning in conditions where short viscosity-adhesion lengths and fast protrusion cause an accumulation of myosin in a small region at the cell rear, destabilizing the axial symmetry of a moving cell.
| To understand shapes and movements of simple motile cells, we systematically explore minimal models describing a cell as a two-dimensional actin-myosin gel with a free boundary. The models account for actin-myosin contraction balanced by viscous stresses in the actin gel and uniform adhesion. The myosin contraction causes the lamellipodial boundary to retract. Actin protrusion at the boundary counters the retraction, and the balance of protrusion and retraction shapes the cell. The models reproduce a variety of motile shapes observed experimentally. The analysis shows that the mechanical state of a cell depends on a small number of parameters. We find that when the contractility is sufficiently strong, cells break symmetry and move steadily along either straight or circular trajectory. Scanning model parameters shows that the contractile mechanism of motility supports robust cell turning behavior in conditions where deformable actin gel and fast protrusion destabilize the axial symmetry of a moving cell.
| Cell motility is a fundamental biological phenomenon that underlies many physiological processes in health and disease, including wound healing, embryogenesis, immune response, and metastatic spread of cancer cells [1], to name a few. Understanding the full complexity of cell motility, exacerbated by complex biochemical regulation, poses enormous challenges. One of them is multiple, sometimes redundant, sometimes complementary or even competing, mechanisms of motility [2]. Many researchers hold the view, which we share, that the way to face this challenge is to study all these mechanisms thoroughly, and then proceed with a more holistic approach.
One of the best studied types of motility is the lamellipodial motility on flat, hard and adhesive surfaces [3], in which broad and flat motile appendages–lamellipodia–spread around the cell body. Biochemical regulation plays an important role in the lamellipodial dynamics, but minimal mechanisms of the lamellipodial motility, such as growth and spreading of a flat actin network wrapped in plasma membrane and myosin-powered contraction of this network, are mechanical in nature [3]. While many cell types exhibit the lamellipodial motility, one model system, the fish epithelial keratocyte cell, contributed very prominently to the understanding of lamellipodial mechanics, due to its large lamellipodium, streamlined for rapid and steady locomotion [4, 5].
There are at least three distinct mechanical states of this system. The cells can be in a stationary symmetric state, with a ring-like lamellipodium around the cell body [6]. Spontaneously, even if slowly, the cells self-polarize, so that the lamellipodium retracts at the prospective rear and takes on a fan-like shape, upon which the cell starts crawling with a constant speed and steady shape [6, 7]. Often, cell’s trajectory changes from straight to circular–the cells start turning [8].
Mechanics of keratocyte movements has been studied extensively [4, 5, 7, 9]. Two principal mechanisms enable the keratocyte motility. First, polymerization of the polarized actin network at the front pushes forward the membrane at the leading edge, stretching the membrane and creating membrane tension at the sides; the membrane then snaps at the rear and pulls forward the depolymerizing actin network [10]. Second, contractile forces generated by myosin, lagging behind in a moving cell, hold the cell sides and retract the rear, allowing the front to protrude [5]. This and stick-slip dynamics of adhesions were recently shown to generate the cell self-polarization [7].
One of the fundamental questions of cell motility concerns dynamics of the cell shape: how do the actin-myosin mechanics in the cell bulk interact with actin growth and membrane mechanics at the boundary to shape, stabilize and propel the cell? This question requires mathematical insight, and in the last two decades, keratocyte mechanics were extensively modeled mathematically. The mathematical problem arising in these models is generally challenging, given that the motile cell is a free-boundary object, in which deformations of the cell shape depend on, and in turn affect, the actin-myosin movements and forces inside the cell. The history of the free-boundary cell modeling was recently reviewed in [11].
To reduce the mathematical complexity of the problem, one can ignore the mechanics in the lateral cross section and consider a simplified one-dimensional (1D) model, essentially representing the cell as a 1D strip of an actin-myosin gel. Mathematical models of this kind [12–15] provided valuable insight into conditions for symmetry breaking, motility initiation, and stabilization of the anterior-posterior length of the moving cell. Modeling the front-to-rear cell mechanics is not the only 1D approach: one can also disregard the bulk of the actin-myosin network and hypothesize that the essential dynamics is concentrated at the very edge of the cell; this allows one to approximate the cell shape by a 1D dynamic contour, which protrudes or retracts locally according to some set of rules. A number of such models [16–19] revealed that a small set of the boundary deformation rules can generate an unexpected variety of dynamic cell shapes mimicking a number of observed motile cell types. The first such model was a Graded Radial Extension mathematical model [20], which integrated experimental data and posited that actin polymerization at the lamellipodial boundary results in protrusion of the cell front and sides in the direction locally normal to the boundary, with spatially graded rate maximal at the center of the leading edge and decreasing towards the sides.
A more accurate mathematical rendering of the lamellipodia is achieved via a full 2D free-boundary model. Its general concept, first introduced in [21], is as follows. Actin-myosin contraction in the bulk of the 2D lamellipodium generates a centripetal actin flow that redistributes myosin powering the contraction; this feedback results in a spatially graded flow that tends to retract the lamellipodial boundary. Actin growth at the boundary results in protrusion countering the retraction, and the balance of protrusion and retraction shapes the lamellipodium, feeding back to the actin-myosin contraction in the bulk of the lamellipodium. The question is: what kind of cell shapes and movements does this model predict?
To address this question, 2D models of actin-myosin mechanics were employed to reproduce steady-state shapes and speed, as well as self-polarization, of a motile cell [5, 7, 22], but a number of important issues have not been adequately explored, including turning behavior and dependence of the motile behavior on the model parameters and boundary conditions. In this paper, we resolve these issues by simulating numerically a minimal free-boundary model described in the next section. We find that 1) cells are stationary when contractility is weak, 2) when contractility is strong, cells break symmetry and move steadily along either straight or circular trajectory, 3) cells exhibit turning behavior when fast protrusion destabilizes the axial symmetry of a planar myosin distribution or cell shape, and 4) motile behavior of a cell is sensitive to conditions of force balance at the cell boundary.
Mechanics plays a dominant role in keratocyte motility, while the role of biochemical regulation is less clear, and is probably of less importance [3]; thus we focus on mechanical formulation. Moreover, at least for the self-polarization phenomenon, the contractile mechanism of motility is dominant compared to the graded actin polymerization [7], and so we concentrate on the myosin generated forces and movements, and for simplicity assume that the actin growth is uniform around the cell boundary.
The model consists of force-balance and myosin transport equations,
ηΔXU+σ∇XM−ξU=0∂TM=∇X⋅(Deff∇XM−UeffM)
(1)
for the velocity of actin flow U(X,t) and myosin concentration M(X,t) defined locally for X ∈ Ω(t), where Ω(t) is a moving 2D domain representing cell geometry. Note, that all model equations are formulated in the lab frame of references. We discuss the validity of approximating the flat lamellipodium as a 2D domain in the thin-cell limit in the Supplemental Material. The model is similar to an actomyosin dynamic model suggested in [5, 23], and to active gel models of the soft matter physics [24]. In the force balance equation, the first term describes the force due to passive viscous stresses in the deforming actin network, where η is the effective actin viscosity. The form of this term corresponds to the viscous shear stress in the Stokes equation of hydrodynamics; we emphasize that the actin polymer mesh is compressible (fluid cytoplasm can be squeezed easily into the dorsal direction in the cell [23]), and so there is no incompressibility condition. In the Supplemental Material, we discuss the conditions under which the effects of hydrostatic pressure and Darcy flow are negligible in the lamellipodium. Because the movement of the cell takes place on the slow time scales, we do not consider viscoelastic effects [23]. Also, as was done in many modeling studies [25], we ignore for simplicity subtle and complex interplay between bulk and shear viscous stresses.
The second term in the force-balance equation describes the divergence of the myosin contractile stress. As in [5, 23, 24], we assume the stress to be isotropic and proportional to the myosin density, with σ denoting the force per unit of myosin density. The third term describes the effective viscous drag arising from creeping movement of F-actin relative to a substrate, mediated by dynamic adhesions and characterized by the viscous drag coefficient ξ. The linear dependence of this drag on actin velocity is a standard assumption made in cell mechanics models; for some cases, this assumption was verified experimentally.
The second of Eq (1) describes myosin transport. Kinetics of myosin can be interpreted in terms of transitions between two states, a state of free myosin diffusing in the cytoplasm and a state in which myosin is bound to the actin network [23]; clusters of the bound myosin both contract the actin network and move with it. For the transitions occurring on a fast time scale, the overall transport is well approximated by a diffusion-advection equation, and we assume this limit in our model here. Additional discussion of the myosin transport is in the Supplemental Material. For low viscosity and slow diffusion, however, using U as an advection velocity and constant diffusion coefficients D results in singular solutions, in which M and U develop Dirac-delta singularities. The effect is reminiscent of the collapsing phenomenon in a 2D version of the Keller-Segel chemotaxis model [26], which is mathematically similar to our mechanical model. The singular solutions are obviously unrealistic, given that myosin molecules have a finite size. The excluded volume effect can be taken into account by introducing effective velocities [27], Ueff = U(1−M/Mmax,u), which approach zero when M → Mmax,u But in a free-boundary problem, the actual maximum of myosin concentrations may significantly exceed Mmax,u. This is because in motile solutions, myosin accumulates at the rear of the cell, where it is also swept forward by a moving boundary; mathematically, this effect originates from the Rankine-Hugoniot boundary condition described in the next paragraph. Thus, the effect of molecular crowding on myosin velocity should generally be written as: Ueff = U(1−M/Mmax,u), if M < Mmax,u, and zero otherwise. Because diffusion is also affected by the crowding, the effective diffusivity in Eq (1) is Deff = D(1−M/Mmax,d), with a value of the cut-off Mmax,d that exceeds the actual maximal densities of myosin [28]. Overall, using a tighter myosin cut-off for advection, Mmax,u < Mmax,d, helps avoiding the singularities and numerical instabilities associated with them in a wide range of model parameters. We also explored the possibility that the anti-crowding effects come from an attenuation of the myosin stress when the myosin density is too high, described in detail in the Supplemental Material.
One of our goals is to investigate how boundary conditions, specified in a free boundary problem at a moving cell boundary ∂Ω(t), affect the model behavior. We explore two types of conditions for the force-balance equation. One of them is the zero actin velocity at the boundary, U|∂Ω(t) = 0, which assumes a very narrow band of very strong adhesions near the cell edge that rapidly adjust their positions to the instantaneous location of the boundary. Many experimental observations indicate the presence of such a band. We will term this version of the model a zero-velocity (ZV) model. Alternatively, in the absence of the sticky band, the force balance is reflected by a zero stress condition, introduced earlier in [5, 23]: n⋅(η∇XU+MI^)|∂Ω(t)=0, where I^ is the unit tensor and n is the outward normal. This assumes that the membrane tension is small relative to the contractile and viscous stresses. We will call this variant of the model a zero-stress (ZS) model. Both versions of the model share a no-flux Rankine-Hugoniot boundary condition for myosin, n ⋅(−Deff ∇XM + (Ueff – Vf)M) |∂Ω(t) = 0, where Vf is the local boundary velocity.
Kinematics of the boundary is modeled by a superposition of locally normal protrusion powered by actin growth and retraction stemming from the centripetal actin flow: Vf = Vpn + U|∂Ω(t). The approximation of the speed of normal protrusion Vp is somewhat different in the two versions of the model, as described below.
In the ZS model, Vp is defined uniformly along the boundary but depends on the cell size: Vp = V0(A0/A)−K(A−A0(A0/A)n), where A = |Ω(t)|, A0 = |Ω(0)|, and n = 2. The first term represents the rate of membrane displacement due to actin growth with a rate constant V0. The cell-size dependence of this term reflects an effective drop in actin concentration in an expanding cell, but this term alone would still produce an infinite cell expansion for large V0. Realistically, cell stretching is limited by membrane tension, which is represented in Vp by −KA term; this is consistent with previous experiments and modeling [4, 5, 29]. The term ∝(A0/A)n reflects cytoplasmic resistance to contractile forces and thus excludes collapsing of the cell in the model with small V0. Mathematically, the quadratic nonlinearity in the area dependence appears to be the lowest nonlinearity preventing the cell collapse in the ZS model. Overall, the second term in Vp, combining the effects of membrane tension and cytoplasm resistance to contraction, plays an area-preserving role (with the parameter K describing sensitivity of Vp to changes in A). Indeed, if A<A0, the membrane tension decreases, whereas actin polymerization accelerates and the resistance to further contraction rapidly grows. On the other hand, if A>A0, the membrane tension increases rapidly stopping the actin growth.
For the ZV model, where U|∂Ω(t) = 0 and Vf = Vpn, there must be a nontrivial variation of Vp along the boundary, since for a uniform Vp, the cell centroid is always stationary. Based on experimental observations and models showing that myosin can impede protrusion by bundling actin filaments at the boundary [18], we hypothesized that the actin growth rate is a decreasing function of local myosin density. Correspondingly, we use the following expression for Vp in the ZV model, Vp = V0(A0/A)(1+M)|∂Ω(t)/M0)−1−K(A−A0), where M0 is a threshold beyond which myosin inhibits actin growth almost entirely. The expression has essentially the same dependence on cell area as in the ZS model, except that for the ZV model, n = 0 proved to be sufficient for preserving the target area. We discuss derivation of the mathematical expression for Vp from the force balance at the lamellipodial boundary and provide additional explanations in the Supplemental Material.
To nondimensionalize the model, we use the following set of units. The length unit L is defined as a characteristic linear size of the cell with a target area, L=A0/π (e.g., if this cell is a circle, L is its radius). We further choose L2D−1 and M0 as the units of time and myosin concentration, respectively. Then the dimensionless variables, differential operator, and current and target cell areas are, respectively, t = TDL−2, x = XL−1, u = ULD−1, m = M/M0, ∇ ≡ ∇x = L∇X, |ω|=|Ω|L−2 and a0 = A0L−2. Correspondingly, Eqs (1) takes the form,
∂tm=∇⋅(deff∇m−ueffm)αΔu+β∇m−u=0,
(2)
where ueff = u(1−m/mmax,u), if m<mmax,u, and zero otherwise, and deff = 1−m/mmax,d. Eqs (2) include two dimensionless parameters: the dimensionless viscosity-adhesion length parameter α = ηL−2ξ−1 and the myosin contractility constant β = σM0(Dξ)−1. Note that the mechanical effect of localized myosin contraction spreads on the length scale η/ξ, so the viscosity-adhesion length parameter α is the ratio of the length scale of the mechanical action to the cell size. The first of Eqs (2), a diffusion-advection equation for myosin, is subject to the mass-conserving zero-flux boundary condition, n ⋅(−deff∇m+(ueff−vf)m)|∂ω(t) = 0, yielding an additional dimensionless parameter μtot=∬ω(t)m⋅d|ω|. The dimensionless boundary conditions for the force-balance equation (the second of Eqs (2)) in the ZV and ZS models are u|∂ω(t) = 0 and n⋅(α∇u+βmI^)|∂ω(t)=0, respectively.
The dimensionless boundary velocity equation is vf = vpn+u|∂ω(t). In this equation, the dimensionless rate of membrane displacement caused by the actin polymerization and area preserving factors is vp = v0a0/a−k(a−a0(a0/a)n), where v0 = V0L/D and a = |ω(t)|. For the ZS model, n = 2, whereas for the ZV model, n = 0 and there is the additional dependence on m in the first term: vp = v0a0/(a(1+m|∂ω(t)))−k(a−a0).
Note that varying parameter k = KL3D−1 is equivalent to rescaling the actin polymerization constant v0. Also, because the myosin contractility constant β enters Eq (2) in combination with m, varying β is similar to rescaling μtot; in fact, β could be formally excluded from the ZS model by employing a different concentration unit, and the same is true for the ZV model defined in a fixed geometry, see section Cell becomes motile when myosin contractility is higher than critical. We therefore focus in our study on the role played by three essentially independent model parameters: α, μtot and v0.
Steady dynamics of a motile cell were explored by solving Eqs (2) in domains with free boundaries. Note that even though the force-balance equation does not involve time derivatives in and of itself, the coupled system (2) constitutes an initial-value problem and one must specify initial conditions for both variables and the domain, m(x,0), u(x,0), and ω(0).
To elucidate processes leading to instability of an initially symmetric stationary state of a motile cell and its transitioning to motility, we used initial conditions based on a stationary steady state of the ZV model in a circular geometry ω(0) = {(x,y):x2+y2≤1} (such that a0 = |ω(0)| = π): u(x,0) = 0, m(x,0) = (μtot/|ω|)(1−gx). Note the linear horizontal gradient, added to a steady-state uniform myosin distribution to probe stability of a stationary state; the gradient steepness g was assigned values from (0,1]. Note also that given the symmetry of ω(0), the definition of m(x,0) ensures that the solution has a prescribed μtot. The initial conditions specified above were used in solving both ZV and ZS models throughout this study.
Numerical solutions of the ZV and ZS models were obtained using a generalized version of a mass-conservative algorithm originally developed for solving parabolic equations in moving domains with known kinematics [30]. The method has been shown to converge in space with an order close to 2 in L2-norm and ensures exact local mass conservation. The latter is achieved by employing finite-volume spatial discretization [31] and natural neighbor interpolation [32]. The algorithm was developed for modeling cell motility in Virtual Cell (VCell), a general-purpose computational framework for modeling cellular phenomena in realistic geometries [33].
To be applicable to a free-boundary problem with the models described above, the original method had to be augmented in several aspects. First, the boundary kinematics is generally not known a priori but rather needs to be computed on the basis of the rates that are functions of state variables—the actin velocities, in the ZS model, and the myosin concentrations, in the ZV model. To approximate values of the variables at the points of the boundary where the boundary velocities need to be evaluated, we used the second-order bilinear extrapolation. Once the boundary velocities are obtained, the cell boundary is advanced using a robust front-tracking technique implemented in FronTier, a freely available C++ library for tracking interfaces in two and three dimensions [34]. Accuracy of the algorithm coupled to FronTier was evaluated using several benchmark examples, one of which was based on the models of this study. The tests have shown that the accuracy of the original algorithm is preserved, if in addition to the second-order extrapolation, the front-tracking routine is also at least second-order accurate.
Second, the system (2), consisting of the coupled parabolic and elliptic equations, is nonlinear. Indeed, the equations are coupled via the advection term of the parabolic equation and myosin-dependent stress term in the elliptic equation, as well as through the boundary conditions at the moving boundary; also, the effective transport parameters are functions of the myosin concentration. To solve the system, we implemented a segregated solution strategy [31], in which equations are solved one at a time and nonlinear terms are treated by fixed-point iterations. One advantage of the segregated solver is that it prevents the matrix of a linearized system from becoming very large even with very fine computational grids. The system was advanced in time using an implicit backward Euler time discretization.
For each time step, the segregated method performs fixed-point iterations in two steps. First, we solve for actin velocities using fixed myosin concentrations from the previous iteration. The obtained velocities are then used as a fixed advection field at the next step, where we solve the linearized transport equation for myosin concentrations. Note that at this step, the values of the myosin concentrations in the discretized time derivative correspond to the previous time step, not to the previous fixed-point iteration. At the end of the iteration, maximum absolute differences of two consecutive iterates are checked for convergence. If they are within prescribed tolerances, the iterations stop and the solver reports the velocities and myosin concentrations as the current time step values, otherwise it proceeds to the next iteration and continues until the iterations converge or an imposed maximum for the number of iterations is exceeded. The algorithm is illustrated below for one time step by the mathematical pseudocode, where mk and uk are the variable values at the kth time step, mnk+1 and unk+1 are the nth iterates for the (k+1)th time step, MaxNumIters is the maximum allowed number of iterations, and ‖⋅‖∞ denotes the L∞-norm.
set m1k+1=mk and u1k+1=uk
for n = 1: MaxNumIters
- solve αΔun+1k+1+β∇mnk+1−un+1k+1=0 to get un+1k+1
- evaluate ueff and deff using un+1k+1 and mnk+1
- solve (mn+1k+1−mk)/Δt=∇⋅(deff∇mn+1k+1−ueffmn+1k+1) to get mn+1k+1
- calculate absolute errors ‖un+1k+1−unk+1‖∞ and ‖mn+1k+1−mnk+1‖∞
- if solution converged, break the loop, else mnk+1=mn+1k+1, unk+1=un+1k+1
end of segregated loop
if n< MaxNumIters mk+1=mn+1k+1, uk+1=un+1k+1, else iterations are stagnant.
The segregated solver was validated against a coupled nonlinear solver implemented in COMSOL Multiphysics [35]. Good agreement was observed, with relative differences below 0.3%.
The computations were performed with the following solution parameters: the mesh sizes h varied between 0.05 and 0.16, whereas the time step was Δt = ch with c varying from 0.0002 to 0.025 (fast-moving cells required smaller mesh sizes and time steps), the tolerance for the differences of consecutive iterates was 1E-10, and the maximum allowed number of iterations, set at 35, was never reached.
The ZV and ZS models were used to simulate transitions of a motile cell from stationary to motile state. For this, as described in section Initial conditions, an initially radially symmetric cell was perturbed by superimposing a linear gradient over a uniform distribution of myosin. The emerged steady states fall into three asymptotically stable mechanical modes. For some parameter values, the cell, after a finite displacement, comes to a stop, with a final radially symmetric shape and myosin distribution, indicating the stability of the stationary state (Fig 1A). For other parameters, the cell irreversibly breaks symmetry, both in terms of its shape, distribution of myosin, and actin velocity field, and either acquires unidirectional motility (Fig 1B) or locks in a rotational mode (Fig 1C).
To analyze conditions for transitioning to different types of motility, we scanned the actin growth constant (v0) and the contractility parameter (μtot) for two values of viscosity-adhesion length parameter, α = 0.5 and α = 1. The values of other model parameters, β = 5, a0 = π, k = 1.5, mmax,u = 15 and mmax,d = 125, were fixed in all simulations; the choice of these values ensures that the corresponding section of parameter space is representative of various states. The results of parameter scanning are presented in Fig 2 showing cell mechanical states as functions of the model parameters.
It should be noted that distinguishing between translational and rotational modes is sometimes ambiguous, particularly for the states near the borders between the corresponding regions in the parameters space. For example, some states of the ZS model shown in the diagrams of Fig 2 as translating were in fact only ‘piecewise unidirectional’, as the cell in those states would on occasion change its direction. Moreover, in some states in the ZS model, identified as translations, also neighboring the rotations in the diagrams of Fig 2, the cell actually changes its direction but very gradually, so the state could be a rotation with a very long radius. As a practical rule, we labeled states as rotations only if the radius of rotation of the cell centroid was comparable to, or less than, the linear size of the cell.
Below we discuss in detail the conditions required for the straight and rotational motility in our models and the underlying mechanisms.
The results in Fig 2 show that cells break symmetry and transition to motility when parameter μtot exceeds a threshold. The threshold behavior originates from a positive feedback between the actin flow and myosin gradients: the contractile forces, generating the centripetal flow of myosin, are proportional to the myosin gradients, which, in turn, are reinforced by the advection of myosin. This positive feedback results in steep myosin gradients and, potentially, symmetry breaking, but below a critical value of μtot, these gradients are prevented by dissipative viscous forces and myosin diffusion, and the cell remains stationary and radially symmetric. Above the critical value of μtot, steep gradients of myosin are developed and the radially symmetric stationary state becomes unstable. While kinematics of a free boundary plays a key role in the symmetry breaking and transitioning to motility in both models, the loss of symmetry in the ZV model also occurs in fixed domains. In the next section, we discuss effects of the boundary conditions for actin flow on solutions in domains with fixed and free boundaries.
Linear stability analysis can be used to estimate critical values of μtot in a simplified ZV model in a fixed domain. Consider a 1D ZV model on the fixed-length segment ∂ω = {x∈(0,1)} in the limit mmax,1, mmax,2 → ∞, so that deff = 1 and ueff = u(x,t). Then, Eqs (2) become αuxx + βmx−u = 0, mt = mxx−(um)x. In this model, varying the contractility constant β is equivalent to rescaling μtot. Indeed, β could be excluded altogether by renormalizing m: m˜=βm, but in what follows parameter β is kept for generality. The symmetric steady state is characterized by uniform myosin distribution and absence of actin flow, u = 0, m = μtot. Its stability is probed by imposing small perturbations, δu(x,t) = δu0 exp(λt+iqx), δm(x,t) = δm0 exp(λt+iqx) with 0<δu0<<1, 0<δm0<<1 and q = π, 2π,…, so that u = δu and m = μtot+ δm. The perturbations satisfy the linearized system of differential equations, αδuxx+βδmx−δu = 0, δmt = δmxx−μtot(δu)x, and the corresponding system of linear algebraic equations, −(1+αq2)δu0+iqβδm0 = 0, iqμtotδu0+(λ+q2)δm0 = 0, yields nontrivial solutions for δu0 and δm0, if λ(q) = q2(βμtot(1+αq2)−1−1). The perturbations grow if λ(q)>0, with the fastest growing mode having the minimal wave number, qmin = π, and thus involving a large-scale redistribution of myosin. Thus, the symmetric state becomes unstable if βμtot>1+π2α or, in the dimensional form, σM0πL2>D(ξL2+π2η).
The instability criterion predicts that the critical value of βμtot is an increasing function of α. The results of Fig 2 indicate that this prediction, while obtained by analyzing a ZV model in a fixed domain, holds for the free-boundary models as well. Indeed, the competition between the myosin contractile stress and dissipative processes, mathematically expressed in the instability condition, drives the initiation of motility in the free-boundary models. As described at the beginning of this section, the transition to motility occurs when the contractility, reinforced by the model positive feedback, prevails over the dissipation. For α≥1, the dimensional form of the instability criterion reduces to σM0L2>πDη: the total myosin stress needs to overcome the smoothing effects of actin viscosity and myosin diffusion, while the adhesion strength does not matter. In the limit of large values of α, the actin network is effectively stiff and thus does not allow for significant actin flows, which makes the cell more symmetric and as a consequence less motile and slower. Our simulations confirm this prediction (Fig 3A and 3B). In the opposite limit of highly deformable actin network, α<<1, the instability criterion reduces to σM0>ξD, so the cell becomes motile if the characteristic myosin stress is able to generate actin flow that overcomes myosin diffusion, which requires the weakening of adhesions and strengthening of myosin, in agreement with the experiment [7].
Finally, it is worth noting that whereas the motility threshold in the ZS model is independent of v0, the critical values of μtot in the ZV model, where the actin growth is affected by myosin, vary with v0 (Fig 2). Indeed, the cell described by the ZV model with v0 = 0 would not transition to motility for any μtot, because in this case, the myosin influence on the boundary is lost. Therefore, in this version of the model motility initiates only for finite values of v0, which drop with the increase of μtot. In the ZS model, the cell with a sufficiently high μtot initiates motility even as v0→0, because the asymmetric myosin pulls the boundary inward asymmetrically, and the area-preserving term causes the effective protrusion.
Similar to the 1D ZV model analyzed in the previous section, the symmetric state of the 2D version of the ZV model becomes unstable for sufficiently high μtot even in a fixed geometry, with myosin relocating to the cell boundary. Fig 4A illustrates the instability of the radially symmetric steady state, in which the maximum of myosin was slightly shifted to the left of the cell center. Qualitatively, because of the zero actin velocities at the boundary, a second, initially small, local maximum of myosin appears at the boundary point closest the main maximum due to slightly faster diffusion. The competition between the two maxima lowers the myosin gradients on the left side of the main one, resulting in a net force acting on it in the left direction. Hence, the relocation of myosin to the left segment of the boundary. In the cell with a free boundary, the redistribution of myosin is conferred to boundary velocity, resulting in slower outward and eventually inward movements of the part of the boundary that becomes the cell rear. The cell movement further skews the myosin towards the rear. For the small to moderate rate of actin growth and cell speeds, the cell maintains a convex shape and a steady unidirectional motion, with myosin forming a wide band at the rear edge (Fig 1B and S1 Movie).
In the ZS model of a fixed symmetric cell, an inward actin flow at the membrane prevents myosin from accumulating there. As a result, the symmetric solution with a myosin maximum at the center remains stable even for μtot above the threshold. Indeed, shifting the maximum from the center in this case increases the myosin gradients and centripetal forces on the ‘shorter’ side and decreases them on the ‘longer’ side, netting a stabilizing force. In the free-boundary problem, however, the symmetric solution, stable at low μtot (Fig 1A), becomes unstable above the motility threshold. Fig 4B and S2 Movie illustrate a transition to unidirectional motility that occurs in the ZS model with sufficiently large μtot and small to moderate v0. As the myosin cluster shifts slightly from the cell center, the closer side is pulled inward faster and becomes the prospective cell rear, while the opposite side, where the protrusions are faster than retractions, becomes the cell front. In the ensued motility, myosin is pressed against the rear and in turn exerts a stronger inward force on the proximal portion of the boundary, developing a local concavity. If the movement is sufficiently fast, the myosin spreads along the portion of the boundary with negative curvature. This positive feedback between the myosin asymmetry, actin flow, and cell movement is the key to the stable motility.
Our simulations show that the aspect ratio of a steadily moving cell varies between 1 and 3 (Fig 3C), in agreement with experimental observations [5]. Note that in contrast to the ZV model, where the cell aspect ratio grows moderately with v0, it becomes essentially independent of model parameters in the ZS model with μtot/π>1. This can be qualitatively understood by noting that myosin, which in a moving cell accumulates in the middle of the rear, exerts comparable forces on the front and side portions of the boundary. Then, because the myosin-generated flow decreases with distance at similar rates in all directions, the distances from the rear to the front and the sides should be on the same order, yielding the average aspect ratio ~ 2.
The ZS model generally predicts significantly higher cell speeds compared to those in the ZV model (Fig 5A). This is because in the ZS model, the fast centripetal flows generated by myosin at the rear boundary tend to decrease the cell area, leading to fast effective protrusion at the front, as actin can grow rapidly against the lowered membrane tension. As a result, the cell speed increases but the cell area decreases with total myosin (Fig 5B). Interestingly, the cell speed in the ZS model decreases slightly with the actin growth rate v0, which can be understood by noting that the cell area in this model increases with v0, thus mitigating the effect of myosin. In the ZV model, the cell speed is virtually insensitive to μtot, for μtot>π, because the term with v0 in the expression for vp becomes inessential for m∼μtot/π>1. For the same reason, the cell area is also insensitive to μtot.
Importantly, our model predicts that there are no short-wavelength instabilities in the cell shape (like fingering instabilities characteristic for some physical free-boundary models), which is supported by the experiment: there are small fluctuations on the experimentally observed cell boundaries, but they mostly do not grow.
The most nontrivial and important property of our models is that they predict rotational states with a radius of rotation comparable to the cell size in large regions of their parameter space (Fig 2). Note that both the radius of rotation and the angular velocity are not particularly sensitive to parameter values (Fig 3D and 3E).
Emergence of cell turning in the models can be qualitatively understood by analyzing the loss of stability of a planar axial symmetry characteristic of the straight moving cell. In the ZV model, the steady rotations are observed for v0 exceeding a threshold that is largely insensitive to either α or μtot. Fig 6 and S3 Movie illustrate, for a particular parameter set, how rotations come about in the ZV model during a transient movement following a ‘nudge’ applied to a stationary cell in the form of an initial horizontal gradient of myosin. The initial convex cell shape is favorable for maintaining a unipolar axially symmetric myosin distribution, and the resulting motility is unidirectional. But due to sufficiently large v0, fast boundary velocities at cell’s sides tend to elongate the cell in y-direction, making it prone to developing a concavity at the cell rear. In such a shape, the myosin spreads along the rear part of the boundary more uniformly (Fig 6B, t = 5), a distribution that is no longer stable. Indeed, even a slight asymmetry in the distribution of myosin, reinforced by the positive feedback from actin velocities and myosin accumulation due to faster movement of the corresponding portion of the rear boundary, breaks the axial symmetry (Fig 6B, t = 15). As a result, a stable asymmetric cell shape emerges as the cell locks in rotations (Fig 6A), with myosin aggregated at a high-curvature portion of the boundary.
While the same mechanisms underlie the turning behavior in the ZS model, the two models yield significantly different results for the parameter regions of rotations. This is due to the differences in boundary conditions that reflect the opposing assumptions about the strength of adhesions at the cell periphery, and in ways of conferring the myosin dynamics onto kinematics of the boundary. Unlike the ZV model, the v0 threshold for rotations in the ZS model strongly depends on μtot and α (Fig 2). In particular, the rotational states may exist for any v0, if α is sufficiently small. Note also, that the concavity of the cell shape does not always destabilize unidirectional motility in the ZS model (Fig 4B).
Fig 7 and S4 Movie illustrate the onset of turning in ZS model. If the contractility due to myosin is strong, myosin forms a radially symmetric aggregate, which in a translating cell is skewed to the cell rear, pulling the rear boundary inward and maintaining the cell propulsion. When the myosin aggregate is sufficiently close to the rear boundary, it pulls the center of the cell rear inward stronger than the sides of the rear edge, creating a ‘dip’ at the center of the rear edge and giving the cell a characteristic keratocyte fan-like shape, in which the sides of the cell lag behind the center. For the parameters in the upper left corner of the parameter space (Fig 2), the cell motion is fast, and in the frame of the cell, myosin is effectively swept towards the rear and ‘pressed’ against the rear boundary; in these conditions, the translational motility remains stable (Fig 4B). For intermediate values of μtot, the cell moves slower and the myosin aggregate maintains its radial symmetry and remains close to the cell centroid (Fig 7A, t = 7). In this position, myosin is able to pull inward not only the rear but also the front of the cell, making the axial symmetry of the system unstable. Indeed, even a slight random asymmetry in either the myosin distribution or the cell shape induces and reinforces the asymmetry of the other. If, for example, the myosin aggregate becomes slightly closer to one side, this side is pulled inward faster than the other, which brings even more myosin to the side that is pulled inward, because the shift of that side effectively sweeps myosin towards it (Fig 7A, t = 9).
Once the axial symmetry of cell shape and the position of the myosin aggregate is broken (Fig 7A, t = 23.5), the boundary velocity field becomes asymmetric as well (Fig 7B). It is then clear that a steady movement of a cell with an asymmetric shape and asymmetric boundary velocities (where faster displacements occur at the location of higher myosin gradients) must involve rotations. Indeed, by connecting consecutively the ends of the arrows representing normal displacements of points of the boundary in Fig 7B, one recovers the same contour in a rotated position, as the centroids of the cell in different positions belong to the same circle (Fig 7B). We also note that the shape of the expanding portion of the cell boundary is reminiscent of spirals described by more abstract models of rotating free boundaries [36].
In this paper we systematically explored the ability of a minimal actin-myosin contractility model [7, 14] to reproduce observed mechanical states of the simplest motile cell. The model analysis has shown that the mechanical state of the cell critically depends on just three dimensionless parameters representing the myosin contractility, characteristic viscosity-adhesion length, and actin growth. For the large viscosity-adhesion length, the actin network becomes effectively stiff and does not allow for significant actin flows, which makes the cell more symmetric and as a consequence less motile and slower, and in the limit of very large values of viscosity-adhesion lengths, the cell is stationary. In the opposite limit of short viscosity-adhesion lengths, myosin forms a very small high-density aggregate, which affects the actin network only locally. In this regime, the steady motility is impossible, and the cell starts to pivot. Thus, an important conclusion is that to move straight and steadily, the cell has to keep the viscosity-adhesion length on the order of unity (to adjust the ratio of actin viscosity to adhesion strength so that it is on the order of the cell area). Interestingly, this conclusion is consistent with estimates based on the experimental data for keratocytes [5].
Intuitively, if myosin contractility is weak, the myosin spreads uniformly and the cell remains stationary and symmetric. Above a contractility threshold, the cells become motile. The mode of motility depends on the boundary conditions. For the zero actin velocity at the boundary and the sufficiently small actin growth constant and cell speed, the convex-shaped cell maintains unidirectional motility, with myosin concentrated in a band at the rear edge. With the rate of actin growth above a certain threshold, the increase of the cell speed is sufficient for the cell to lose its planar axial symmetry and start rotating. With the zero-stress boundary conditions, rotations occur for intermediate contractility strengths, whereas in the high contractility range, the fast cells stabilize their unidirectional movement, as myosin being effectively compressed into a long band at the rear edge. We found that both explored boundary conditions explain general features of the keratocyte motility, but there are interesting differences in the predicted behaviors, as discussed above.
The main finding of our study is that the contractile mechanism of motility results in a very robust turning behavior of the cell: in the models with both explored boundary conditions, the cell moving along a circular trajectory is not an anomaly but rather a solution that exists in a large region of the model parameter space. Broadly speaking, the cell starts turning in conditions of breaking the planar axial symmetry of its myosin distribution; in the ZV model the transition to rotation is controlled by the rate of actin growth, whereas in the ZS model–by all three independent model parameters. Turning motile behavior is an important part of the cell mechanical response in chemotaxis [37] and galvanotaxis [38], and our model generates intuition about the turning mechanism.
One important test of our model is that the solutions exhibit a characteristic fan-like keratocyte shape, with the side-to-side distance greater than the front-to-rear distance and aspect ratio between 1 and 3, in excellent agreement with the observations [5]. Moreover, the predicted aspect ratio in the ZS version is nontrivial and biphasic, reaching a maximum at intermediate myosin contractility and decreasing at very weak or strong contractility, indeed observed in [5]. Similarly, the model predicts that the lamellipodial area increases at higher adhesion and lower myosin contractility, and that the cell speed increases with myosin contractility, as observed [5]. Lastly, in agreement with the experiment [7], higher myosin contractility and/or lower adhesion strength are predicted to promote the cell polarization and motility initiation. One significant prediction of our model is that both self-polarization of the cell and its turning behavior can, in principle, occur, without complex adhesion dynamics. While it was observed that a nonlinear stick-slip adhesion behavior accompanied cell polarization [7], it remains an open question whether this nonlinear behavior is essential.
The model predicts that the stable motile behavior of the cell requires tight regulation of the total lamellipodial area. We hypothesized that this regulation is mechanical, through the membrane tension. Indeed, perturbations of the total membrane area and membrane tension were found to change the lamellipodial area in a predictable way, and drastic perturbations destabilized the cell [29]. We find that cell polarization may not depend on the cell ability to move: in the ZV model, myosin distribution and actin flow become asymmetric even in the stationary symmetric domain. However, for the motility initiation, protrusion of the boundary is obviously essential (note that while motility in the ZS free-boundary model with sufficiently high μtot can be initiated even with v0→0, the area-preserving term of Vp in this limit effectively induces protrusion of the front, which is less affected by myosin, see sections Model and Cell becomes motile when myosin contractility is higher than critical).
Keratocyte motility and especially the peculiar and steady shape of the moving cell inspired a great deal of free-boundary modeling in the past decade. Our model is based on the well-justified assumption that the mechanical force balance determines cell shape and movements. Conceptually, our model is similar to the active gel models [24, 39], originating in the soft-matter physics. Some models were based on the viable idea that certain self-organized chemical patterns are upstream from the actin-myosin machinery [22, 40, 41], but majority of studies explored mechanical models [42–44]. A variety of numerical techniques–Potts models [40, 45], phase-field method [42, 44], immersed boundary method [43]–were used in respective simulations. Keratocyte polarization was modeled in [7, 46]. Alternative turning mechanisms, very different from the one predicted by our model, were computationally explored in [47, 48]. The fact that the majority of the models reproduce the keratocyte shapes and motile behavior corroborates the existing biological intuition about the keratocyte lamellipodium as the most basic, streamlined and robust actin-myosin motile structure [49]. Each of the cited studies added invaluable insights to understanding multifaceted aspects of cell motility; a relative advantage of our model is in that it is most easily connected to the experimentally observed biophysics of force balance and myosin transport in keratocytes [5, 7].
The minimal model we explored already predicts a wealth of motile behaviors (Fig 8). It is known, however, that even the cell as streamlined for locomotion as keratocyte has complexities that far exceed our minimal model. The two main aspects that need to be added to the model to make it more realistic are: spatially graded actin polymerization independent of myosin [10] and dynamic nonlinear adhesions. Complex effects of dynamic and non-homogeneous adhesions already attracted special attention and were simulated in [7, 46, 50]. It will also be interesting to explore how the predicted cell dynamics depend on actin density [51], more complex constitutive relations for the actin-myosin stress [52], membrane curvature [53, 54], elastic [55] and anisotropic [49] effects in the actin network. Our minimal free-boundary model might be useful for future modeling of other modes of cell motility [56] and collective cell movements [57, 58]. Lastly, for decades, research focused on understanding cell movements on flat 2D surfaces, and only recently exploration of cell crawling through three-dimensional (3D) matrices, more physiologically relevant, has begun experimentally [59] and theoretically [60–62]. Extension of our model to 3D will be a challenging, yet necessary, effort.
|
10.1371/journal.pgen.1001136 | Germline Variation Controls the Architecture of Somatic Alterations in Tumors | Studies have suggested that somatic events in tumors can depend on an individual's constitutional genotype. We used squamous cell carcinomas (SCC) of the skin, which arise in high multiplicity in organ transplant recipients, as a model to compare the pattern of somatic alterations within and across individuals. Specifically, we performed array comparative genomic hybridization on 104 tumors from 25 unrelated individuals who each had three or more independently arisen SCCs and compared the profiles occurring within patients to profiles of tumors across a larger set of 135 patients. In general, chromosomal aberrations in SCCs were more similar within than across individuals (two-sided exact-test p-value ), consistent with the notion that the genetic background was affecting the pattern of somatic changes. To further test this possibility, we performed allele-specific imbalance studies using microsatellite markers mapping to 14 frequently aberrant regions of multiple independent tumors from 65 patients. We identified nine loci which show evidence of preferential allelic imbalance. One of these loci, 8q24, corresponded to a region in which multiple single nucleotide polymorphisms have been associated with increased cancer risk in genome-wide association studies (GWAS). We tested three implicated variants and identified one, rs13281615, with evidence of allele-specific imbalance (p-value = 0.012). The finding of an independently identified cancer susceptibility allele with allele-specific imbalance in a genomic region affected by recurrent DNA copy number changes suggest that it may also harbor risk alleles for SCC. Together these data provide strong evidence that the genetic background is a key driver of somatic events in cancer, opening an opportunity to expand this approach to identify cancer risk alleles.
| Tumors exhibit DNA copy number gains and losses, many of which alter the dosage of genes that promote or suppress tumorigenesis. Evidence from familial cancer syndromes and animal models have shown that DNA copy number changes acquired somatically during tumor progression can be controlled by the constitutional genotype. The genetic heterogeneity among humans makes it difficult to systematically assess the extent of this effect. We used a unique clinical scenario of squamous cell carcinoma (SCC), which can arise in high multiplicity within patients, to compare the pattern of somatic alterations on a homogeneous genetic background. We examined the genome-wide pattern of DNA copy number changes of tumors from individuals who had three or more independent SCCs. We identified multiple chromosomal regions that showed higher frequency of change in SCCs within patients than across patients, suggesting that the genetic background of the individual is important in driving these changes. We further confirmed this by identifying eight regions with strong evidence for a selection of loss or gain of a particular allele within patients. Together these data demonstrate that the genetic background of an individual influences the pattern of somatic alterations in tumors, offering a novel approach to map susceptibility alleles.
| Human solid cancers are characterized by the presence of numerous genetic alterations that accumulate during the evolution of the disease. While the mutation spectrum within biologically related cancer subtypes often shows similarities with regards to the patterns of genetic alteration, each individual cancer has a unique combination of alterations. The forces that shape the genomic landscape of individual cancers are in part determined by the nature of the initiating oncogenic alterations and the sequence in which they occur. However, the constitutional genotype of the cell acquiring the first pathogenetically relevant mutation is likely to play a role in influencing which somatic alterations will undergo positive or negative selection. The influence of inherited alterations on the pattern of somatic mutations found in evolved cancers has been demonstrated in several cancer types. In breast cancers from individuals with inherited BRCA1 mutations one finds more frequent losses on 4p, 4q, 5q, Xp and Xq and gains of 10p and 16q compared to breast tumors from individuals without BRCA1 mutations [1], [2]. In melanoma, patients with germline variations in MC1R have a higher frequency of somatic BRAF mutations in their melanomas than patients without MC1R variants [3], [4]. These examples of interactions between predisposing germline alterations and acquired mutations in the tumor occur between different genes (trans-effects). Several studies have also identified cis-effects, in which somatic alterations affect specific inherited variants. Examples include two genes identified through mouse mapping studies: AURKA, which shows allele-specific gains of the T91A allele in human colon tumors [5], [6] and PTPRJ, which shows allele-specific losses of the A1176C allele in human colon tumors [7]. In addition, rs6983267, a SNP on 8q24 found through several genome-wide association studies to be associated with susceptibility to colorectal cancer, shows allele-specific imbalance [8].
Together these data suggest that inherited variation as well as somatic mutations arising early in progression help shape the pattern of somatic changes that occur subsequently during tumor evolution. One way to more systematically assess the effect of the constitutional genotype on the pattern of somatic alterations is to compare cancers of the same type that arose independently on the common genetic background of a single individual. Only a few cancer types arise frequently enough to render such analysis practical. Basal cell carcinoma and squamous cell carcinoma (SCC) of the skin often develop in multiplicity. Furthermore, the incidence of SCC in particular is dramatically increased in immunosuppressed patients. Specifically, in organ transplant recipients (OTRs) the risk of SCC is 65 to 250 fold increased compared to the general population [9]. As a consequence, some patients develop dozens of separate primary carcinomas. In this study, we exploited the unique property of the OTR population to test the hypothesis that tumors arising on a common genetic background will have somatic alterations that are more similar to each other than to those found in similar tumors that developed in different individuals and whether this scenario can be exploited to discover predisposing genetic factors.
We obtained copy number profiles as measured by array comparative genomic hybridization (aCGH) from tumors arising in individuals with multiple independent cutaneous squamous cell carcinomas (SCCs) or keratocanthomas (KAs) (intra-group) and copy number profiles of SCCs and KAs of separate individuals (inter-group). 305 independent tumor samples from 181 patients were included in this initial study (Figure 1). As previously reported [10], [11], focal genomic aberrations were rare in these tumors and DNA copy number aberrations consisted mostly of the loss or gain of whole chromosome arms. As the resolution of copy number changes using aCGH is around 1Mb, it is possible that we missed focal amplifications or deletions in this study.
We compared aCGH profiles between the three types of skin tumors in our study, SCC, SCC in situ (Bowen's Disease) and keratocanthoma. There were no statistically significant differences in frequency of clone loss or gain between the SCC and keratocanthoma profiles; however there were several loci which showed differences between the SCC in situ profiles and profiles from the other two tumor types (data not shown). Because of this, we focused our comparative analysis on SCCs only. Our data set included 222 SCCs from 135 individuals. From 25 of those individuals, three or more SCCs (median = 4.2; range 3–6) were analyzed to compare the intra-group and inter-group similarities of DNA copy number changes.
We found a significantly higher concordance of chromosomal aberrations in SCCs within than between patients [two-sided T-test p-values: 6.97×10−8](Figure 2). Interestingly, certain chromosomal regions (4q, 11q, and 17q) were preferentially affected by this concordance (individual arm p-values<0.05; Table 1). The intra-group correlation coefficients (ICC) for the array elements of these regions were compared but did not allow narrowing the genomic region to specific loci within these regions. This is not unexpected, considering that most of the tumors showed copy number changes affecting large genomic regions, such as entire chromosomal arms or chromosomes.
We rationalized that any inherited variants that promote cancer in an allele-specific manner would result in allele-specific DNA copy number changes reflected by preferential loss or gain of one specific chromosome in the tumors of an individual patient. By contrast, dosage events affecting genes that promote cancer in allele-independent manner, e.g. loss of CDKN2A or gain of MYC, were expected to display random somatic alterations of either allele [12]. To determine the presence of allele-specific changes occurring within tumors of individual patients, we performed loss of heterozygosity analyses of 45 microsatellite markers covering 14 chromosomal regions that were chosen based on the frequency of aberration as measured by aCGH and without prior knowledge of regions showing more similarity within versus across patients. 270 tumors from 65 individuals with a minimum of three independent tumors were included in this analysis. The constitutional genotype was determined from DNA extracted from blood leukocytes of each patient. Allelic imbalance was defined as a tumor to normal DNA allelic ratio of greater than 1.5 or less than 0.67. Statistical analyses for preferential imbalance were conducted for individuals who were heterozygous for a given marker and had two or more tumors showing imbalance; two examples are illustrated in Figure 3. Thirteen markers representing eight different genomic regions showed significant skewing towards one allele as determined by a Bayesian/frequentist approach (Text S1). Markers demonstrating significant preferential allelic imbalance mapped to chromosomal locations 3p24, 3q21-26, 5q23, 7p12-21, 7q31, 8q24, 9p21, 11q24, and 18q22 (Table 2, Figure 3). These data indicate that the increased similarity of copy number changes within individuals is at least in part due to inherited variation within the same region as the copy number change.
The next question we addressed was whether variations in any known tumor susceptibility genes were driving allele-specific imbalance at the loci identified through our studies. Several genome-wide association studies (GWAS) have been performed for multiple cancers including breast, prostate, colon, and melanoma [13]–[15]. Variants at 8q24 identified via GWAS have been associated for cancer risk for multiple cancer types [8], [13], [16]–[21]. To determine if any of these were candidates for the observed allele-specific imbalances at 8q24, we tested three variants, rs13281615, rs1447295, and rs6983267, for allele-specific imbalances in matched normal and tumor DNAs from individuals with SCC. Of these, only rs13281615 showed statistically significant evidence of allelic skewing (Table 3). Of 35 heterozygous tumors showing imbalance for rs13281615, 28 of them showed an imbalance in favor of the A allele while only 7 showed an imbalance favoring the G allele (p-value 0.012). A second SNP, rs6983267, showed a similar trend that did not reach statistical significance (p-value 0.157). These data raise the possibility that rs13291615 may be a candidate susceptibility allele for SCC. Our results suggest that the use of preferential allelic imbalance may be an efficient approach to map susceptibility variants in specific clinical settings.
In summary, our finding of an increased concordance of DNA copy number changes together with the presence of allelic-specific imbalance within separate cancers of individuals strongly suggests that the somatic changes occurring in tumors are in part affected by underlying characteristics of the individual host. An in depth comparison of allele-specific genomic changes occurring in multiple tumors of individual patients may offer a unique route to uncover cancer susceptibility loci.
The allele-specific LOH data from both microsatellite analysis and from SNP analysis indicate that the increased similarity of copy number changes within individuals is at least in part due to inherited variation within the same region as the copy number change. By contrast, not all loci that were frequently affected by concordant aberrations within individuals showed evidence of preferential allelic imbalance. This could be due to trans-effects between inherited variants elsewhere in the genome and a cancer gene in the region affected by the copy number alteration. For example, 13q12-q21, containing the Rb tumor suppressor gene, showed frequent loss in SCCs and a high intra-group concordance but did not show evidence of preferential imbalance.
There are some alternative explanations for the greater similarity of changes in tumors within versus between individuals. In our study we defined tumors as being independent based on arising in different anatomical sites. This should reduce the probability that tumors are related via a shared clonal origin. It is unlikely, but not impossible that tumors arising on different sites might have a common precursor which would explain the results of this study. It also remains possible that other pathogenetic factors such as ultraviolet light exposure or the presence of human papilloma virus may also influence the similarity of somatic alterations of tumors arising within an individual that do not show allele-specific imbalance. Finally, different immunosuppressive drugs may result in specific mutations occurring in tumors which might manifest as similar copy number patterns in tumors from within an individual. Another explanation of our results is that environmental exposures may result in differential selection between alleles which could result in allele-specific imbalances. Despite these possibilities, our study strongly supports the notion that the constitutional genotype of an individual exerts a strong influence on the somatic alterations that arise in cancer. Genetic analyses of cancer that arise at high multiplicity may offer a novel route to the discovery of cancer susceptibility genes.
All study participants signed informed consent and the study was approved by University of California San Francisco (UCSF) and Ohio State University (OSU) Institutional Review Boards. Participants were eligible if they had available SCC and normal tissue available for study. To reduce the possibility that tumors from the same individual might be related clonally, we chose tumors from different anatomical locations when they were excised on the same day. Tumors excised on different dates also needed to be excised from different anatomical locations. Re-excisions were not included in the study. Tumor DNA was microdissected from formalin-fixed paraffin embedded tissue sections containing at least 70% tumor cells and the concentration was measured using TaqMan analysis [22]. Blood DNA was used as a source of normal reference DNA for loss of heterozygosity analyses.
We obtained aCGH profiles from a total of 305 tumors from 181 patients and these consisted of one actinic keratosis, 37 Bowen's disease, 45 KAs and 222 SCCs. We focused our subsequent analyses on 222 SCCs from 135 individuals. The cohort included 25 patients who had 3 or more independent tumors that were examined by aCGH (number of tumors, n = 104). Tumor genomic DNA (1000ng) and reference DNA (600ng) (Promega) was labeled with Cy3 and Cy5, respectively using random primers essentially as previously described [23], [24]. The labeled tumor and reference DNA was pooled and applied to Hum3.2 BAC arrays for 48 hours. The arrays contained 2464 BAC clones with an average resolution of 100 Mb. Analysis of the arrays was carried out using R/Bioconductor software [25], [26]. Prior to analysis the data was normalized with respect to GC content and geometrical position on the arrays [27]. Regions of equal copy number were defined by segmenting the data using circular binary segmentation (CBS) [28]. The scaled median absolute deviation (MAD) of the difference between the observed and segmented values was used to estimate the sample-specific experimental variation; samples with a MAD of greater than 0.2 (n = 18) were considered unsuitable for inclusion in the study. The gain and loss status for each probe was defined using the merged level procedure [29].
For each autosomal arm, correlation coefficients based on log2 ratio values were computed for each pair of samples for those patients who had at least three independent SCC samples. Only those sample pairs were considered where at least one of the samples had 20% of clones with absolute value greater than 2 times sample MAD and another 20% below 2 times MAD. This ensured that the correlation was not driven by a flat sample profile. Only those arms were considered where there were at least 40 clones with non-missing values and at least 20 sample pairs in each of intra and inter groups; arms excluded from analyses include 2p, 3q, 4p, 5p, 6p, 6q, 7p, 7q, 9p, 9q, 10p, 10q, 11p, 12p, 12q, 15q, 16p, 16q, 17p, 18p, 18q, 19p, 19q, 20p, 21q and 22q. Two-sided, two sample t-tests were performed comparing the intra and inter patient groups for each of those arms where there was no significant difference in group variances. Brown-Forsythe version of the Levene-type test [30] was used to test for unequal group variances. The t-test p-values were then adjusted for multiple testing by Holm's method. Since a number of arms had unequal group variances, t-tests with Welch's approximation were also performed on each arm. Theoretical p-values were then adjusted by Holm's method. Genome-wide p-value was similarly computed by considering clones from all autosomes. There was no difference in group variances when considering whole genome. The correlation coefficients when considering individual arms and also whole genome had near normal distributions.
We identified clones having high within vs. between patient effects by estimating intraclass correlation coefficient (ICC) which captured the within-patient similarity. A random effects model Yij = μ + αi + εij, where the response variable is the CBS value with original log2ratio if a clone is an outlier in that segment, j and i represent the tumor and patient respectively, μ is an unobserved overall mean, αi is an unobserved random effect shared by all tumors in patient i, and εij is an unobserved noise term, was fit for each clone and ICC calculated as σα2/(σα2+σε2) where σα2 and σε2 are the variances of αi and εij respectively.
Matched normal and tumor DNAs from 65 individuals were genotyped for 45 microsatellite markers mapping to regions of common chromosomal loss or gain. A total of 270 skin tumors were studied for microsatellite LOH analyses. For allelotyping, we chose microsatellite markers with a high degree of heterozygosity that can be readily quantified. To allow efficient amplification from fixed tissue we selected microsatellite markers with PCR product sizes less than 200 bp [12]. Fluorescently labeled, multiplexed PCR products were analyzed on an ABI 377 DNA sequencer using GeneMapper v3.7 (Applied Biosystems) in the OSU Comprehensive Cancer Center Nucleic Acids Shared Resource. An allelic imbalance ratio (R) in each tumor sample for each marker was calculated using a standard protocol: R = (TA/TB)/(NA/NB), where TA is a peak height from tumor DNA of the larger sized allele, TB is the peak height area from tumor DNA of the smaller sized allele, NA is the peak height area from normal DNA of the larger allele, and NB is the peak height area from normal DNA of the smaller allele. As described by others, when R was greater than 1.5 or less than 0.66, the sample was considered to have allelic imbalance [31], [32]. When R was between 1.25 and 0.85 the sample was considered to have no imbalance. Other values for R were treated as uncertain.
Allelic imbalance data using microsatellite markers cannot simply be compiled across individuals, due to the heterozygosity of allele sizes across individuals each individual is likely to have a different combination of genotypes. We used a Bayesian/frequentist approach which we developed specifically for these data to determine if any given marker showed preferential allelic imbalance within and across patients (Text S1). In brief, we evaluated patient-specific odds of preferential imbalance as an indicator of randomness in a given individual using a Bayesian method. We then combined these odds into a sample to assess the evidence in favor of preferential imbalance in the general population using a frequentist method. A Wilcoxon rank sum test was then performed and all loci that rejected the null hypothesis at a 5% level of significance were deemed to show preferential imbalance.
We conducted quantitative genotyping of matched normal and SCC tumor DNA pairs using Sequenom MassARRAY Iplex gold genotyping technology. It is highly quantitative and is extremely sensitive for detection of allelic gains or losses in tumors and has been used for allelic imbalance studies [33]. All genotypes of poorer quality (aggressive calls) and those for whom a water sample had a strong call were eliminated from further analysis. Genotypes were also discarded from analysis if one of the two paired normal/tumor DNAs did not work resulting in genotypes included in analysis from 299 SCCs from 130 individuals for rs13281615, 110 SCCs for 70 individuals for rs1447295, and 175 SCCs from 84 individuals for rs6983267. Genotypes and peak area data for each allele were analyzed to identify regions of genomic imbalance between each matched normal and tumor DNA. An allelic imbalance ratio (R) to measure imbalance in each tumor sample for each SNP was calculated as described for the microsatellite LOH studies. Duplicate SNPs were included for quality control and two control samples and two no template controls were used. Chi-squared analyses were used to determine significance of observed allelic imbalances compared to expected 50∶50 imbalances indicative of random allelic imbalance.
|
10.1371/journal.pcbi.1006787 | A saturated reaction in repressor synthesis creates a daytime dead zone in circadian clocks | Negative feedback loops (NFLs) for circadian clocks include light-responsive reactions that allow the clocks to shift their phase depending on the timing of light signals. Phase response curves (PRCs) for light signals in various organisms include a time interval called a dead zone where light signals cause no phase shift during daytime. Although the importance of the dead zone for robust light entrainment is known, how the dead zone arises from the biochemical reactions in an NFL underlying circadian gene expression rhythms remains unclear. In addition, the observation that the light-responsive reactions in the NFL vary between organisms raises the question as to whether the mechanism for dead zone formation is common or distinct between different organisms. Here we reveal by mathematical modeling that the saturation of a biochemical reaction in repressor synthesis in an NFL is a common mechanism of daytime dead zone generation. If light signals increase the degradation of a repressor protein, as in Drosophila, the saturation of repressor mRNA transcription nullifies the effect of light signals, generating a dead zone. In contrast, if light signals induce the transcription of repressor mRNA, as in mammals, the saturation of repressor translation can generate a dead zone by cancelling the influence of excess amount of mRNA induced by light signals. Each of these saturated reactions is located next to the light-responsive reaction in the NFL, suggesting a design principle for daytime dead zone generation.
| Light-entrainable circadian clocks form behavioral and physiological rhythms in organisms. The light-entrainment properties of these clocks have been studied by measuring phase shifts caused by light pulses administered at different times. The phase response curves of various organisms include a time window called the dead zone where the phase of the clock does not respond to light pulses. However, the mechanism underlying the dead zone generation remains unclear. We show that the saturation of biochemical reactions in feedback loops for circadian oscillations generates a dead zone. The proposed mechanism is generic, as it functions in different models of the circadian clocks and biochemical oscillators. Our analysis indicates that light-entrainment properties are determined by biochemical reactions at the single-cell level.
| Circadian clocks in various organisms are composed of cell autonomous gene expression rhythms with a nearly 24-hour period. Transcriptional-translational feedback loops (TTFL) in single cells drive such rhythmic gene expression [1–3]. One of the most important roles of circadian clocks is to entrain behavioral and physiological rhythms in organisms to the light-dark (LD) cycle on earth. A light signal shifts the phase of the clocks by affecting the biochemical reactions in the TTFLs that regulate circadian gene expression. Such phase responses of circadian clocks to light signals allow their entrainment to the LD cycle.
Most organisms maintain their behavioral rhythms under constant dark (DD) conditions, indicating that their circadian clocks set subjective day and night. Subjective day and night under DD conditions can be defined by referring to the rhythms in an LD cycle. The phase responses of the circadian clocks to light signals have been examined by exposing organisms to short light pulses under DD conditions. Phase shift as a function of the timing of light exposure characterizes the entrainment property of the circadian clocks and is referred to as the phase response curve (PRC) [4, 5]. Intriguingly, the PRCs of different organisms have several common features [4, 6]. First, a light pulse at subjective morning advances the phase of the clock. Second, a light pulse at subjective evening and night delays the phase. Third, a light pulse at subjective daytime hardly changes the phase. This time window during daytime, when the phase of the clock is insensitive to light pulses, is referred to as the dead zone.
Previous theoretical studies revealed the PRC shape that is optimal for robust light entrainment. This optimal PRC is similar to those observed in several organisms and, remarkably, includes a dead zone during daytime [6–9]. This is because the presence of a dead zone increases the robustness of the clock against external fluctuations, such as fluctuation of light intensity [6] and daylight length [9], by reducing the responsiveness of circadian clock systems to external signals. However, the mechanism whereby a dead zone is created during daytime while preserving phase responses during the night remains unclear.
Two possible mechanisms may underlie the creation of a dead zone during daytime. One possibility is the gating of light input to reduce its influence on circadian clock genes [10, 11]. Gating is an elaborated mechanism, as it must be clock-dependent to distinguish day from night. Molecules that may be responsible for such gating have been identified previously [12, 13]. Gating makes circadian clocks robust against external perturbation as described above by reducing input signals detrimental to clock gene expression [6, 9]. However, gating is ineffective against internal perturbation that arises inside the gate, such as noise in gene expression and physiological states in cells [7, 8]. The alternative and more beneficial mechanism is the use of biochemical reactions in the TTFL to directly decrease the responsiveness of the phase of the clocks to stimuli during daytime. A dead zone formed in this way can make the clocks resistant to both external and internal perturbation because of the unresponsiveness of the phase of the clocks [7, 8]. Hence, some organisms should evolve to create a dead zone by biochemical reactions in TTFLs. Here we propose a design principle for creating such dead zones by analyzing the circadian clock systems of different organisms.
The biochemical reactions modulated by light signals in the TTFL for the circadian oscillation differ between species. In some organisms, such as Drosophila, light signals increase the degradation of circadian clock proteins, which we refer to as the degradation response (Fig 1A) [14, 15]. In Drosophila, the transcription of Period (Per) and Timeless (Tim) genes is induced by the CLOCK/CYCLE complex (Fig 1A). The PER/TIM complex then represses the transcriptional activity of CLOCK/CYCLE, forming a negative feedback loop (NFL) [16]. By this NFL, the abundance of TIM protein oscillates under both LD and DD conditions [14, 17]. Light signals activate Chryptochrome (Cry) and it degrades TIM protein [18–21]. This light-induced degradation of TIM allows the Drosophila clock to entrain to the LD cycle. As a result, the levels of TIM protein are lower during the day and higher during the night under LD conditions [14]. Differently, in other organisms such as mammals and Neurospora, light signals induce the transcription of repressor mRNA (Fig 1B) [10, 11, 22, 23]. In mammals, the transcription of Per and Cry is induced by the CLOCK/BMAL1 complex (Fig 1B). The PER/CRY complex then represses the transcriptional activity of CLOCK/BMAL1, forming an NFL as in the case of Drosophila. This NFL generates self-sustained rhythms of Per expression under both LD and DD conditions. Light signals induce the transcription of Per genes through the activation of CREB (Fig 1B) [10, 11, 24–26], allowing the mammalian clock to entrain to the LD cycle. In an LD cycle, Per expression levels are higher during the day and lower during the night [27]. We referred to this type of light response as an induction response.
Although light-responsive reactions differ between Drosophila and mammals, the mechanism for phase shifting at night is predicted to be the same: light signals induce repressor mRNAs when their concentrations are decreasing due to the strengthened feedback repression by the abundant repressor proteins. In Drosophila, degradation of TIM protein by light signals relieves transcriptional repression, leading to the induction of Tim mRNA. In mammals, light signals induce the transcription of Per via CREB. Thus, the elevation in repressor mRNA levels by light signals leads to phase shifts during the night in both systems.
On the other hand, there seems to be apparent differences in the effects of light signals on the dynamics of repressor mRNA and protein during daytime. In Drosophila, light signals increase the degradation of the repressor protein when its concentration is lower. In contrast, light signals in mammals increase repressor protein synthesis by inducing repressor mRNA when its concentration is already higher. These differences raise the question of whether the mechanisms for dead zone generation with different light responses are common or distinct. For the degradation response in Drosophila, several previous theoretical studies reproduced a dead zone without a clear explanation of its mechanism [28–30]. A reason for the elusiveness of this mechanism may be because an NFL with the degradation response can naturally create a dead zone without the inclusion of any additional reactions, as we discuss in this study. For the induction response in mammals, although various theoretical models have been proposed [29, 31–35], dead zone originated from unresponsiveness of phase of a clock has not been paid attention. Therefore, the mechanism underlying dead zone generation for the induction response observed in mammals should be clarified and compared with that for the degradation response observed in Drosophila to address the above questions.
Here, we reveal that the saturation of a biochemical reaction in the repressor synthesis in an NFL is a common mechanism to cancel the effect of light signals and create a dead zone in different organisms with the distinct light responses. The location of the saturated reaction in the NFL depends on the types of light responses. It is the saturation of repressor transcription that generates a dead zone with the degradation response, whereas it is the saturation of its translation with the induction response. In short, these saturated reactions in the repressor synthesis are located next to the light-responsive reactions in the NFL, suggesting a design principle for the dead zone generation during daytime.
We start with dead zone generation with the degradation response as observed in the Drosophila circadian clock (Fig 1A). As the neurons in the central pacemaker tissue are considered to determine the phase responses of individuals by entraining cells in peripheral tissues in general [36–39], we model a negative feedback loop in these pacemaker neurons. Because rhythms of these neurons are synchronized with each other by intercellular interactions [36, 40], they can be approximated as a single oscillator for simplicity. Previous theoretical studies reported a dead zone in PRCs with the degradation response [28–30]. However, which reaction processes are critical for the dead zone generation has not been clarified yet. To reveal the key determinants of the dead zone, we first consider the following dimensionless three-variable Goodwin model (Fig 1C):
1τdx(t)dt=11+(z(t)/K1)n−x(t),
(1)
1τdy(t)dt=γ1x(t)−γ2y(t),
(2)
1τdz(t)dt=γ2y(t)−(γ3+γl(t))z(t)Km+z(t),
(3)
where x, y and z are the concentrations of the repressor mRNA, repressor protein in cytoplasm and repressor protein in nucleus, respectively. In this model, the repressor protein is translated in the cytoplasm and transported into the nucleus. In Drosophila, these variables correspond to the levels of Tim mRNA and proteins in each cell compartment. K1 and n in Eq (1) are the threshold and Hill coefficient for transcriptional repression, respectively. γ1 is the translation rate and γ2 is the transport rate of repressor protein from the cytoplasm to the nucleus. We assume the saturation of nuclear protein degradation in Eq (3). γ3 and Km in Eq (3) are the maximum degradation rate of nuclear repressor protein and Michaelis constant, respectively. γl is the rate of degradation induced by transient light pulses and is specified below. The time constant τ can tune the period of oscillation without affecting other properties of a limit cycle. The three-variable model in the absence of light signals can generate stable limit cycles (Fig 2A). We set the origin of the horizontal axis in Fig 2 such that the levels of mRNA x take a minimum value at t = 0. The levels of Tim mRNA in Drosophila become lowest around dawn (~ CT 0) [17]. Hence, t = 0 in Fig 2 corresponds to the subjective dawn.
To examine the PRC, we consider the following form of light-induced perturbation in a reaction parameter (Fig 1D):
γl(t)={εltl≤t≤tl+Td0elsewhere
(4)
where tl is the onset of a light pulse and Td is the pulse duration. The parameter εl represents the rate of a light-induced biochemical reaction. For Eq (3), it is the rate of light-induced degradation of nuclear protein. We consider that εl reflects the strength of light. The value of εl becomes larger for a stronger light signal.
To obtain the light-induced phase shift Δϕ, we compute the difference in peak timing between perturbed and unperturbed systems (Fig 1E). A positive value of Δϕ indicates phase advance, whereas a negative value indicates phase delay. Typically, we run simulations for about 50 cycles after perturbation and measure the phase shift. Note that Δϕ quantifies the phase difference in terms of time. In this study, we examine the influence of each reaction parameter on the PRC. A change in the value of a reaction parameter may change the period of oscillation Tp. For clearer comparison of PRCs between different parameter values, we compute the phase shift normalized by the period of oscillation, Δϕ/Tp. In addition, the duration of the light pulse Td in Eq (4) is also scaled with the period Tp.
PRCs obtained by the above procedure may depend on the functional form of light-induced perturbation in biochemical reactions γl(t). Therefore, it is desirable to characterize phase responses to perturbation based solely on a limit cycle of the unperturbed system as a complement. If a perturbation by a light signal is sufficiently small, the properties of a PRC can be well characterized by the phase sensitivity of a limit cycle [6, 41]. Phase sensitivity describes how a small increase in state variables at given time t shifts the phase of oscillation. Namely, the modulus of phase sensitivity represents the responsiveness of a clock to perturbation. Suppose φ is the phase of oscillation defined in radians (0 ≤ φ < 2 π) and χ(φ) = (xχ(φ), yχ(φ), zχ(φ)) is a limit cycle solution of Eqs (1)–(3) in the absence of perturbation. See S1 Text for the details of the definition of phase in the entire state space. A small perturbation to the state variables at time t can be described as x(t) = χ(φ(t)) + μ η where η is a unit vector that specifies the direction of perturbation in the state space and μ is the modulus of the perturbation (μ ≪ 1). Then, the phase shift δφ caused by this perturbation reads:
δφ=φ(χ(φ)+μη)−φ(χ(φ))≈μ∂φ∂x|x=χ(φ)⋅η=μZ˜(φ)⋅η,
(5)
where Z˜(φ)=(Z˜x(φ),Z˜y(φ),Z˜z(φ))≡∂φ(χ(φ))/∂x. Thus, the 2π periodic function Z˜(φ) specifies the magnitude and direction of phase shift and is referred to as phase sensitivity [41]. A positive (negative) value of Z˜i (i ∈{x, y, z}) indicates that an infinitesimal increase of the variable i advances (delays) the phase of oscillation. Note that Z˜(φ) can be determined for a limit cycle in the unperturbed system. With this phase sensitivity, the phase shift Δϕ quantified by a peak phase difference between perturbed and unperturbed systems can be approximated as (see S1 Text for details):
Δϕ≈Tp2π∫tltl+TdZ˜(φ(t))⋅G(t,φ(t))dt,
(6)
where G(t, φ) is the perturbation in biochemical reactions by the light signal evaluated on the limit cycle χ. For example, G(t, φ) = (0, 0,–γl(t)zχ(φ)/(Km+zχ(φ))) for Eqs (1)–(3). Hence, if Z˜z(φ) for Eqs (1)–(3) involves an interval where Z˜z(φ)≈0, the interval will form a dead zone in the PRC. Although a dead zone can be formed in an interval where Gz(t, φ) ≈ 0 as well, such dead zone formation was examined previously [9, 42] and is out of the scope of the present study. Because the phase shift Δϕ is measured as the peak time difference, we introduce Z(t)≡(Tp/2π)Z~(φ(t)) to quantify the phase sensitivity in a unit of time. We compute the phase sensitivity for a limit cycle with the adjoint method as described in S1 Text.
Fig 2B shows the PRC of the model Eqs (1)–(3) when a short light pulse is administered at each time point tl. We use pulse duration Td = 0.5 Tp/24 in Fig 2. For example, if Tp = 24h, Td = 0.5h with this parametrization. During the night when the abundance of nuclear repressor protein z is higher, light signals shift the phase of oscillation (Fig 2B). A light pulse delays the phase of oscillation at which the levels of mRNA x are near their peak and those of nuclear repressor protein z are increasing (Fig 2B and S1A Fig). The reduction of repressor protein by a light signal during this time causes an increase in the transcription rate 1/(1+(z/K1)n), resulting in excess accumulation of mRNA (S1A Fig). Consequently, the nuclear repressor protein peaks later, delaying the initiation of the next cycle (S1A Fig). In contrast, a light pulse near the peak of z advances the phase of oscillation (Fig 2B and S1B Fig). The decrease in repressor protein during this time allows transcription to start earlier (S1B Fig). Thus, light signals induce the transcription of repressor mRNA by relieving transcriptional repression. The magnitude of phase shifts within these time windows becomes larger as the rate of light-induced degradation εl increases (Fig 2B). These results for the phase shifts are qualitatively consistent with previous experimental observations for the Drosophila circadian clock [14, 43].
In contrast, during daytime when the abundance of z is lower, the phase of oscillation does not change with light-induced degradation, indicating the presence of a dead zone (Fig 2B and S1C Fig). In this time window, the transcription rate of repressor is saturated at its maximum value due to the lower concentration of z (Fig 1C and S1C Fig). This saturation of transcription cancels the effect of light signals (S1C Fig), creating a dead zone. We also compute the phase sensitivity Z = (Zx, Zy, Zz). Because the nuclear protein z is decreased by a light signal through enhanced degradation, |Zz| is relevant to the magnitude of a phase shift. We consider –Zz to match the phase advance and delay zones between phase sensitivity and the PRC (Fig 2D). –Zz > 0 indicates phase advance by the decrease of z through light-induced degradation, while –Zz < 0 indicates phase delay. The magnitude of Zz is almost zero at the trough of nuclear protein concentration, confirming the existence of a dead zone (Fig 2D). The presence of a dead zone in phase sensitivity Zz indicates that the dead zone in the PRC shown in Fig 2B is not created merely by the lower rate of light induced degradation εl z/(Km + z) for z ≈ 0, but, indeed, lower phase sensitivity of the limit cycle to perturbation.
The continuous PRCs in Fig 2B are referred to as type 1. As the rate of light-induced degradation εl further increases, the PRC becomes discontinuous (S2A Fig), which is referred to as type 0. In S2A Fig, the breaking point (i.e., transition point of Δϕ/Tp from –0.5 to +0.5) is at around tl / Tp = 0.85, where the levels of nuclear protein are near their peak. Even in this stronger light-induced degradation, a dead zone is maintained. The breaking point and the dead zone length are rather insensitive to the change in εl once the type 0 PRC is created (S2A Fig). With these larger values of εl, the levels of repressor protein z become almost zero immediately after receiving a light pulse, meaning that the effect of light signals is saturated. In addition, the model can be entrained to a LD cycle (S2B Fig). The wave form of nuclear protein z peaks at night while reaching troughs in the daytime, which is consistent with experimental observations for TIM proteins under LD conditions [14, 15].
We then study how the dead zone length depends on the parameters in Eqs (1)–(3). We define the dead zone based on the magnitude of phase sensitivity Z. We detect a spanned time window where the phase of oscillation is insensitive to change in biochemical reactions induced by light signals:
{t1≤t≤t2||Zi(t)|<θ,|Zi(t1)|=|Zi(t2)|=θ},
(7)
where i ∈ {x, y, z} and θ is the threshold for phase irresponsiveness to perturbation. For the degradation response, phase sensitivity for nuclear protein Zz is relevant to phase shifts by light signals, i = z in Eq (7). We set θ = 10−1 throughout the study. Any time window that satisfies Eq (7) is considered to be a dead zone and we measure its length Ld = (t2 –t1)/Tp. Note that Ld is defined as the time interval normalized by the period of oscillation Tp.
We first examine the dependence of the dead zone length Ld and the amplitude of phase sensitivity –Zz on the Michaelis constant for protein degradation Km (Fig 3). The amplitude of phase sensitivity decreases as the value of Km decreases (Fig 3A and 3C). Instead, Ld monotonically increases as Km decreases (Fig 3A and 3C). When the value of Km is smaller and degradation is strongly saturated, the minimum levels of repressor proteins at troughs zmin are close to zero, zmin/K1 ≪ 1 (Fig 3B). The effect of light signals diminishes at that time because the light-induced degradation of nuclear protein does not further increase the transcription rate, 1/(1+(z/K1)n) ≈ 1 in Eq (1) (Fig 3D, S1C and S3B Figs). Thus, the result confirms that the saturation of transcription is required for dead zone generation. The other requirement is quick recovery of the levels of nuclear protein after the light-induced degradation. If the recovery is slow, the duration of transcription is extended due to the lower levels of nuclear protein caused by light-induced degradation (S3A Fig). This longer duration of transcription results in phase shifts. These requirements are more likely to be satisfied when zmin ≈ 0. Thus, a dead zone tends to be long as the time interval where z(t) ≈ 0 becomes long. These results suggest that the strong saturation of TIM degradation lengthens the dead zone of the PRC in the Drosophila circadian clock.
Next, we study the dependence of the dead zone length Ld on the other parameters in Eqs (1)–(3) (S4 Fig; S1 Text). Typically, Ld depends on reaction parameters non-monotonically because the minimum value of nuclear protein zmin changes non-monotonically as the value of each parameter changes. In general, the amplitude of oscillation becomes smaller near a Hopf bifurcation point that sets the lower and upper bounds of an oscillatory parameter range. In the vicinity of Hopf bifurcation points, zmin is near the steady state and is more likely to be well above zero. Hence, Ld tends to be non-monotonic between the lower and upper Hopf bifurcation points. In addition, zmin becomes larger before the Hopf bifurcation points due to the accumulation of nuclear protein with, for example, faster nuclear protein transport (larger value of γ2) and slower degradation (smaller value of γ3), further reducing the dead zone length. The details of the dependence of Ld and amplitude of phase sensitivity –Zz on each reaction parameter are described in S1 Text. For all the parameters, we find that Ld tends to be large when the values of z(t) at trough phase are close to zero. This observation suggests that each parameter influences the dead zone length by affecting the wave form of nuclear protein z(t).
In summary, the light-induced degradation of nuclear repressor protein induces transcription of the repressor mRNA. The elevation in mRNA levels result in phase shifts. A dead zone is formed if the light-induced degradation does not lead to a significant increase in x. Such time window arises when the nuclear protein concentration is significantly lower than the threshold for transcriptional repression K1. Thus, it is the saturation of repressor transcription that cancels the effect of light signals and creates a dead zone for the degradation response. Reaction parameters in the NFL affect the dead zone length by modulating the wave form of nuclear protein z.
To confirm the generality of the above results, we also analyze the dead zone in another model of Drosophila circadian clock [30]. The model includes interlocked feedback loops of PER/TIM and CLOCK/CYCLE. Qualitatively, the same results are obtained using this more complex Drosophila model (S5A–S5D Fig; S1 Text). Furthermore, we also consider a biochemical oscillator other than circadian clocks. We adopt a repressilator model for the synthetic oscillator [44]. We obtain same results using the repressilator model (S5E–S5H Fig; S1 Text), indicating that the proposed mechanism is robust and generic. Finally, we note that other previous models that realized dead zones also included the saturation of repressor mRNA synthesis and that of repressor degradation [28, 32, 45]. Thus, our current analysis highlights the relevance of saturation of repressor mRNA synthesis to dead zone formation.
Next, we consider a model for the induction response (Fig 1B). As in the case of the degradation response, we model a negative feedback loop in central pacemaker neurons. In mammals, neurons in the suprachiasmatic nucleus (SCN) in the brain receive light signals from the eyes and determine the phase responses of individuals by entraining peripheral clocks [37–40]. Light signals induce the transcription of Per genes in these neurons. We describe this light response with the following dimensionless differential equations:
1τdx(t)dt=1(1+(z(t)/K1)n)+γl(t)−x(t)
(8)
1τdy(t)dt=γ1x(t)−γ2y(t),
(9)
1τdz(t)dt=γ2y(t)−γ3z(t)Km+z(t),
(10)
where γl(t) in Eq (8) is the induction rate of a clock gene by a light signal. γl(t) is the same function as defined in Eq (4). The light signal induces transcription of repressor mRNA at rate εl independent of the concentration of repressor protein in Eqs (4) and (8). This may represent the induction of Per genes by CREB through CRE element in the mammalian circadian clock (Fig 1B) [24, 25]. The inclusion of light-induced transcription in this form differs from previous theoretical studies [6, 29, 32]. These previous studies assumed that the nuclear repressor protein also repressed the light-induced transcription. Therefore, in these models, the effect of light signals was diminished when the protein levels were high. Because light signals only influence the transcription of repressor mRNA in Eqs (8)–(10), phase sensitivity for x, Zx underlies phase shifts.
The model Eqs (8)–(10) generates stable limit cycles with appropriate parameter sets (Fig 4A). In Fig 4, we set t = 0 to the time at which the levels of mRNA x are at the minimum value. In the mammalian SCN, the expression levels of Per genes are lowest at around CT20 [27]. Hence, the origin of the horizontal axis in Fig 4 may correspond to the subjective midnight.
We then examine the phase shifts with the induction response (Fig 4B and 4C). We do not find an extended dead zone in either the PRC (Fig 4B) or the phase sensitivity Zx (Fig 4C) of Eqs (8)–(10). Rather, Δϕ and Zx intersect with zero steeply at a single time point. Phase delays are caused by light signals near the peak of mRNA. An increase in mRNA near its peak time results in an increase in the levels of nuclear protein and lengthens the duration of repression (S6 Fig). We further examine whether a dead zone is formed in Eqs (8)–(10) with other different parameter sets. For this, we randomly generate 2000 parameter sets from uniform distributions with which Eqs (8)–(10) can generate stable limit cycle oscillations (S1 Text). We compute the phase sensitivity Zx for each random parameter set and check the length Ld of the spanned time window that satisfies the condition Eq (7). Ld of all the 2000 parameter sets examined is less than 1/24 (e.g., for Tp = 24h, Ld = 1/24 indicates a dead zone of 1h). Thus, our numerical results suggest that the NFL model Eqs (8)–(10) with the induction response does not form an extended dead zone in the PRC.
The analysis of the degradation response described in previous sections implies that a dead zone can be formed when a light signal does not increase the levels of nuclear repressor protein. For this, cancellation of the influence of mRNA induction by light signals may be required. This consideration leads us to introducing a saturation of a biochemical reaction in the NFL. We first test the saturation of protein transport from the cytoplasm to the nucleus, but it does not generate a dead zone (S7 Fig; S1 Text). We then test the saturation of mRNA degradation (S8 Fig; S1 Text). In this case, although a dead zone is formed at night when the concentration of repressor mRNA is lower, a daytime dead zone is not generated (S8 Fig). Finally, we introduce a Michaelis-Menten function in the translation process in Eq (9):
1τdy(t)dt=γ1x(t)Kt+x(t)−γ2y(t),
(11)
where Kt is the Michaelis constant for translation. Translational regulation by certain RNA binding proteins [46, 47] may cause Michaelis-Menten type nonlinearity as assumed in Eq (11). Although a previous theoretical study examined the effect of a saturated translation term mainly on the period of oscillation [48], its effect on phase responses to light signals has not been studied. We simulate the model Eqs (8), (10) and (11), and find that it can generate sustained oscillations (Fig 5A). Remarkably, the saturated translation can generate an extended dead zone in the PRC at subjective daytime when the levels of mRNA x(t) are near their peaks (Fig 5B). The dead zone appears robustly even when we use different values of induction strength εl in Eqs (4) and (8) (Fig 5B). To quantify the degree of saturation, we define a saturation index for translation sx = x(t)/(Kt + x(t)) [49]. The value of sx is close to 1 when the translation is saturated and close to 0 when less saturated. The time series of sx shows that the translation is indeed strongly saturated within the dead zone (Fig 5C). We also find that the dead zone is present in the phase sensitivity Zx (Fig 5D). We then check whether other parameter sets can create similar dead zones. Of 2000 randomly generated parameter sets, 54 (2.7%) form similar dead zones of length Ld greater than 1/24. The reason for this relatively small percentage of longer dead zones is that the values of Kt are large in most of those 2000 random parameter sets, meaning that translation is not saturated strong enough to generate a longer dead zone. Larger values of Kt are favored because the saturation of translation tends to suppress oscillations [49], as we discuss in the discussion section.
Changes in time series of nuclear protein z generate phase shifts. The induction by light at the early increase phase of mRNA can increase the levels of nuclear protein near its trough (S9A Fig). Due to the excess amount of nuclear protein, the forthcoming peak of mRNA decreases. Accordingly, the levels of nuclear protein at the forthcoming peak are also decreased. Hence, the repression of mRNA synthesis relieves faster, advancing the phase. A light pulse at the late increase, the peak, and the early decrease phases of mRNA only weakly influences the levels of nuclear protein z due to saturation of translation (S9B Fig). An unaltered time series of z does not cause a phase shift. The induction of mRNA by light at the late decrease phase of mRNA can increase the peak levels of nuclear protein (S9C Fig). This lengthens the duration of transcriptional repression, delaying the phase.
This model can realize type 0 PRC at stronger light intensity (S10A Fig). Unlike the type 0 PRC in the degradation response (S2A Fig), the breaking point and shape of the PRC change depending on the strength of light induction εl (S10A Fig). Excess induction of mRNA lengthens the duration satisfying x > Kt. This allows for the production of excess protein and delays its peaks, resulting in phase delays of greater magnitude (S10A Fig). We also confirm that the model can be entrained to an LD cycle (S10B Fig). The levels of mRNA peak during daytime in the LD cycle, which is consistent with experimental observations of Pers in mammals.
We then examine the parameter dependence of the dead zone length by computing the phase sensitivity Zx for Eqs (8), (10) and (11) (Fig 6 and S11 Fig). We start with the dependence of Zx on the Michaelis constant for translation Kt in Eq (11) (Fig 6A). Smaller values of Kt lead to stronger saturation of translation when the levels of mRNA are higher (Fig 6B). For each value of Kt, we measure the length of a time window Ld, within which the absolute value of phase sensitivity satisfies |Zx| < θ (i = x and θ = 10−1 in Eq (7)). Ld is larger for smaller values of Kt and it decreases monotonically with an increase in Kt (Fig 6C). The amplitude of Zx also monotonically decreases as Kt increases (Fig 6D). Thus, the saturation of repressor translation increases both the dead zone length and phase sensitivity.
We next study the dependence of the dead zone length Ld and the amplitude of phase sensitivity Zx on the other parameters in Eqs (8), (10) and (11) (S11 Fig). Overall, the dependence of Ld on each parameter is similar to that observed in the degradation response (Fig 3 and S4 Fig). The reason for this observation is as follows. In the induction response, the larger amplitude and wider wave form of repressor mRNA x extend the dead zone by lengthening the time interval during which translation is saturated. To achieve this condition, the levels of nuclear protein z at its trough must be near zero. This common requirement underlies the similarity in the parameter dependence of Ld between the degradation and induction responses.
Furthermore, as observed in the degradation response, the dead zone length Ld often changes non-monotonically as the value of a parameter increases (S11 Fig). For example, Ld depends on the maximum translation rate γ1 non-monotonically (S11A Fig). When γ1 is smaller, the amplitude of mRNA x is small, resulting in smaller Ld values. In contrast, when γ1 is larger, the amplitude of x is large whereas the width of its wave form narrows due to the higher levels of nuclear protein z. As a balance of these two contributions, Ld peaks near the lower Hopf bifurcation point (S11A Fig). A similar trend can be seen in the dependence on the transport rate γ2 where Ld peaks near the lower Hopf bifurcation point (S11B Fig) and the maximum degradation rate of nuclear protein γ3 where Ld peaks near the upper Hopf bifurcation point (S11C Fig). The dead zone length also depends non-monotonically on the threshold constant for transcriptional repression K1 (S11D Fig), as the amplitude of mRNA x changes non-monotonically between the lower and upper Hopf bifurcation points. As in the degradation response, the dead zone length becomes longer monotonically for smaller values of the Michaelis constant for protein degradation Km (S11E Fig). The result indicates that the stronger saturation of protein degradation is more likely to lead to the generation of a dead zone.
Each reaction parameter also influences the amplitude of phase sensitivity Zx (S11 Fig). Changes in the values of the maximum translation rate γ1 and protein degradation rate γ3 strongly influence the magnitude of phase delay rather than causing phase advancement (S11A and S11C Fig). Changes in the values of the other parameters affect the magnitudes of both phase advance and delay (S11B, S11D and S11E Fig). As observed in the degradation response, the amplitude of phase sensitivity becomes larger near the Hopf bifurcation points (S11 Fig).
To further study the effect of nonlinearity in translation on the dead zone generation, we extend the Michaelis-Menten function in Eq (11) into a Hill function:
1τdy(t)dt=γ1x(t)hKth+x(t)h−γ2y(t),
(12)
where h is the Hill coefficient and the parameter Kt can now be interpreted as the threshold level of repressor mRNA required for translation to occur. Translation does not occur as long as x/Kt ≪ 1 for a large value of h. Time evolution of mRNA x and nuclear protein z are given by Eqs (8) and (10).
We study the dependence of the dead zone length on the Hill coefficient h with the same parameter set used in the analysis of the Michaelis-Menten translation function except for Km. For better illustration of the influence of h, we set Km = 0.053 in Fig 7, which is larger than the value used in Fig 5 (Km = 0.025). We find that the increase in h lengthens the dead zone (Fig 7A and 7B). This is because a larger h extends the interval of x where xh/(Kth+xh) ≈ 1. In addition, larger values of h increase the amplitude of x (Fig 7C). The amplitude of the PRC also increases with h (Fig 7A and 7D). The sharp transition of translational activity near x ~ Kt set by the Hill function more strongly influences the levels of nuclear protein z, resulting in a larger phase shift by a light signal. These results suggest that a switch in translation extends the dead zone in the PRC.
Finally, to confirm the generality of the above results for the induction response, we analyze the dead zone in a more complex model of the mammalian circadian clock. The model includes the NFL of Per and Bmal1 and that between Bmal1 and Rev-erb (S1 Text). Saturation of translation creates a dead zone in this complex model (S12A and S12B Fig; S1 Text). In addition, we obtain a dead zone with the repressilator model including translational saturation and the induction response to external signals (S12C and S12D Fig; S1 Text). In summary, when light signals increase the mRNA synthesis independent of the clock states, the saturation of translation is required to generate the daytime dead zone.
The biochemical reactions that are influenced by light signals in circadian clock systems vary between organisms. In this study, we considered the degradation and induction responses as observed in Drosophila and mammals, respectively. Despite the difference in light responses in these two different animals, light signals induce the transcription of the repressor and cause phase shifts during night as described below. In the degradation response, light signals increase the degradation of repressor protein, thereby relieving transcriptional repression. Subsequent elevation of mRNA levels determines the phase shifts of the clock (S1 Fig). In the induction response where light signals directly induce the transcription of repressors, it is the subsequent increase in protein levels that determines the phase shifts (S6–S9 Figs). In contrast, a dead zone is formed in both types of light responses in the daytime. Previous studies demonstrated the importance of a dead zone during the daytime to make the circadian clock systems resistant against internal and external perturbation [6–9]. However, whether and how the TTFLs for the circadian clocks create the dead zone has remained unclear. In this study, we revealed that the saturation of a biochemical reaction in repressor synthesis is a common mechanism for the different light responses to reduce the phase sensitivity of a limit cycle and create a dead zone (Figs 2, 3 and 5). The degradation response requires the saturated transcription of repressor mRNA to generate a dead zone, whereas the induction response requires saturated repressor translation. Our theoretical results suggest that locating a saturated reaction in repressor synthesis next to a light-responsive reaction in an NFL is a design principle for dead zone generation.
The saturation of biochemical reactions in an NFL influences the generation of oscillations [49–51]. For example, the saturation of translation and transport of repressor protein from the cytoplasm to nucleus suppresses oscillations [49]. Conversely, the saturation of mRNA and protein degradation facilitates oscillations. Therefore, the location of the saturated reaction for dead zone generation should affect the ability of the NFL to generate oscillations. It is ideal if the saturated reaction for dead zone generation can facilitate generation of sustained oscillation. However, because the location of the saturated reaction for dead zone generation is constrained by that of the light-responsive reaction in an NFL as described above, it may not be optimal in terms of rhythm generation in some organisms. For the degradation response, the saturation of transcription at its maximum rate is required to create a dead zone (Figs 2C and 3D and S1C Fig). The saturated transcription can realize an effective transcriptional switch, as the resultant accumulation of repressor protein causes a subsequent rapid drop in the transcription rate (Fig 3D). Such a switch in transcription is favorable to oscillation, as the Hill function with a larger Hill coefficient in transcription facilitates oscillation [50, 52]. For the induction response, however, the saturation of translation is essential to create a dead zone (Figs 5 and 6). Although strong saturation of translation deprives the NFL of the ability to generate oscillation as described above (S13 Fig; S1 Text) [49], other additional saturations such as that of protein degradation could compensate (S13 Fig). Importantly, the saturation of protein degradation not only facilitates oscillations but also lengthens the dead zone in a PRC in both the degradation and induction responses (Fig 3, S3 and S11 Figs). A recent theoretical study proposed that the saturated degradation of molecules can be regarded as a positive feedback [51]. This suggests that there may be some positive feedbacks that support dead zone generation by the saturation of repressor synthesis, as the saturated degradation does.
Crucially, our analysis of the induction response indicates that not every saturated reaction in an NFL can create a dead zone during daytime. For example, the saturation of repressor transport from the cytoplasm to the nucleus does not generate a dead zone in the PRC (S7 Fig). The saturation of mRNA degradation does create a dead zone but only at night when the levels of mRNA are near the trough (S8 Fig). According to these results, the translation of repressor protein is the most plausible reaction to saturate for dead zone generation during the daytime (Fig 5). Additionally, we note that nonlinear functions of protein translation other than saturation cannot create a dead zone in the daytime, as we demonstrate that a Hill function for translation with a larger threshold Kt results in the generation of a dead zone at night (S14 Fig; S1 Text). A previous experimental study demonstrated that PER protein synthesis induced by vasoactive intestinal polypeptide (VIP) indeed saturated [53]. However, because the quantification was performed by the bioluminescence of the luciferase reporter fused with the functional PER2 protein, of which synthesis is under the control of endogenous transcriptional and translational regulations [54], it remains open whether this saturation occurred at the translation step of the PER protein. Recent experimental studies have revealed the importance of posttranscriptional regulations in the circadian clock [55]. Several micro RNAs inhibit the translation of clock gene transcripts by leading them to degradation. For example, micro RNAs regulate the onset of circadian rhythms in mouse embryonic tissue by determining the localization of Clock mRNA [56]. On the other hand, some protein molecules are known to promote translation. For example, mammalian LARK binds to Per1 mRNA to promote translation of the PER1 protein [46]. Mouse heterogeneous nuclear ribonucleoprotein Q (mhnRNP Q) binds to 5'-UTR of Per1 mRNA, which is necessary for internal ribosomal entry site mediated translation [47]. The saturated translation assumed in the present study may be caused by these mRNA binding proteins. Interestingly, although the levels of Lark and mhnRNP Q mRNAs are constant, the protein levels of both are rhythmic [46, 47]. Therefore, it is an important future work to reveal the significance of such rhythmic translational activities together with saturation in the phase responses of the circadian clock.
In general, a dead zone in the PRC can be generated in several ways. One way is to gate the light input to the circadian clock genes at a particular phase of oscillation. Because light signals do not influence the expression of circadian clock genes in the presence of gating, the phase sensitivity of the limit cycle does not need to be near zero to form a dead zone [9, 42]. Previous theoretical studies realized such gating by assuming that circadian clock proteins repress the light-induced transcription [9, 29, 42]. In mammals, light signals do not induce the expression of Per genes in the central pace maker tissue SCN during subjective day [10, 11], suggesting that the SCN gates the light input. However, the molecular mechanism of gating was not elucidated. Moreover, if the circadian system only utilizes gating, the system would remain susceptible to internal perturbation such as fluctuations in gene expression and physiological states in cells. In this study, we proposed an alternative mechanism of dead zone generation that is effective against both external and internal fluctuations [6–9] and functions in different organisms with different light responses. We note that the mechanism proposed in the present study can function together with gating and further improve clock precision. A dead zone like interval has been observed in the PRC of firing rate of neurons in the rat SCN explants treated with VIP for few minutes [57]. PRCs of single cells derived from peripheral tissues are typically type 0 [58–61], most likely due to strong phase resetting effects of applied perturbations (S10 Fig). A dead zone tends to be obscured in a type 0 PRC (S10 Fig), making it hard to perform direct comparison with type 1 PRCs of the current study. A future study will address whether SCN can create a dead zone at a single cell level by the saturated translation of Per mRNAs. Treatment of lower concentration of VIP or forskolin, which also activates CREB, to a cell line derived from the rat SCN [62] can be used to realize type 1 PRCs in single cells and address this question.
In conclusion, the saturation of a biochemical reaction in repressor synthesis is a simple and generic mechanism for the generation of a dead zone for light signals. Several environmental cues other than light signals change the phase of the circadian clock by influencing the rates of biochemical reactions. The PRCs for those cues may include a dead zone as observed in the responses to light pulses [63]. Our findings indicate that the saturation of biochemical reactions should also function in such dead zone generation.
|
10.1371/journal.pcbi.1003944 | Spinal Mechanisms May Provide a Combination of Intermittent and Continuous Control of Human Posture: Predictions from a Biologically Based Neuromusculoskeletal Model | Several models have been employed to study human postural control during upright quiet stance. Most have adopted an inverted pendulum approximation to the standing human and theoretical models to account for the neural feedback necessary to keep balance. The present study adds to the previous efforts in focusing more closely on modelling the physiological mechanisms of important elements associated with the control of human posture. This paper studies neuromuscular mechanisms behind upright stance control by means of a biologically based large-scale neuromusculoskeletal (NMS) model. It encompasses: i) conductance-based spinal neuron models (motor neurons and interneurons); ii) muscle proprioceptor models (spindle and Golgi tendon organ) providing sensory afferent feedback; iii) Hill-type muscle models of the leg plantar and dorsiflexors; and iv) an inverted pendulum model for the body biomechanics during upright stance. The motor neuron pools are driven by stochastic spike trains. Simulation results showed that the neuromechanical outputs generated by the NMS model resemble experimental data from subjects standing on a stable surface. Interesting findings were that: i) an intermittent pattern of muscle activation emerged from this posture control model for two of the leg muscles (Medial and Lateral Gastrocnemius); and ii) the Soleus muscle was mostly activated in a continuous manner. These results suggest that the spinal cord anatomy and neurophysiology (e.g., motor unit types, synaptic connectivities, ordered recruitment), along with the modulation of afferent activity, may account for the mixture of intermittent and continuous control that has been a subject of debate in recent studies on postural control. Another finding was the occurrence of the so-called “paradoxical” behaviour of muscle fibre lengths as a function of postural sway. The simulations confirmed previous conjectures that reciprocal inhibition is possibly contributing to this effect, but on the other hand showed that this effect may arise without any anticipatory neural control mechanism.
| The control of upright stance is a challenging task since the objective is to maintain the equilibrium of an intrinsically unstable biomechanical system. Somatosensory information is used by the central nervous system to modulate muscle contraction, which prevents the body from falling. While the visual and vestibular systems also provide important additional sensory information, a human being with only somatosensory inputs is able to maintain an upright stance. In this study, we used a biologically-based large-scale neuromusculoskeletal model driven only by somatosensory feedback to investigate human postural control from a neurophysiological point of view. No neural structures above the spinal cord were included in the model. The results showed that the model based on a spinal control of posture can reproduce several neuromechanical outcomes previously reported in the literature, including an intermittent muscle activation. Since this intermittent muscular recruitment is an emergent property of this spinal-like controller, we argue that the so-called intermittent control of upright stance might be produced by an interplay between spinal cord properties and modulated sensory inflow.
| The maintenance of upright quiet stance is a challenging task for the central nervous system (CNS), and the objective is to achieve the control of an intrinsically unstable biomechanical system under the effect of gravity. Posture control is a position control problem in which the CNS, leg muscles, and different sensory systems (e.g., the muscle proprioceptors) interact to maintain the horizontal projection of the centre of mass (COM) within a region bounded by the feet (for a review of the basics of posture control, see [1]). Sensory systems, such as the vestibular, visual, and somatosensory, play a significant role in the aforementioned motor task, so that disorders in any of these systems may lead to postural instability [2], [3].
A conceptual question under debate in the literature concerns the manner the CNS controls upright stance in adults (i.e., human beings who already learned to walk). Some researchers argue in favour of a negative-feedback continuous control, with the leg muscles reflexively activated in response to drifts or perturbations away from an equilibrium position [4]–[6]. Conversely, others suggest that an anticipatory (feedforward) mechanism is necessary to explain some findings, such as the low feedback gains observed in some experiments [7]–[9]. More recently, the view that human upright stance is controlled by an intermittent mechanism has grown [10]–[15]. The latter is based in part on results from an experiment involving the manual balancing of a virtual inverted pendulum controlled by a subject through a computer joystick [13]. Additionally, other recent studies showed that motor units (MUs) of the Medial Gastrocnemius (MG) muscle exhibited an intermittent recruitment during upright quiet stance with a pattern closely linked to COM and centre of pressure (COP) fluctuations [14], [16]. It is noteworthy that the previously referred papers used a wide-sense/physiological conceptualisation of intermittent postural control, which is understood as one exhibiting discrete (bursty) actions of the neuronal controller, producing ballistic-like (phasic) muscle activations. In addition to this more qualitative/physiological view of intermittency, there is a strict, quantitative, definition of intermittent control that has also been used to analyse motor control problems, such as stick balancing and postural control [17]. In the present study we will adopt the former (physiological) view of intermittency.
From a theoretical standpoint, mathematical models have been developed to describe and to investigate human postural control. These models explicitly incorporate the above-mentioned control strategies hypothesised to be adopted by the CNS during upright standing control, but do not identify which part of the CNS the control resides, e.g. cortex, brainstem, cerebellum, spinal cord. For instance, [4], [5], [18]–[20] represented the postural control system as a negative-feedback continuous control system, so that proprioceptive information of body position and velocity were fed back to the CNS. Alternatively, the models proposed by [21]–[23] adopted a continuous predictive control system, while others [11], [12], [24], [25] represented the control of posture as an intermittent control system. Despite these fundamental differences, all these models were based on a control engineering framework, whereby the whole system was simplified and the CNS was represented by a PD/PID (proportional-derivative/proportional-integral-derivative) controller, an optimal continuous controller or an intermittent controller that activated the muscles in discrete bursts. Despite their advantages of a relative simplicity and the power of explanation, it is not easy to translate their results in terms of the underlying physiological mechanisms involved in the control of human posture.
The present study aims at providing a complementary approach to those mentioned above in that the fundamental focus is on biology (anatomy and physiology). A large amount of physiological knowledge, from the behaviour of neuronal ionic channels to the dynamics of muscle contraction, was funnelled into a large-scale mathematical model of the nervous, muscular, and biomechanical systems involved in posture control. In this vein, a biologically based large-scale neuromusculoskeletal (NMS) model was developed and used to investigate the problem of postural control from a more neurophysiological standpoint. The model encompasses: i) a spinal neuronal network, which includes conductance-based models of both motor neurons (MNs) and interneurons (INs); ii) Hill-type muscle models to represent the viscoelastic properties of the Soleus (SO), MG, Lateral Gastrocnemius (LG), and Tibialis Anterior (TA) muscles; iii) models of both muscle spindle and Golgi tendon organ (GTO); iv) afferent fibres providing Ia, II, and Ib feedback; and v) an inverted pendulum model, which is a first approximation of upright quiet stance [26]. Here the proprioceptive feedback is provided by muscular proprioceptors (i.e., spindles and GTOs), since they seem to be largely responsible for the position sense of the limb [27], [28]. The first hypothesis of the present study is that stance control might be properly achieved by a spinal-like controller (SLC, approximated here by the developed NMS model) based on a proprioceptive feedback. An associated second hypothesis is that the activation of leg muscles by this SLC is a continuous process yielding the maintenance of human body equilibrium.
Another relevant issue related to postural control that the present study analyses is related to recent experimental findings of a “paradoxical” behaviour of the calf's muscle fibre lengths during postural sway [29], [30]. The behaviour was called “paradoxical” because the muscle fibre lengths were negatively correlated with COM/COP displacements [31]. These studies proposed that reciprocal inhibition from antagonistic (TA) Ia afferents might be responsible for this unexpected motor behaviour. With the availability of the detailed NMS model employed in the present study, the correlation between muscle fibre length and COM/COP displacements could be tackled without much difficulty. Therefore, a third hypothesis of the present study is that a model structure without the reciprocal inhibition pathway (from TA Ia afferents to Triceps Surae MNs) would exhibit positive correlation coefficients between muscle fibre length and COM/COP displacements. A complement of this hypothesis is that the reciprocal inhibition neural circuit may increase the probability of occurrence of the “paradoxical” relation mentioned above. This combined third hypothesis is, therefore, evaluated here using the complex NMS physiological model in order to verify a hypothesis put forward by previous authors on the basis of heuristics and experimental results from humans [29]–[31].
To the best of our knowledge, this is the first study to address the control of an intrinsically unstable neuro-biomechanical system associated with the maintenance of human quiet standing by means of a complex large-scale system mostly based on known physiology. However, others have also used biologically based reductionist models of the NMS system to investigate how the CNS controls other motor tasks [32], [33]. A minor part of this material was already published as conference abstracts [34], [35].
Typical biomechanical and neuronal outputs of the NMS model are presented in Figure 1. The model's responses resemble qualitatively those frequently reported in postural control studies (e.g., [9], [29], [36]). Irrespective of model structure (i.e., Model 1 or Model 2 - see Methods for details), the inverted pendulum leaned about 5 deg forward (equilibrium point), so that COM and COP displacements oscillated around a basal value of 80 mm (see Figure 1A). The basal plantar-flexion torque (negative torque) was 10% of the maximum isometric torque produced by the model. One can notice that COM and COP (Figure 1A) oscillated in anti-phase with respect to the muscle torque (Figure 1B), i.e. when the body leaned forward from its equilibrium position the plantar-flexion torque increased (more negative). Conversely, muscle activations (EMG envelopes in Figure 1C-E) were modulated approximately in phase with postural sway. In the simulations, TA muscle was silent during postural sway (not shown).
A quantitative analysis was performed to validate the model with respect to the available data from the literature. Typical time-domain metrics were calculated from the COP time series and compared to data from normal subjects and vestibular loss patients standing on a force plate without visual information (see Table 1). Root mean square (RMS) and mean velocity (MV) of simulated COP were higher than the values observed experimentally in normal subjects, but compatible with data from vestibular loss patients. Another quantitative validation was based on a cross-correlation analysis performed between the COM and COP time series (Figure 2A-B), as well as between COP and EMG envelopes (Figure 2C-D). COM and COP were highly correlated () at lag zero. COP and EMG envelopes were positively correlated with the correlation peak occurring at a positive lag. Correlation coefficients () and cross-correlation peak lag values were compatible with experimental data from healthy subjects (see Table 1). In general, correlation coefficients were higher for Gastrocnemii in comparison to the SO, and muscles' activations (EMGs) were advanced by approximately 200–300 ms in relation to the postural sway (COP). The 50% power frequency () estimated from the COP power spectrum (see Figure 2E-F) resulted quite similar to the value from healthy subjects (see Table 1). COP power spectra of both model structures were limited to 1 Hz.
A final quantitative validation was based on the pooled histogram of COM displacements (1-mm bins) as shown in Figure 3 (data are from the simulations of Model 2). The histogram shape was bimodal, with two peaks around the equilibrium position of the inverted pendulum (value 0 in the abscissa). The Jarque-Bera goodness-of-fit test was applied to verify if this histogram could be fitted by a typical Gaussian probability density function [11]. The null-hypothesis (the histogram comes from an unimodal Gaussian function) was rejected (). The same result was obtained for Model 1.
Figures 4 and 5 show how the spike trains from spinal MNs, INs, and afferent fibres were modulated during postural sway. An interesting qualitative finding was that MUs from the MG muscle were intermittently recruited/de-recruited as the inverted pendulum swayed forward/backward (Figure 4B). This intermittent pattern of MU recruitment was similar for the LG muscle (not shown), but less evident for the SO muscle (see Figure 5A). The degree of intermittency for the MG and SO MUs was quantified by the activation ratio (see [16] and Methods for details). The median (range) activation ratios calculated for 90 randomly selected MG MUs (30 MUs were chosen per simulation) from Model 1 and Model 2 were 0.69 (0.44–0.80) and 0.65 (0.47–0.81), respectively. For 90 randomly selected SO MUs the activation ratios were 0.97 (0.75–1) and 0.96 (0.79–1) for Model 1 and Model 2, respectively. Because of these results, the MG and LG muscles were considered to have ballistic-like activations (see EMG envelopes in Figures 1D-E and 4B), while the SO muscle was mostly tonically (continuously) active during the maintenance of an upright posture (see Figures 1C and 5A).
In order to quantify the intermittent recruitment of MG MUs, the interval between successive recruitments was computed for a subset of 30 randomly chosen MUs (10 MUs were chosen per simulation). In accordance with the method used by [14], intermittent recruitment was considered if a given MU was discharging at a rate lower than 4 Hz (i.e., interspike intervals higher than 250 ms). For Model 1, 899 intervals of 30 MG MUs were evaluated and the mean (modal) interval between successive recruitments was equal to 511 (274) ms [i.e. a mean (modal) rate equal to 1.96 (3.65) Hz]. Similarly, for Model 2, 846 intervals of 30 MG MUs had a mean (modal) value of 505 (277) ms [1.98 (3.61) Hz]. Therefore, both model structures produced a similar intermittent recruitment pattern on MG MUs. A low number of LG MUs was recruited (less than 30) and the SO MUs were mostly tonically active during the simulation of postural control (see the activation ratios in previous paragraph), hence the intermittency of the MUs from these muscles were not quantitatively evaluated here.
Panels C-I in Figure 4 and panel B in Figure 5 show typical results of how proprioceptive feedback (encompassing afferent fibres and spinal INs) was modulated during sway (Model 2 was used for this simulation). The activity of the Ia afferents from the MG muscle (Figure 4C) was highly modulated, following approximately the COM/COP displacement (note the firing rate modulation for three different Ia afferents, as indicated by the thin lines). Since there was little variation in the MG muscle torque (RMS value 2.50% of the maximum MG muscle torque) and the mean MG muscle fibre length was maintained at an approximately steady value (i.e., there was little change in the static component of the muscle fibre length - see below), the activities of Ib and type II afferents were minimally modulated (see panels G and H in Figure 4). The proprioceptive pathways responsible for the reciprocal inhibition were also highly modulated during postural sway (see panels F and I in Figure 4). Inhibitory Ia INs discharged phasically when the inverted pendulum swayed backward and this contributed to a decrease in the ankle joint torque generated by the plantar flexor muscles. Conversely, Ia afferents from the SO muscle spindles were poorly modulated in the posture control task (see Figure 5B).
The intermittent recruitment of MG MUs was evaluated on the basis of two phase plots that relate angular velocity and muscle torque with ankle angle data obtained from the postural control model (Figure 6). Figure 6A shows that most of the MG MUs (60%) were recruited when the inverted pendulum was leaning forward from its equilibrium position irrespective of its velocity (first and fourth quadrants of the angle-velocity phase plots). Nonetheless, a large number of MUs (28%) were recruited when the inverted pendulum was at a backward position but with a positive velocity (second quadrant), i.e., the pendulum was starting to return to a forward position. Similarly, most of the MG MUs (50%) were recruited when the pendulum was leaning forward and producing a higher plantar flexion torque (fourth quadrant in the Figure 6B), i.e., the pendulum was at a forward position and decelerating. Almost 35% of the MUs were recruited when the pendulum was at a backward position and with a lower (more positive than the mean value) plantar flexion torque. In general, the discharges in the first and third quadrants were mainly involved in the generation of a basal torque, while the discharges in the second and fourth quadrants represented the phasic corrective torque control produced by the MG muscle.
Two model structures were adopted to investigate the effect of the reciprocal inhibition pathway on postural control (see Methods for details). According to the data presented in the previous section, there were no dramatic differences in the time- and frequency-domain metrics, as well as in the MU recruitments. The COP power spectrum calculated from the model without reciprocal inhibition (Model 1) was slightly narrower than that from Model 2. In addition, correlation coefficients between COP and EMG envelopes were lower when the reciprocal inhibition pathway was included in the model. Nonetheless, both model structures produced fluctuating outputs that resembled experimental data, suggesting that reciprocal inhibition from TA Ia afferents is not a strict requisite for the control of upright standing.
In the present study, we tested the hypothesis that reciprocal inhibition may be responsible for the negative correlation between muscle fibre length and COM/COP displacement [29]–[31]. This was tested by performing a correlation analysis between COM displacements and muscle fibre lengths for the SO, MG, and TA. Typical signals representing these variables are shown in Figure 7. For this typical simulation (Model 2) one can notice that MG muscle fibre length was positively correlated with COM, while TA muscle fibre length and COM displacement were negatively correlated. The latter results are a typical “orthodox” behaviour that has been shown for some healthy subjects during upright quiet standing [29]. For the SO muscle, a more quantitative analysis was performed (see Figure 8). Correlation analysis between SO muscle fibre length and COM displacement was performed on 3-s windows (see dashed vertical lines in Figure 7) according to the method adopted in [29]. Correlation coefficients calculated from Model 1 and Model 2 were pooled into two groups: positively correlated () and negatively correlated (). For Model 1 most of the intervals (80%) showed a positive correlation between SO muscle fibre length and COM displacement (i.e., “orthodox” behaviour), whereas for Model 2 the number of negatively correlated intervals increased to 50%. -test revealed a statistically significant difference () between the responses of the two model structures. This suggests that, at least for the SO muscle, reciprocal inhibition might be involved in the genesis of the “paradoxical” behaviour of muscle fibre length. Nonetheless, a small percentage of negatively correlated intervals (20%) might also be generated in the absence of any reciprocal inhibition from antagonistic Ia afferents.
A large-scale NMS model was applied in the present study to the problem of controlling upright standing in humans. A different feature of this approach in comparison with most of the previous studies in the literature (e.g., [4]–[6], [10]–[12], [18], [19], [22]–[24]) is that the structure and behaviours of each element were based on known physiology, anatomy, and biomechanics encompassing important parts of the postural control system. Therefore, the control strategy employed by the modelled CNS emerged from the interaction between several neuromechanical elements involved. While the overall mechanisms that control the inverted pendulum sway are beyond an analytical understanding, the increased biological realism provides important clues regarding some putative neurophysiological mechanisms underlying the posture control task. In the following sections, the results presented earlier shall be discussed with respect to relevant experimental findings reported in the literature.
The results presented here showed that a SLC was effective in maintaining the equilibrium of an intrinsically unstable biomechanical system (see Figure 1 for an illustrative example). This leads to the first contribution of this study, which is to support the hypothesis that human upright quiet standing may be properly controlled by spinal mechanisms, for example, without any cortical involvement. This view is consistent with several studies, which suggest that cortical control is decreased or may be absent when a motor task is well trained (e.g., [30]–[40]). Normal quiet postural control, under no special restrictions (as standing on a narrow beam or during dual tasks), is certainly a candidate for a well-trained task that would not require cortical control. Since the scope of the present study is limited to the investigation of neurophysiological mechanisms underlying the control of quiet standing, the assumed lack of supraspinal neural structures should not limit the ensuing interpretations. A separate section below (see Model Limitations and Future Research) presents and discusses the limitations both of the modelling as well as the conclusions derived from the simulations.
Time- and frequency-domain metrics obtained from the NMS model were compatible with experimental data (see Table 1). The equilibrium values of the inverted pendulum (mean angle, torque, and COM/COP displacements) varied slightly both within and between simulations due to programmed (randomised) changes in system configuration. Between-simulation changes occurred due to changes in neuronal connectivities and intrinsic properties (e.g., action potential thresholds) that were randomly attributed at the beginning of each simulation [41]. On the other hand, within-simulation variations were mainly related to the number of recruited MUs, which varied stochastically due to neuronal noise. Therefore, the mean equilibrium position of the body depends on the overall instantaneous configuration of the postural control system. The analysis of the COP time series showed that the model was less stable (i.e., presented a larger postural oscillation) than healthy young subjects [42]. Notwithstanding, simulated data were compatible with those from vestibular loss subjects standing on a stable surface without visual information [3]. As a consequence, the simulation results reinforce that the increased postural oscillation observed in patients may be due to the lack of other sensory inputs providing information to the CNS, such as vestibular and visual sources. Or, in other words, the proprioceptive feedback gain in such patients is not sufficient to replace the other missing sensory feedback modalities. Interestingly, the variability observed in the simulated postural sway was exclusively generated by the variability in sensory afferents and descending commands, which results in random fluctuations of motoneuronal discharges. Therefore, it is predicted that most of the biomechanical variability (sway) observed during upright standing has a neuronal origin and is less influenced by internal disturbing forces (e.g., heartbeats and respiration) as proposed elsewhere [4], [6], [8], [11], [25].
The cross-correlation analysis (Figure 2) showed that EMGs from the Triceps Surae (TS) muscles were positively correlated with postural sway (as measured by the COP). Simulation results are in accordance with experimental data that showed higher correlation coefficients between EMGs from Gastrocnemii and COP [9], [43]. Additionally, the time lags between COP and EMGs were within a range of 200–300 ms, which is also compatible with experimental data [9], [43]. This is in some sense a remarkable result that emerged from the NMS model since the sum of the afferent and efferent action potential propagation delays is much smaller than this time lag between COP and EMG. Albeit qualitative, or semi-quantitative, this is a relatively strong sign that the model was able to capture at least a part of the overall system dynamics. In [9] it is argued that the existence of this lag between the mechanical and neuronal responses would be due to an anticipatory action of the neuronal controller, i.e., the postural control is mediated by a feedforward mechanism. However, theoretical and computer simulation studies [19] showed that even in the absence of any feedforward mechanism, lags between neuromechanical signals may be obtained depending on the parameters of the continuous feedback system and stochastic features of the input signals. The results presented here corroborate the latter view, since no feedforward mechanism was incorporated into the NMS model.
Another experimental finding from human postural control studies that was reproduced by the model was the bimodal distribution of COM displacements (see Figure 3). In [11] most of the data from young subjects exhibited double-peaked histograms of COM displacements with a local minimum in between (see their Figure 5). The authors of the above-mentioned study [11] showed that the bimodal distribution of COM displacements was only obtained when they represented the postural control system by a mixture of both continuous and intermittent (with a phasic controller operating at a 3–4 Hz burst rate) control mechanisms. A continuous postural control model produced unimodal Gaussian-like histograms. Therefore, they argued that the postural control in humans is not mediated exclusively by a continuous control mechanism. Our results corroborate this proposal. However, in our approach the control structure (e.g., continuous, intermittent, or a mixture of both) was not imposed a priori. The mixture of continuous (SO motor nucleus and muscle fibres) and intermittent (mostly Gastrocnemii motor nuclei and muscle fibres) control behaviours was a result of the interactions of the several interconnected neuromusculoskeletal elements of the model and their respective dynamics. Further discussion on the issue of continuous and intermittent control mechanisms operating during postural control shall be presented in a separate section below.
It is worth mentioning that a key parameter for stabilizing the inverted pendulum model was the constant level of the fusimotor activity that adjusted the sensitivity of muscle spindles for each simulation. Without a proper value for both static and dynamic fusimotor activities the pendulum fell at the beginning of the simulation. As previously mentioned, some parameters were randomly distributed in each simulation run [41], hence producing a different set of initial conditions in different runs. This explains why the mean values of fusimotor activities vary (slightly) across the different simulation trials (see Methods). In the context of human postural control, the dependence of an effective control upon the fusimotor activity suggests that the CNS must properly set the muscle spindle sensitivity for the performance of the task [27], [44], [45].
Another point that should be stressed is that the NMS model was not stabilised without the neuronal activity, i.e., the pendulum fell when the SLC was turned off. This is compatible with the current view that the viscoelastic properties of the muscles around the ankle joint are not sufficient to control upright standing [8], [46]–[48]. In the model, the intrinsic passive ankle joint stiffness was about 70% of the critical toppling torque (see Methods for details), which is in accordance with the experimental estimates at ankle joint rotation angles similar to those obtained in our model (about 0.60 deg on average) [46], [48]. The adoption of a constant passive stiffness is a typical simplification (see for instance [5], [11]) that was also adopted in the present study. In an analytical study [49], the conclusion was that the feedback provided by the muscle spindles and GTOs is not sufficient to stabilize an inverted pendulum representing the human body. However, the authors did not considered any passive mechanism at the joint level and, hence, the total torque was generated by the active neural controller. Here, as the passive properties were included, the demand of the CNS was about 30% of the necessary stabilizing torque, which is compatible with the experimental estimates [11], [46].
The most remarkable result obtained from the proposed NMS model is shown in Figure 4. The intermittent recruitment of MG MUs is a phenomenon recently observed in human experiments [14], [16] and the postural control model reproduced this behaviour with a high degree of fidelity. The experimental study in [14] reported that MG MUs were intermittently recruited with a modal frequency of 2 Hz (pooled data from 7 subjects), which is similar to the value observed in [13] for actions of a human subject manually controlling an inverted pendulum. A central hypothesis raised by several recent studies postulates the involvement of an intrinsic predictive mechanism used by the CNS in the performance of postural control [13], [15], [50]. This intrinsic mechanism, sometimes named as a “refractory response planner” [51] and involving a “psychological refractory period” [15], [50], [51], would be responsible for the intermittent actions of the neural controller during the equilibrium maintenance of an unstable load. The simulation results showed that even in the absence of any predictive mechanism (or an internal time setting neural circuit), actions of the neuronal controller occurred at a mean (modal) rate of 2 (4) Hz, i.e., MUs from the MG muscle were recruited with a mean (modal) interval of 500 (250) ms, irrespective of model structure (i.e. Model 1 and Model 2). The models adopted in this study are interpreted as representing a single subject instead of a population of subjects, and the MU intermittence rates observed in the simulations are within the experimental range (2–4 Hz) reported elsewhere (e.g., [11], [13], [14]). These data suggests that the interplay between a SLC and the muscles involved in the task being performed is sufficient to provide a mechanism underlying the intermittent actions of the CNS during postural control. No complex central mechanism (e.g., predictive, response planner) was needed in our model for the genesis of this control pattern.
As discussed in [14], [16], the MG muscle seems to be mostly involved in balance control during standing, while the SO muscle provides a basal torque due to its mostly continuous activity (see Figure 5 and Figure 1C). Regarding the LG muscle, a recent finding showed that this muscle has a minimal or absent activation during the postural control task [16]. The simulation data are in agreement with these experimental results, and suggest that the differences in the organisation of the MG and SO motor nuclei might be responsible for their different actions during postural control. Additionally, the LG muscle was minimally activated during the simulations (the reason why we performed quantitative analysis only on MG MUs), although the recruited LG MUs (less than 30 per simulation) followed a pattern similar to the MG MUs. This different behaviour between the lateral and medial parts of the Gastrocnemius might be explained by a different number of MNs innervating each muscle. The LG muscle has approximately 60% less MUs than the MG muscle (see Methods), and hence for a similar effective synaptic current the number of recruited MUs with similar intrinsic characteristics would be lower for the LG muscle in comparison to the MG. In the model, the SO muscle is more homogeneous, having a high number of low-threshold and smaller twitch amplitude MUs, while MG and LG muscles have an equal proportion of low- and high-threshold MUs generating higher twitch amplitudes (and mostly briefer twitches). The lower proportion of low-threshold MUs (with a similar recruitment range - see [16]) might naturally produce the intermittent recruitment due to fluctuations in sensory feedback. The rationale is that for a similar lower sensory inflow (or effective synaptic current), a low number of MG/LG MUs are recruited, hence producing a low torque at the ankle joint. Conversely, a higher number of SO MUs are recruited producing a basal torque sufficient to counterbalance the static toppling torque. During a forward sway, any small modulation of sensory inflow is sufficient to recruit additional higher-threshold MG/LG MUs, counterbalancing the postural perturbation. As the inverted pendulum returns to a backward position, sensory inflow decreases and the recruited MUs are de-recruited (see Figure 4). For the SO, these oscillations in sensory inflow seem to be lower during postural sway (see Figure 5B). However, the analysis is quite complex since the system is operating in closed loop, so that any argument based on causality may lead to logical difficulties. In spite of this limitation, the simulation results indicate that upright standing could be controlled by means of proprioceptive sensory information feedback and a mixture of continuous (SO muscle) and intermittent (mostly MG and LG muscles) action of the CNS. Therefore, the second working hypothesis raised in the present study (that efferent actions of the CNS are continuous during postural control - see Introduction) turned out to be half true, i.e., the leg muscles are activated by a combination of both continuous and intermittent processes.
The results in Figure 6 showed that MG MUs were preferentially recruited when the body leaned forward (panel A in Figure 6) and decelerated (panel B in Figure 6), i.e., MG MUs were mostly recruited in order to counteract the toppling torque due to gravity, pushing the body to a backward position. These simulation results are similar to those obtained from human subjects, as reported in [14]. The authors of the referred study proposed that a strategy of MU recruitment instead of MU rate modulation during upright standing would be generated by the CNS due to the postural task demands. The data from the simulations in Figure 6 reinforce this view. The preference for recruitment coding would be due to the same mechanisms discussed in the previous paragraph, i.e., structural features of the MG motor nucleus and modulation of sensory information due to perturbation from a mean equilibrium position. On the other hand, recent experimental and computer simulation studies have shown that during isometric contractions, the TS torque control relies mainly on rate coding [52] and the variability observed in both torque and EMGs is highly dependent on the MU discharge rate variability. Therefore, the same muscle group (i.e., the TS) is probably being driven according to two different laws depending on the motor task: rate coding for isometric torque control in a very stable condition, and recruitment coding (for the MG/LG muscles) in a more challenging condition, such as erect posture. Interestingly, recent experimental data relating postural sway with isometric torque variability (at similar mean torque values) in young subjects found that they have a positive correlation [53] albeit the first is much larger in magnitude than the latter. As the isometric torque control (seated subjects) involved almost certainly only continuous feedback (mostly from the SO) this experimental result gives support to the dual control mode (continuous and intermittent) that was found in the present simulations for standing posture control.
Two model structures were simulated in order to investigate whether postural control may be influenced by the reciprocal inhibition pathway (see Methods for details). Recent studies have discussed the importance of reciprocal inhibition in movement control. For instance, [29], [30] hypothesised that this inhibitory pathway may be a better source of feedback control since TA proprioceptive activity is unmodulated by the homonymous muscle activation during postural sway.
An interesting result was that in comparison to the model without reciprocal inhibition (Model 1) the complete model (Model 2) showed an increased number of intervals in which SO muscle fibre length was negatively correlated with the COM displacement (see Figure 8). This “paradoxical” behaviour was reported in some experimental studies [29]–[31] and was used as evidence to postulate the significant role of reciprocal inhibition in the control of upright quiet standing [29]. The simulation results corroborate the hypothesis that the “paradoxical” behaviour of muscle fibre lengths may be generated by the reciprocal inhibition pathway. Nevertheless, no interval with negative correlation was found between MG and LG muscle fibre lengths and COM. In [29], the authors reported that two out of eight subjects showed a larger number of positively correlated intervals for the MG muscle, and they discussed that these subjects oscillated in a more forward position. For the physiologically-constrained set of parameters adopted in the present model the mean equilibrium position was 5 deg forward, which is similar to experimental findings [36]. Therefore, further studies are necessary to better understand the real significance of “paradoxical” muscle fibre behaviour and how it emerges during upright stance control. Yet, it is interesting that a highly complex and unexpected biological phenomenon may be partly explained/reproduced by a biologically plausible NMS model, and, therefore, providing neurophysiological clues to its genesis.
Regarding basic postural sway metrics (e.g., COP RMS, MV, and spectral contents) the simulation results did not show large differences between the two model structures (see Table 1), suggesting that reciprocal inhibition is not a fundamental mechanism for postural control.
In spite of the suggestion that TA muscle spindles must be a better (“cleaner”) source of ankle angle feedback than TS muscle spindles [29] the simulation results from Model 1 (without reciprocal inhibition) showed that even “noisy” sensory feedback from the TS muscle receptors is sufficient for an adequate postural control. The TS spindle feedback is “noisy” in the sense that the TS muscle receptors are signalling a mixture of information from ankle angle changes as well as changes in muscle length and tension due to the MN pool activation.
One conclusion that can be reached from the present simulation results is that mechanisms beyond those included in the model are not strictly necessary to reproduce experimental data from other studies. However, it is not possible to exclude that, despite theoretically not necessary, such mechanisms play a role in human postural control. Specifically, contributions from additional sensory modalities, such as foot soles, joint and skin receptors, vision, and vestibular system, certainly contribute by varying degrees to postural control depending on the particular experimental conditions [2], [3], [42], [54]. Additionally, one cannot rule out the involvement of supraspinal centres (e.g., brainstem, basal ganglia, primary motor cortex) [51], [55], specially if the maintenance of upright standing is being learned, such as in infants and adults recovering from a serious medical/neurological disease. Modulations of fusimotor [44], [56] and presynaptic inhibition activities [57], [58] are examples of important spinal-related mechanisms that certainly play relevant roles too.
In a general context, proprioceptive information from the legs is provided by muscle, joint and cutaneous receptors [54], [59]. However, in the NMS model presented here the proprioceptive information was provided exclusively by muscle receptors (i.e. muscle spindles and GTOs), which are postulated as being primary sources of sensory information in response to limb movement [27], [28]. The degree of influence of joint and cutaneous receptors on postural control is a controversial issue in the literature. Some experimental findings showed little change in postural sway after ischemia or anaesthesia [60], [61], while others showed that stimulation of cutaneous afferents evoked postural changes during quiet standing [54]. The simulation results presented here are in accordance with the former view, but further theoretical/computational and/or experimental studies are required to investigate what is the relative contribution of additional neuronal structures for the maintenance of human upright standing.
Regarding the biomechanics of the human body, it is well known that the inverted pendulum is an approximation for the human body during quiet standing [4], [26]. Other expansions, for instance, multi-link and/or multi-dimensional (e.g., including medial-lateral oscillations) models and analyses [24] are very interesting research avenues. However, any biomechanical expansion in large-scale models such as that used in the present study should envisage an increasing number of neuronal and musculoskeletal elements, along with the complexities of their interactions.
Finally, it is noteworthy that we simulated postural control during relatively brief periods (about 30 s). Prolonged unconstrained standing is associated with large changes in body equilibrium position along time [62]. A NMS model to provide approximate postural control during prolonged standing would probably require a reasonably higher complexity than the one employed in the present model and this is certainly an interesting challenge for future research.
Large-scale modelling has been adopted in several fields of modern neuroscience research (e.g., [32], [33], [63]–[67]). To our knowledge this is the first study aimed at modelling the NMS system involved in the control of human upright standing from a more neurophysiological perspective. The main conclusion drawn from the simulation results is that posture control might be, at least in part, mediated by spinal mechanisms, with proprioceptive information being fed back to the spinal neuronal circuitry. Additionally, the results provided evidence that complex phenomena observed in human experiments, such as intermittent actions of the CNS, might not depend on intricate control strategies of complex structures within the CNS. The structure and organisation of the spinal cord, i.e., the types of MUs, synaptic connectivities, the ordered recruitment of MUs, as well as the modulation of proprioceptive information could be sufficient to explain how the CNS controls body position during upright quiet standing in a general sense.
The model proposed in this study encompasses four main subsystems (neuronal controller, muscles, proprioceptors, and body biomechanics) that were interconnected to represent the NMS system involved in the control of human upright stance. An overview of this large-scale model is depicted in Figure 9. It is worth mentioning that the model is aimed to study body sway in the sagittal plane during unperturbed stance. In this condition, the posture control task relies mainly on afferent and efferent activities associated with the muscles around the ankle joint (ankle strategy) [9]. Figure 9A shows a schematic view of the neuronal circuitry composed of groups of spinal MNs and INs (see mathematical description below), referred to as the SLC (Spinal-Like Controller). MNs were assembled in motor nuclei associated to the TS (i.e., SO, MG, and LG) and TA muscles, which is the most relevant antagonist group of muscles involved in postural control during ankle strategy [9], [29], [36]. Stochastic descending commands represented part of the synaptic inputs from the brain that drive the spinal MNs during the maintenance of upright standing. Musculotendon units (MTUs) were represented by Hill-type models (see mathematical description below), which were driven by the spike trains from the spinal MNs (Figure 9B). Muscle spindles were placed in parallel with the muscle fibres and received commands from Gamma motor neurons (-MNs), while GTOs were placed in series with the tendon. Proprioceptive feedback was provided by Ia, II and Ib axons mediating: i) monosynaptic Ia excitation; ii) di-synaptic Ib inhibition; iii) di-synaptic II excitation; and iv) reciprocal inhibition from antagonistic Ia afferents. These are fundamental pathways frequently associated with different motor tasks, including upright standing [58]. An inverted pendulum was adopted to represent the body biomechanics (Figure 9C), which is a first approximation for unperturbed quiet standing [4], [5], [10], [11], [20], [25], [26]. In the following sections, the mathematical details concerning each of these models will be provided.
The models described above were written in Java programming language (Oracle Corp) and simulated in an open-source web-based application developed in the Eclipse IDE (The Eclipse Foundation). Model source code is available for download at the website http://remoto.leb.usp.br/ and at a public repository (http://dx.doi.org/10.6084/m9.figshare.1029085). Differential equations were numerically solved by a fourth-order Runge-Kutta method with a 50 µs time step.
Recent experimental studies argued that the reciprocal inhibition from antagonistic Ia afferents (from TA muscle spindles to TS MNs - see Figure 9 A) has an important role in the postural control due to the “orthodox” modulation of TA length and the lack of TA contraction during standing [16], [29]. To investigate this potential role of reciprocal inhibition pathway, two different model structures were used in this study. The first one, hereafter referred to as Model 1, excluded the reciprocal inhibition pathway so that all proprioceptive information was provided exclusively by the TS sensory afferents (autogenic pathway). Conversely, Model 2 is the complete model fully described above. All common parts in both models have the same parameter values.
Three independent simulations (30-s duration) were carried out with each model. During the first second of simulation, the movement of the inverted pendulum was restricted so that the neuromuscular system reached its steady-state condition. This prevented instabilities due to the lack of neural activity. In order to provide a basal activation, descending commands were modelled as 400 homogeneous stochastic Gamma point processes with an individual mean rate equal to 50 Hz and a shape factor equal to 25 (coefficient of variation equal to 20%). With this descending drive (without any feedback information) the MN pools produced a small basal muscle torque equal to approximately 2% of the maximum torque. As mentioned earlier, the fusimotor drive was modelled as a Gaussian stochastic process with a small variance. The mean values of static and dynamic fusimotor activities were adjusted for each simulation so as to produce outputs in which the inverted pendulum was oscillating around an equilibrium point for the whole simulation duration. For Model 1 the values of static (dynamic) fusimotor activities were 31.10 (33.30), 31.20 (33.30), and 31.50 (33.60). Similarly, for Model 2 the values of static (dynamic) fusimotor activities were set to 32 (33.80), 32 (33), and 31.70 (32.70).
Several output variables were analysed in the present study: i) Spike trains from MNs, INs and afferent fibres; ii) EMGs from the TS muscles; iii) COM and COP anteroposterior displacements; iv) Muscular torque; v) Muscle fibre lengths; vi) Joint angle and angular velocity. These raw variables are available for download at a public repository (Model 1 - http://dx.doi.org/10.6084/m9.figshare.1027609; Model 2 - http://dx.doi.org/10.6084/m9.figshare.1029084). Analyses were performed on 22.50-s duration time series, so that the initial 5 s and final 2.50 s were removed due to transient responses and data filtering (see below).
COM, COP, joint angle, angular velocity, muscle fibre length, and torque time series were detrended and downsampled to 2 kHz. In order to compare model behaviour with experimental data recorded with force plates, time- and frequency-domain metrics were estimated from the COP, for instance, RMS, MV and 50% power frequency [42]. Power spectra were estimated by Welch's method. EMGs were downsampled to 2 kHz, rectified, and lowpass filtered at 2 Hz (zero-phase fourth-order Butterworth filter).
Similar to the studies reported in [9], [19], [43], cross-correlation analysis was performed between COM and COP time series, as well as COP and EMGs. In addition, similar to the analysis reported in [29], correlation coefficients were computed between COM and muscle fibre lengths. For the latter analysis COM and muscle fibre length time series were binned to 3-s duration intervals (without overlap) so that the correlation coefficients were calculated for each bin (see Figure 7).
In order to quantify the degree of intermittency of a given muscle, the activation ratio of individual MUs was measured. This ratio was calculated as the percentage of time a given MU had interspike intervals lower than 250 ms [16]. An activation ratio equal to 0 indicates that the MU was inactive during the whole simulation time, whereas a ratio equal to 1 indicates a continuous activity. Therefore, lower activation ratios represent an intermittently discharging MU. Additionally, phase plots were constructed to evaluate the MU recruitment during quiet standing. Recruitment phase plots were constructed by binning joint angle, angular velocity, muscular torque and spike trains in 100-ms non-overlapping windows. Within each window, the number of recruited MUs was counted and a circle with diameter proportional to the counting number was plotted in 2-dimension angle-velocity (see Figure 6A) or angle-torque graphs (see Figure 6B). Data from the three simulations were pooled together in the same phase plot. A similar analysis was performed in [14].
When applicable, statistical analysis was performed with a significance level set at 0.05. All analyses were performed in Matlab (The MathWorks Inc) and SPSS Statistics (IBM Corp).
|
10.1371/journal.pcbi.1001124 | Quantifying the Relative Contributions of Divisive and Subtractive Feedback to Rhythm Generation | Biological systems are characterized by a high number of interacting components. Determining the role of each component is difficult, addressed here in the context of biological oscillations. Rhythmic behavior can result from the interplay of positive feedback that promotes bistability between high and low activity, and slow negative feedback that switches the system between the high and low activity states. Many biological oscillators include two types of negative feedback processes: divisive (decreases the gain of the positive feedback loop) and subtractive (increases the input threshold) that both contribute to slowly move the system between the high- and low-activity states. Can we determine the relative contribution of each type of negative feedback process to the rhythmic activity? Does one dominate? Do they control the active and silent phase equally? To answer these questions we use a neural network model with excitatory coupling, regulated by synaptic depression (divisive) and cellular adaptation (subtractive feedback). We first attempt to apply standard experimental methodologies: either passive observation to correlate the variations of a variable of interest to system behavior, or deletion of a component to establish whether a component is critical for the system. We find that these two strategies can lead to contradictory conclusions, and at best their interpretive power is limited. We instead develop a computational measure of the contribution of a process, by evaluating the sensitivity of the active (high activity) and silent (low activity) phase durations to the time constant of the process. The measure shows that both processes control the active phase, in proportion to their speed and relative weight. However, only the subtractive process plays a major role in setting the duration of the silent phase. This computational method can be used to analyze the role of negative feedback processes in a wide range of biological rhythms.
| As modern experimental techniques uncover new components in biological systems and describe their mutual interactions, the problem of determining the contribution of each component becomes critical. The many feedback loops created by these interactions can lead to oscillatory behavior. Examples of oscillations in biology include the cell cycle, circadian rhythms, the electrical activity of excitable cells, and predator-prey systems. While we understand how negative feedback loops can cause oscillations, when multiple feedback loops are present it becomes difficult to identify the dominant mechanism(s), if any. We address the problem of establishing the relative contribution of a feedback process using a biological oscillator model for which oscillations are controlled by two types of slow negative feedback. To determine which is the dominant process, we first use standard experimental methodologies: either passive observation to correlate a variable's behavior to system activity, or deletion of a component to establish whether that component is critical for the system. We find that these methods have limited applicability to the determination of the dominant process. We then develop a new quantitative measure of the contribution of each process to the oscillations. This computational method can be extended to a wide variety of oscillatory systems.
| Biological systems involve a large number of components that interact nonlinearly to produce complex behaviors. How can we determine the role that a component plays in producing a given behavior of the system? We approach this question in the relatively simple context of relaxation oscillations, since relaxation oscillator models and their extensions are used to describe a wide variety of biological behaviors [1], such as the cell cycle [2], electrical activity of cardiac and neural cells [3], [4], circadian patterns of protein synthesis [5], metabolic oscillations [6] and episodic activity in neuronal networks [7]. Specifically, we use a model developed to describe the rhythmic activity of developing neural networks and whose formalism also applies to cellular pacemakers [8]. The activity of the system can be either high or low, and slow negative feedback processes switch the system back and forth between the active and silent states. Hence the rhythm consists of episodes of high activity separated by silent phases, repeated periodically. While relaxation oscillator models usually contain one negative feedback process to regulate the rhythmic activity, in biological systems two or more feedback processes are often present. Thus, we consider a model with two different types of negative feedback: divisive and subtractive. In the context of an excitatory network, synaptic depression (weakening of synaptic connections between neurons) is a divisive feedback (decreasing the slope of the network input/output function) while activation of a cellular adaptation process (decreasing the neurons' excitability) can be a subtractive feedback (shifting the network input/output function) [8], [9]. With both types of negative feedback in the model, we seek to determine the contribution that each makes to episode initiation and termination.
We begin by using two strategies based on the two broad types of experimental protocols. The correlative strategy seeks to detect associations between the time course of a variable and the system's behavior. To use the example of episodic activity generated by an excitatory neural network, we compare the variation of the fraction of undepressed synapses (weighted by the synaptic conductance) to the activation of the cellular adaptation current (scaled by its conductance). Intuitively, the process that shows the greatest changes would be considered to affect activity the most, and thus contribute more to episode initiation/termination. The faster process covers a wider range during the active and silent phases [8]. This predicts that the faster a process and the larger its associated weight, the more it contributes to episode initiation and termination. The second strategy is to block one process, with the expectation that this will result in changes in activity that are directly related to the contribution of that process. Perhaps contrary to intuition, we find that blocking a slow process may provide little information on the role of that process in the rhythm generation, and that the correlative and blocking strategies may even lead to contradictory conclusions.
We then develop a new strategy based on the idea that if a negative feedback process contributes significantly to episode termination, then increasing its time constant should significantly increase episode duration. Similarly, if recovery of such a process contributes to episode initiation, then increasing its time constant should significantly delay episode initiation. We develop a measure of the respective contribution of each process based on these ideas. This measure reveals that if the divisive and subtractive feedback processes have similar time scales and similar weight they contribute similarly to episode termination. In contrast, the subtractive process controls episode initiation, even if it is slower or has less weight. This also means that the divisive process only plays a minor role in episode initiation. This unexpected result was not revealed using the standard approaches, and demonstrates the utility of the new measure in pulling out the key dynamics involved in rhythm generation. These results demonstrate that the characteristics of the correlative and blocking methods limit their usefulness in the determination of which feedback process controls rhythmic activity. Instead, this question requires computational tools such as the ones developed here. Finally, we point out in Discussion that hybrid systems such as the dynamic clamp may allow experimental use of our method.
We consider a mean field type model describing the activity of an excitatory neural network subject to both synaptic depression and cellular adaptation as described previously (Tabak et al., 2006). The variables of the model are a, the network activity (firing rate averaged across population and time; a = 0 corresponds to all cells silent, a = 1 means all cells fire at their maximal frequency); s, the fraction of undepressed synapses (s = 0 means all synapses are depressed, s = 1 means all synapses are operational); and θ, a cellular adaptation process that raises the neuronal firing threshold (θ = 0 means no adaptation so the cellular threshold is at its baseline level θ0, θ = 1 is the maximal adaptation). The model equations are:(1)(2)(3)where a∞ is an increasing sigmoidal network input/output function (Table 1). The two parameters w and θ0 set the global network excitability [8]. Connectivity (w) represents the amount of positive feedback due to excitatory connections, i.e., it determines the fraction of network output (activity) fed back as input to the network. The average cellular threshold (θ0) measures the cellular excitability, i.e., it biases the cells' responses to synaptic inputs.
In Eq (1) we see that synaptic depression, which decreases s, acts as a divisive factor, decreasing the amount of positive feedback, while cellular adaptation, which increases θ, is a subtractive factor. An additional parameter, g, can be adjusted to scale the strength of the adaptation process. Unless mentioned otherwise, g is set to 1. The steady state functions s∞ and θ∞ are decreasing and increasing sigmoidal functions of activity, respectively. Thus, when activity is high, s decreases and θ increases, both of which contribute to active phase termination. During the silent phase, s increases and θ decreases, eventually initiating a new active phase. The active phase is defined as the period of activity for which a is above an arbitrarily determined threshold (0.35). Below this threshold the system is in the silent phase.
The network recruitment time constant, τa, is arbitrarily set to 1 and the time constant for the variations of s and θ are assumed much larger than τa. That is, s and θ are slow processes. All parameter values are given in Table 1. Equations were solved numerically using the 4th order Runge Kutta method (dt = 0.05) in XPPAUT [10]. The simulation code is freely available on RB's website http://www.math.fsu.edu/~bertram/software/neuron.
To assess the contributions of slow divisive and subtractive feedback to episode onset and termination we first test two methods based on measurements and manipulations that can be performed experimentally. We use a mean field model of rhythmic activity in an excitatory neural network regulated by both synaptic depression and cellular adaptation, defined by Eqs. 1–3, to generate synthetic data. These data show the time courses of the network activity, a, and the two negative feedback processes, s and θ (Figure 1AB). When we ask what is the contribution of a process to the episodic activity, we ask two questions: what is its contribution to episode initiation, and what is its contribution to episode termination. To clarify the meaning of “contribution”, we see in Figure 1A or 1B that during an episode s decreases and θ increases. These effects decrease network excitability and eventually the activity cannot be sustained, so the high-activity episode stops. But which effect is more important in terminating an episode? Was it the decrease in s or the increase in θ? Can we quantify this notion? Similarly, during the silent phase both processes recover (i.e., s increases and θ decreases), until a new episode is initiated. Again, can we quantify the effects on episode initiation of the increase in s vs. the decrease in θ?
The rationale for this first approach is that if a process varies greatly during the high-activity episodes (active phases) and the inter-episode intervals (silent phases), then it is likely that it contributes significantly to episode termination and onset. On the other hand, if the variations are small, it is likely that the contribution of the process is small. This approach thus relies on observing a relationship between the time course of a process and the system's behavior. Its pitfall, that correlation does not imply causation, is well known.
Experimentally, one can record spontaneous or evoked postsynaptic potentials or currents in target neurons [11], [12], [13], [14]. The variations of this postsynaptic response during the interval of time between two episodes of activity would represent the variations of the effective connectivity, or available synaptic strength, w.s. Similarly, one may record the degree of adaptation or the current responsible for this adaptation at various times during the silent phase [11], [14]. The variations of the current with time would be equivalent to the variations of g.θ. Here we assume that there are only two slow feedback processes, represented by s and θ, which can be measured unequivocally and with sufficient precision. This is an ideal situation that will not often be encountered experimentally; we show that even with such ideal conditions we may not be able to determine the contributions of the two slow processes using the correlative approach.
If s varies by Δs and θ by Δθ over one phase of the oscillation, then according to the correlative approach the ratio (4)measures the contribution of s relative to that of θ. We have shown previously [8] that if s and θ vary exponentially with time constants τs and τθ, then Δs/Δθ ≈ τθ/τs . Thus, (5)
Assuming that w and g are similar – we set w and g to 1 unless noted otherwise – the ratio of the contributions of the two processes to the rhythmic activity is inversely proportional to the ratio of their time constants, so the faster process contributes more than the slower process. This is illustrated in Figure 1AB where we plot the variations of a (network activity), s and θ for the cases r = τθ/τs = 0.1 (A) and r = τθ/τs = 10 (B). In the case shown in Figure 1A, we expect s (red curve) to contribute more to episode onset/termination because it is the faster process, while in the case shown in Figure 1B θ (blue curve) is faster and thus expected to have the major contribution.
We define a quantitative measure of the contribution of the two processes by (6)(or, using the approximation given by Eq. 5, c = (r−1)/(r+1)). C varies between −1 and 1. If C is near 1 then s determines the episode onset and termination (i.e., θ has no role). If C≈−1 then θ controls episode onset and termination. Intermediate values of C indicate that both processes contribute. This measure is plotted as a function of r in Figure 1C, and clearly demonstrates the shift of control (according to the correlative definition) from θ to s as the s dynamics are made progressively faster relative to θ. The filled circles result from simulations with the cell excitability parameter θ0 set to 0 (relatively high cell excitability). The open circles were obtained using θ0 = 0.18 (low cell excitability). The differences are very small, showing that, according to this measure, the respective contributions of the two processes depend very weakly on θ0. The dashed curve in Figure 1C is obtained by plotting c = (r−1)/(r+1). Since the points obtained from plotting C lie almost on this curve, one concludes that, according to the correlative approach, the contributions of the two slow processes depend only on the ratio r = (w/g) (τθ/τs). Thus, the faster that one process is relative to the other the greater its contribution will be to rhythm generation. Similarly, the greater the relative weight of a process, the greater its contribution. Finally, since each process covers the same range during the active and silent phase, these results do not distinguish between episode initiation and termination. That is, the correlative approach predicts that the contribution of each process is the same for episode initiation and termination.
The rationale for this second approach is that if a process is important to a system's behavior, then removing it will have a large effect. This type of experiment is widely used in biology and includes pharmacological block, surgical ablation, and gene knockout. If, for example, θ represents the activation of a potassium current responsible for cellular adaptation, then one could block this current pharmacologically or genetically and measure the effect on network activity. We block the θ process by setting g = 0 and observe the effect on the length of both the active and silent phases after transient effects have died down. If we see a large increase in the active phase duration, then we conclude that this process is important in terminating the active phase. Similarly, if after the block we see a decrease in silent phase duration then we conclude that recovery of this process is important for episode initiation. The pitfall of this approach is that after blocking a process we obtain a different system.
Figure 2 illustrates the results obtained with this approach, for different values of the parameter θ0. Figure 2A shows the time course of network activity before and after blocking θ in the case τθ = τs. When cell excitability is too high (e.g., θ0 = 0.06), synaptic depression alone cannot bring the network to a low activity state and rhythmicity is lost after the block. For lower cell excitability (higher θ0, middle and right columns), blocking θ leads to changes in the lengths of both the active and silent phases, to various degrees. These changes in active and silent phase durations (AP and SP), after transient effects have died out, are represented on Figure 2B for different values of the ratio τθ/τs. Can we infer the importance of θ variations on rhythm generation from these changes?
We first note that for low θ0 rhythmic activity is lost after blocking θ, for all values of the ratio τθ/τs. Thus, variations in θ are required for rhythm generation in these cases. In the other cases shown, blocking θ has large effects on the active and silent phase durations, but these effects are difficult to interpret. For instance, we expect the block to increase the active phase in proportion to θ's contribution to episode termination. Thus, it seems that θ contributes significantly to episode termination in cases vi, viii and ix (where there is a large increase in AP after the block), but does not contribute much to episode termination in case iii (where there is no change in AP after the block). In cases ii and v the active phase duration actually decreases after the block, which is hard to interpret. Similarly, we expect the decrease in silent phase duration following θ block to be in accordance with θ's contribution to episode initiation, since residual adaptation delays episode onset. Thus, we would say that θ contributes significantly to episode onset in cases ii, iii, v, vi and viii. But again, we have an unexpected case (ix) where SP increases after the block.
The blockade experiment illustrated in Figure 2 suggests that there are more cases where θ has a significant contribution on episode initiation (ii, iii, v, vi, viii) than on episode termination (vi, viii, ix). This is in contradiction with the correlative approach, which suggested that θ had a similar contribution to both episode termination and initiation. There are also cases, such as vi and viii, where the effects of the block are similar, suggesting that θ's contribution to episodic behavior is similar in those cases. But cases vi and viii correspond to different values of the ratio τθ/τs. According to the correlative approach, the contribution of each process should vary with τθ/τs (Figure 1C), so again the blockage approach and correlative approach disagree. Finally, on each row of Figure 2B the effect of the blockage varies with the value of the parameter θ0. This again contradicts the correlative analysis, which showed little dependence on θ0.
The strong perturbation to the system effected by the block is responsible for the counterintuitive decrease in AP observed in cases ii, v and increase in SP observed for case ix. These changes reflect system compensation; after the block and after transients have died out, the unblocked process, s, covers a different range of values, so AP and SP are modified. This compensation could be avoided by measuring AP and SP just after the block instead of letting it equilibrate. This is illustrated in Figure 2Aii, where the block initially increases AP, then decreases it as SP is decreased by the absence of θ. Interpretation of the block experiment would therefore be facilitated by considering only transient behavior, but this would be difficult to do experimentally in most cases. For instance if we block a K+ channel pharmacologically then the kinetics of drug application and binding to the channels will interfere with the transient effects.
In summary, we find that the correlative and blockage approaches suggest different interpretations about the contributions of the negative feedback processes to rhythm generation. In the following, we show that neither approach gives a satisfactory description of the contributions of the slow processes. This is because each approach suffers its own pitfall. The first approach is purely correlative, i.e., it links variations in one process to the behavior of the system, but cannot establish causation. To obtain causation it is necessary to determine how the system responds to a perturbation to one of these processes, as in the blocking approach. Unfortunately, by perturbing the system, we change it. The loss of periodic activity after blocking θ (as in cases i, iv, vii in Figure 2) shows that this process may be necessary for maintaining rhythmic activity, but it does not indicate what was the contribution of θ before the block.
The goal here is to derive a measure that allows one to draw a causal link between each slow process and the activity pattern that does not involve a strong perturbation to the system. Suppose that s is the only negative feedback process regulating episodic activity, so it contributes 100% to both episode termination and initiation. Then doubling τs will (approximately) double both AP and SP. If s is not the only negative feedback process and therefore has only a partial contribution to episode termination and initiation, then doubling τs will still increase AP and SP but by a smaller factor. Thus, the contribution of s to the episodic activity can be determined by the fractional change in AP and SP durations following a change in τs. To illustrate this idea, we plot both AP and SP durations as either τs or τθ is varied in Figure 3A.
Figure 3Ai shows that AP varies more with τs than does SP. This suggests that s has more influence on episode termination than on episode initiation. The variations of AP and SP with τθ (Figure 3Aii) show the opposite trend, suggesting that θ has more influence on episode initiation than on episode termination. These trends are also illustrated by the variations of the duty cycle ( = AP/(AP+SP)) with τs and τθ (Figure 3B). The duty cycle increases with τs, but decreases with τθ. Finally, comparing Figures 3A i and ii, we observe that the variations of AP with τs and τθ are similar, suggesting that s and θ have comparable contributions on episode termination. On the other hand, SP varies more with τθ than with τs, suggesting that θ has a stronger influence on episode initiation than does s. This example suggests that the contributions made by the slow processes to the episodic activity can be determined by varying the time constants of the processes and observing the effects on AP and SP durations. We now use this idea to construct a quantitative measure of these contributions.
We first construct a measure of the contribution of s to episode termination, as illustrated in Figure 4. At the beginning of an episode, τs is increased by δτs. If s contributes to episode termination, slowing down s increases AP by δAP. We can quantify the contribution of s to episode termination by evaluating the ratio of the relative change in AP, δAP/AP, divided by the relative change in τs, δτs/τs. We thus define the normalized contribution of s to episode termination as (7)
If s has no influence on episode termination, slowing it down has no effect and δAP = 0. If s is the only process contributing to episode termination, then the active phase duration is the time it takes for s to decrease from its value at the beginning of an episode to its value at the transition between AP and SP. Since we consider relaxation oscillations, the transition time between active and silent states is negligible. Thus, a fractional change in τs leads to the same fractional change in AP (δAP/AP = δτs/τs) so that CsAP = 1. Therefore, CsAP has a value between 0 (s does not contribute to episode termination) and 1 (s is the only process contributing to episode termination). We quantify the contribution of s to episode initiation similarly using (8)
We define the contributions of θ to episode termination and initiation in a similar way:(9)(10)
These measures have the same motivation as the blockage experiment, but can be computed with small perturbations to the system. We use δτ/τ = 4% so the perturbation is small but nevertheless has a measurable effect. In addition, we look at the acute effect of the perturbation, i.e., we do not wait until the system equilibrates.
Figure 5A shows the contributions of s to episode termination (CsAP) and initiation (CsSP) as the ratio τθ/τs is varied, determined through numerical simulations as shown in Figure 4. CsAP increases as this ratio is increased, that is, s contributes more to episode termination as it becomes faster relative to θ. When s is much slower than θ, CsAP is close to 0. For s much faster than θ, CsAP is close to 1. When s and θ have similar speed CsAP is close to 0.5, suggesting that the divisive and subtractive feedback processes contribute equally to episode termination when their time constants are similar. This relationship between the contribution of feedback processes to episode termination and the ratio of their time constants is in agreement with the prediction from the correlative approach (Figure 1C). However, the contribution of s to the silent phase, CsSP, varies differently with τθ/τs. Although it increases with τθ/τs, this increase is so weak that CsSP is below 0.1 even if s is 10 times faster than θ. This consistently low CsSP suggests that regardless of the relative time constants of the two negative feedback processes, s never contributes significantly to episode onset, in sharp contrast with the prediction from the correlative approach.
Figure 5B shows that the contributions of θ to episode termination (CθAP) and initiation (CθSP) vary in the opposite way to CsAP and CsSP. If τs is much larger than τθ then s does not affect AP while θ strongly affects AP. As the ratio τθ/τs increases, the contribution of s to episode termination increases while the contribution of θ decreases, in such a way that the sum of the contributions of s and θ stays around 1 (CsAP + CθAP ≈ 1) as shown in Figure 5C. The effect of θ on SP is always strong, while the effect of s is weak, regardless of τθ/τs. The sum of the contributions of s and θ to episode initiation also stays around 1 (CsSP + CθSP≈1). Thus s and θ have complementary contributions to the episodic activity and our measure is self-consistent. The relationship CsxP + CθxP≈1 is a consequence of the fact that s and θ are the only processes controlling AP and SP. That is, if we increase both of their time constants by a factor k, then AP and SP both increase by the same factor k (Figure 3Aiii). This can be written, in the case of the active phase, as: AP(k τs, k τθ) = k AP(τs, τθ). Application of Euler's theorem for homogeneous functions yields: and, after dividing each side by AP, results in CsAP + CθAP = 1.
Since we are dealing with only two slow processes, we can combine the measures defined for s and θ (Figures 5A and 5B) into single measures by defining(11)(12)
With this definition, CAP and CSP vary between −1 to 1. A value close to −1 signifies that θ is the dominant process; a value close to 1 signifies that s is the dominant process; a value near 0 means that s and θ have similar contributions. These are plotted in Figure 5D as a function of τθ/τs. We see that CAP rises from −1 to near 1 as τθ/τs increases, indicating that θ dominates the AP when it varies more rapidly than s, and s dominates when it varies more rapidly than θ. This agrees with the result obtained with the correlative approach (dashed curve, c = (r−1)/(r+1)). In contrast, the SP is controlled by θ for the full range of τθ/τs; this was not predicted by the correlative approach.
The contribution measures defined above are meaningful only if specific conditions are satisfied. The most important condition is that each variable or process contributes to the same aspect of system behavior. For instance we cannot compare the contribution of a slow negative feedback process, such as our s or θ, which terminates an episode of activity, to the contribution of a fast negative feedback variable that could be responsible for fast cycling during the high activity phase. Second, the variables must vary monotonically during each phase of the activity. If not, then increasing their time constant may not increase the duration of a phase in a predictable way and the sum of the contributions of the variables to that phase may not equal 1.
We use a relaxation oscillator with a clear distinction between active and silent phases. The measure can be applied to other types of oscillations, as long as active and silent phases can be clearly distinguished. In more complex cases, it may be necessary to divide a period of activity into more than two phases. More generally, the method could be applied to non-oscillatory systems, for example to determine the contribution that different variables make to return the system to an equilibrium following a perturbation. Also, the measure is not limited to two negative feedback processes. We have chosen feedback processes of different types, subtractive and divisive, because we find the problem of disentangling their relative contributions to be quite challenging. This measure can be applied with feedback processes of the same type, as long as they contribute to the same behavior. We have used the method to compute the respective contributions of two subtractive feedback processes to burst generation and shown that the results can be used to predict the occurrence of phase-independent resetting [15]. Finally, we use a deterministic model. Noise would not qualitatively affect our measure, as long as it does not affect the mechanisms for the transitions between phases. If noise is part of the transition mechanism [16] our method cannot be applied as it is, since noise would also contribute to the transitions.
Since the measure requires a model of the system, the validity of its results depends on the validity of the model. Models may incorporate various degrees of realism, so it is important that the measure be robust to model details. For instance, if we add a fast variable to the relaxation oscillator model, so that fast oscillations (spikes) are produced during each active phase, the two slow negative feedback processes may still terminate episodes (bursts) like in the relaxation case. Thus, the relative contributions of each slow variable to burst onset and termination should not change qualitatively. We have demonstrated such robustness with a model of bursting in pancreatic islets [15].
We now evaluate how the parameters that control network excitability, w (network connectivity) and θ0 (average cellular threshold), affect the contributions of s and θ to rhythm generation. Variations of CAP and CSP with w are represented in Figure 6, for three different values of θ0 (and for τθ/τs = 1). Clearly, CAP increases with w, i.e., synaptic depression contributes more to episode termination when network connectivity is high. However, this is not true for episode initiation, as CSP is almost unaffected by w. There is in fact a slight tendency for CSP to increase at the lowest values of w, which is more visible if s is faster than θ (not shown). Changes in θ0 do not affect either CAP or CSP significantly. This is in agreement with the correlative approach, but in contrast to the results of the blockade experiment (Figure 2).
In summary, the ratio τθ/τs and connectivity w – but not θ0 – strongly affect CAP, while none of these have a significant effect on CSP. In general, both feedback processes s and θ play roles in the episode termination, but only θ controls episode initiation. The relative influence of s and θ to episode termination varies with parameter values. The correlative approach is roughly correct for predicting the contributions of the two processes to episode termination, but not to episode initiation. This approach makes a direct comparison between the time scales of the two processes, scaled by their relative strength (w and g), evaluating r = (w/g) (τθ/τs). But this ratio is not the ratio of the contributions of the two processes to episode initiation. In fact, we show below that the weighted time scales cannot be compared directly but must be rescaled, the correct ratio beingwhere the scaling factor ak is the activity level at the transitions between active and silent phases. At episode termination, ak≈1 so the correlative approach is approximately right. However, at episode onset ak≈0, so rrescaled ≈ 0, meaning that s does not contribute significantly unless r >> 1. Looking back at Eq. 1, it is evident that s generally has little effect when activity is low. Such a simple fact was not revealed using the correlative and blockade approaches, stressing again that these standard experimental approaches are not always useful for determining the contributions of different variables to rhythmic activity.
The analysis above suggests that the correlative approach can reasonably estimate the contribution of each process to episode termination, but misses the fact that s contributes little to episode onset (Figure 5D). Results from both the blockade simulations and the analysis above suggest that θ is more important for episode initiation than episode termination. However, we have seen that the blockade approach does not typically provide a good indication of the contribution of θ to the AP and SP durations (Figure 2B). To further demonstrate this, we plot in Figure 7 the variations of CAP and CSP with g (curves), the maximal “conductance” of the adaptation process θ, in four of the cases illustrated in Figure 2B (v, vi, viii, ix). The values of both CAP and CSP decrease as g is increased, indicating that the influence of θ in the control of the rhythm increases with g. As g decreases towards 0, both CAP and CSP increase toward 1 since s is the only slow process when g = 0. This is true for all four cases. However, CSP only increases noticeably when g approaches 0, illustrating again that the subtractive feedback process controls the silent phase in most cases.
Comparing Figure 7A–B, we see that the CAP curve is similar in both panels, as is the CSP curve. The bar plots show the effects of a blockade simulation, where g = 1 before the blockade and g = 0 afterwards. In Figure 7A the blockade results in a 50% reduction in the AP duration, while in Figure 7B there is a very large increase in the AP duration following blockade. Yet, according to the CAP curves the contribution of θ to the AP duration is nearly the same in both cases when g = 1 (green and yellow boxes). Similarly, CSP is similar in panels C and D for g = 1, yet the blockade results in decreased SP duration in C, but increased SP duration in D. Thus, the effects of the blockade on AP and SD durations do not provide much information on the respective contributions of the two processes before the blockade.
Next, we compare cases shown in Figure 7B and 7C. We notice that CAP differs between the two cases, showing that when g = 1 the s variable contributes significantly to episode termination in one case (Figure 7B) but not the other (Figure 7C). Yet, after blockade the changes in AP/SP (bar plots) are similar in both cases. Again, results from the blockade approach do not indicate what was the contribution of each process before the blockade.
For the mathematically simple system used in this work, we can use a geometrical argument to derive approximate formulas for CSP and CAP. If the system is two-dimensional with one slow process, s, the trajectory could be drawn in the a,s-phase plane and would follow the a-nullcline (except for fast jumps at the transitions between active and silent phase). For the three-dimensional system presented here, the trajectory in the three-dimensional a,s,θ-phase space follows the surface defined by da/dt = 0 [8]. We can project the three-dimensional trajectory and surface into the a,s-plane. This results in a two-dimensional trajectory that follows a dynamic a-nullcline (Figure 8A). The effect of the third variable (θ) in this two-dimensional representation is to move and deform the dynamic a-nullcline (the thin, black S-shaped curve in Figure 8A). Increasing θ moves the nullcline rightward.
At the end of the active phase, the trajectory falls from the high- to the low-activity state and the dynamic nullcline is at its rightmost position (thick, discontinuous, grey S-shaped curve on the right of the diagram). During the silent phase, s increases so the system's trajectory moves to the right while θ decreases so the a-nullcline is transformed leftward. When the trajectory passes the low knee (LK) of the nullcline, the trajectory jumps to the upper branch. At this point the nullcline has reached its leftmost position (the thick grey S-shaped curve on the left), since θ will now again begin to increase and the a-nullcline will be transformed rightward.
To compare the contributions of s and θ to the termination of the silent phase, we can therefore compare the length traveled by the trajectory (controlled by s) with the length traveled by the low knee (controlled by θ). Assuming that their speeds are nearly uniform, we can compare the instantaneous variation of the trajectory's position ds to the instantaneous variation of the knee dsk due to the variation of θ, dθ. We can show [8] that ds ≈ dθ (τθ/τs) and that dsk ≈ (g/w) (dθ/ak) where ak is the activity level at the knee (its value varies little with θ). Thus, the ratio of the contributions of s and θ is (13)
This formula applies to both active and silent phases, however the activity level at the knee, ak, differs between the two phases. During the silent phase, ak is close to 0 so ds/dsk is very small, i.e., s generally contributes little to the termination of the silent phase. On the other hand, during the active phase ak is close to 1, so ds/dsk ≈ (τθ/τs) (w/g). If (τθ/τs) (w/g) ≈1 then the two slow processes contribute similarly to active phase termination. This shows that the relative contributions of s and θ are qualitatively different for the different phases of activity. It also explains why the intuitive approach illustrated in Figure 1 is correct for the active phase (where ak≈1), since from Eq. 5 and Eq. 13 r ≈ ds/dsk. If (τθ/τs) (w/g) ≈ 1, then r ≈ 1 and the correlative approach predicts equal contributions of the feedback variables (Figure 1C). On the other hand, during the silent phase ak≈0 so r is not a good approximation to ds/dsk and the correlative approach is invalid.
To compute ds/dsk for both phases, we must compute ak (Eq. 13) for both knees of the dynamic a-nullcline shown Figure 8A. For this we note that the nullcline is defined by da/dt = 0. Solving for s, we obtain(14)
For each value of θ, the knees are defined by and differentiating Eq. 14 gives:(15)which has two solutions ak, each corresponding to a knee, provided the right hand side is greater than 2. The values of θ at onset and termination of the episodes, to be used in Eq 15, were obtained from the durations of the active and silent phases obtained from simulations [8]. Finally, when θ0 is changed there is a similar but opposite change in the range of variation of θ, so θ + θ0 is not affected much by a change in θ0. Thus the solutions of Eq. 15 are not very sensitive to θ0. This explains why the relative contributions of s and θ are little affected by θ0, as seen in Figure 6.
Since we identify ds/dsk to the ratio of the contribution of the two slow variables for each phase, CsxP/CθxP, the combined measures CxP defined in Eq 11–12 correspond to the ratios (ds/dsk − 1)/(ds/dsk + 1). These ratios are computed for both active and silent phases as a function of w and shown on Figure 8B. Comparison with Figure 6 (middle panel) shows that this geometric measure of the contributions of the slow processes is in good agreement with the empirical measure constructed above using sensitivities to the slow variables' time constants.
Finally, we point out that there are rare situations when the two measures (Eqs. 7–12 vs. Eq. 13) do not give similar results. Such a case is shown in Figure 8C, for which the parameter θ0 is large (average cell excitability is low) and τs is 10 times greater than τθ. Because θ0 is large, even when θ decreases to its minimum during the silent phase, s may not be sufficiently large for an episode to start, particularly if the connectivity is low. In that case, an episode is not started until s reaches the value corresponding to the low knee. Even if this is a small distance, it can take a long time since s is so slow. Thus, changing τs can have a strong effect on the silent phase and CSP determined from Eq. 12 becomes positive (Figure 8C, left panel) instead of close to -1 as computed using Eq. 13 (Figure 8C, right panel). In other words, using a measure based on time indicates a strong contribution of s in that particular situation, while a measure based on geometry indicates a marginal contribution of s to episode initiation. This discrepancy between the two measures appears because θ does not vary uniformly. It slows down considerably as it approaches its asymptotic value, “waiting” for s to reach the low knee. Thus the dynamics of s now play a major role in terminating the silent phase. Note that θ still has a strong effect on the s dynamics during the silent phase (it determines the location of the low knee of the a-nullcline in Figure 8A), but θ's dynamics do not affect the silent phase duration much, so the measure that relies on perturbing the time constants finds it has little contribution.
Biological systems are characterized by the interactions between many components. Often, several processes contribute to regulate the same behavior. The purpose of this work was to develop a method for determining how two different negative feedback processes contribute to the generation of relaxation oscillations in biological systems such as excitatory neuronal networks. We gave a precise meaning to the contribution of a given process to episodic activity in an excitatory network regulated by two activity-dependent negative feedback processes. Namely, a process contributes significantly to the termination of a phase (active our silent) of the activity if an acute change to its time constant at the beginning of the phase significantly lengthens that phase. To illustrate this concept we have used a mean field model of an excitatory neuronal network in a relaxation oscillation regime, regulated by two types of negative feedback, divisive (synaptic depression) and subtractive (cellular adaptation). The measure developed here shows that there is differential control of the two phases by the two feedback processes. Both divisive and subtractive feedback processes contribute similarly to episode termination, as long as their time constants and strengths (i.e., associated conductance) are in the same range. In contrast, only the subtractive feedback process contributes significantly to episode initiation in most cases. This difference in the control of the active and silent phases arises from the very nature of the divisive feedback: acting as a multiplicative factor to the activity level, its influence is much lower during the silent phase when activity is low. Thus during the silent phase the dynamics of the subtractive process play a larger role.
We have first attempted to use approaches inspired from experimental methodology to determine the relative contributions of the two feedback processes to rhythm generation. These included comparison of the time course of each process (the correlative approach) and blocking one of the processes.
The correlative approach simply compares the amount of variation of each process, scaled by each process' strength or conductance. Since the two processes vary by the same amount during the active and silent phase, this approach does not distinguish between active and silent phase. According to this approach, the relative contribution depends only on the ratio of their time constants (τθ/τs) and on the ratio of their strength (w/g). It predicts that if these two ratios are close to 1 then both feedback processes contribute similarly to the rhythm. In the example shown here this is a good approximation for the active phase. However, for the silent phase, this intuitive rule fails, because an additional scaling factor must be introduced to compare the contributions of the two different negative feedback types. This scaling factor is significantly different from unity for the silent phase; it reflects the fact that the divisive feedback process, being a multiplicative factor to the activity, has very little effect at low activity (i.e., during the silent phase).
The blockade approach suggests that the subtractive process might be more important in setting the silent phase duration, since blocking this process affected the silent phase duration more often than the active phase duration. In this way it provides a piece of information that is missed by the correlative approach. However, similar effects of the blockade on AP and SP durations were found in cases where the ratio of time constants was different (and different effects when that ratio was identical), contradicting the correlative approach and, as shown in Figure 7, contradicting our measure of the relative contributions of each process. Furthermore, unlike the correlative approach, the blockade experiment suggests a strong effect of θ0 (which biases the input/output relationship of the system). In general, however, this parameter has little effect on the contribution of each process (cf. Figure 6).
These disappointing results from the two experimental approaches are due to their well known pitfalls: passive observation only establishes an association without proving a causal relationship, while perturbations to the system, such as blockade experiments, can qualitatively change the system being studied. The use of total blockade may be considered extreme. A partial block can potentially be more informative than a complete block because a small enough perturbation may indicate a trend in a component's influence and preclude switching the system to a different mode of operation (see e.g., [17], [18], [19]). In other words, if the perturbation is small enough the effect on the activity may be close to linear so the effect of the partial block can be quantified and provide information on the role of the process that is partially blocked. However, partial blockade cannot provide a quantitative measure with the properties (summation to 1) of the C values developed here.
Our approach, instead, is to use small perturbations to the time constants of the feedback processes and look at the effect immediately following the perturbations. This minimizes the perturbation to the system, while quantifying the relative contribution of the two slow processes to the rhythmic behavior. This method could be applied to many oscillatory systems that rely on the interplay between positive feedback and several negative feedback processes. However, for most known experimental conditions, this method seems impossible to implement. To apply the method requires 1) the ability to change the time constants of the variables of interest one by one, 2) these changes must remain small but have measurable effects and 3) the system's behavior immediately after the changes must be measured, without waiting for transients to die out. For example, in the context of a neural network, there is currently no technique available to change the time constant of synaptic depression by a small amount, quickly and without affecting other network parameters. Thus, in many cases, the question of determining the contributions of different negative feedback processes in rhythm generation (using our approach) may only be addressed with computational models.
One example in which our approach could be used in an experimental setting is the electrical oscillatory activity of single cells. The mathematical formalism used to describe the mean activity of an excitatory network is similar to the Hodgkin-Huxley formalism commonly used to describe the electrical activity of excitable cells [8], [20], [21], [22]. In excitable cells, the sodium or calcium channels generate voltage-dependent inward current, providing fast positive feedback that increases membrane potential, while the delayed activation of outward potassium currents and inactivation of the inward currents provide negative feedback. An outward current has an opposite influence to the excitatory inward current and therefore provides subtractive feedback; on the other hand the inactivation of an inward current is a multiplicative term reducing the amount of positive feedback and therefore is a divisive feedback process. Preliminary results with the Hodgkin-Huxley model of nerve excitability [20] in a repetitive spiking mode suggest that while both sodium current inactivation and potassium (K+) current activation contribute to terminating an action potential, it is mostly the de-activation of the K+ current that initiates the next spike (J. Tabak, unpublished results). This could be verified experimentally for electrically compact cells using the dynamic clamp technique, which allows one to introduce a model-generated ionic current into a cell [23], [24]. For example, one could pharmacologically block the Na+ current, then re-introduce it into the cell using the dynamic clamp. Because the added current is computed from a model, it would be possible to change its inactivation time constant by a desired amount and measure the effect of this perturbation on the duration of the spike or interspike interval. To our knowledge, a similar experiment has been done only once, to show that increasing the inactivation time constant of a low-voltage-activated calcium current would result in longer bursts in invertebrate neurons [25]. While both divisive and subtractive feedback can in principle terminate bursts in neurons [26] it is usually the latter that is considered to regulate bursting, in the form of slow, calcium-activated K+ currents. The experiment described in [25] provided strong support for a role of low-voltage-activated calcium current inactivation (divisive feedback) in burst termination.
Modeling is being established as an essential tool for understanding complex biological systems [27], complementing experimental approaches. But more than mere simulations of systems of differential equations, which are akin to experiments, it is the qualitative analysis of the models that provides new insights into a system's dynamics. Qualitative model analysis techniques include phase plane and bifurcation analysis, but these techniques become more difficult to apply as the number of variables increases. The commonly used fast-slow analysis, which simplifies model analysis by formally separating the equations into fast and slow subsystems, may have limited usefulness when many variables operate on the same time scale.
An extension of fast-slow analysis that can deal with many variables operating on the same time scale is the Dominant Scale Method (DMS) [28]. This method follows one variable of interest along an oscillatory trajectory (for instance, voltage in a cellular oscillator model) and determines the sensitivity of this variable at each point on its trajectory to other variables that are present in its differential equation. During different epochs of time, only a few variables may significantly affect the primary variable, so the model can be reduced to a few variables during each epoch. Thus, a complex model is transformed into a sequence of simpler models using only the dominant variables, and qualitative analysis of the dynamics is possible for each successive epoch [29]. The DMS can evaluate the relative contributions of variables that have different roles, unlike the measure presented here. However, our approach uses the sensitivity of observable features of the system behavior (AP and SP), not the sensitivity of a variable to other variables. For this reason, one may use our approach to identify cases where a variable has very little effect on the primary variable but nevertheless controls the duration of a given phase of the activity (as discussed in last section of Results).
Our approach to measure the contribution of feedback processes to rhythmic behavior is to compute the sensitivity of the AP and SP to the time constants for these processes. Other techniques that use sensitivities of observables of a system to control parameters are Metabolic Control Analysis and Biochemical Systems Theory [30], [31], which have been used to analyze metabolic and gene regulatory networks. Important features of these approaches include summation theorems, for instance the sum of the sensitivities of the level of a metabolite to control coefficients is equal to 1. A similar summation theorem holds in our analysis, where the contributions of the two slow variables to the AP or SP duration sum to 1. These techniques are usually applied to the control of steady states, but they have also been used to describe how observables such as the period and amplitude of an oscillatory system are regulated by control parameters [32], [33]. The control of these observables is usually distributed across control parameters [33]. Here, we found that the control of the active phase is distributed across the divisive and subtractive feedback processes, but control of the silent phase is mostly operated by the subtractive process, θ. That is, θ is the “rate limiting factor” in the termination of the silent phase.
Finally we mention parameter search techniques, which are usually developed to find parameter sets that lead to a target behavior. These techniques can also be used to determine what parameter changes must be done to qualitatively affect a system's activity and provide information about the robustness of such activity [34]. Furthermore, by finding different parameter sets that produce similar system behavior, it is possible to determine the relationships between parameters that allow a behavior to be maintained [35] or to evaluate how each model parameter influence a given characteristic of the behavior using nonlinear regression [36]. This “database approach” indirectly provides information about the role played by some variables of the system and how a variable can take over when another variable is eliminated. It can be used to explore the behavior of a model in different regions of parameter space [37]. An intriguing observation is that different parameter combinations in a wide area of parameter space may produce similar oscillatory patterns [38]. If two distinct parameter sets produce the same system behavior, does this mean that a variable might have different roles in different networks that produce similar activity? This question could be answered with a combination of the database approach and the analysis technique developed here.
We have developed a computational method to quantify the relative contributions of feedback processes to active and silent phases of episodic activity. We have considered a case involving both subtractive and divisive processes. If both processes have similar strength and time scales, they contribute equally to terminate the active phase. This is consistent with our intuition and predicted by the correlative approach. Interestingly, it is the recovery from the subtractive process that sets the duration of the silent phase. This is because the divisive feedback is a multiplicative factor to the system's activity and therefore plays little role during the silent phase. Thus, different phases of the activity are controlled differently by the negative feedback processes. Experimental methodologies do not in general provide this type of information, so the determination of the relative contributions of different variables to a biological system's activity will usually require the development of a computational model. The method presented here can be applied to a wide array of oscillatory systems.
|
10.1371/journal.pcbi.1003087 | Combinatorial Clustering of Residue Position Subsets Predicts Inhibitor Affinity across the Human Kinome | The protein kinases are a large family of enzymes that play fundamental roles in propagating signals within the cell. Because of the high degree of binding site similarity shared among protein kinases, designing drug compounds with high specificity among the kinases has proven difficult. However, computational approaches to comparing the 3-dimensional geometry and physicochemical properties of key binding site residue positions have been shown to be informative of inhibitor selectivity. The Combinatorial Clustering Of Residue Position Subsets (ccorps) method, introduced here, provides a semi-supervised learning approach for identifying structural features that are correlated with a given set of annotation labels. Here, ccorps is applied to the problem of identifying structural features of the kinase atp binding site that are informative of inhibitor binding. ccorps is demonstrated to make perfect or near-perfect predictions for the binding affinity profile of 8 of the 38 kinase inhibitors studied, while only having overall poor predictive ability for 1 of the 38 compounds. Additionally, ccorps is shown to identify shared structural features across phylogenetically diverse groups of kinases that are correlated with binding affinity for particular inhibitors; such instances of structural similarity among phylogenetically diverse kinases are also shown to not be rare among kinases. Finally, these function-specific structural features may serve as potential starting points for the development of highly specific kinase inhibitors.
| The kinases are a group of essential signaling proteins within the cell and are the largest family of enzymes encoded by the human genome. The high degree of binding site similarity shared across the protein kinases has made them difficult targets for which to design highly selective inhibitors, but kinome-wide binding site analysis can help predict unintended off-target inhibitions. Given the increasingly large number of available kinase structures, kinome-wide comparative analysis of binding sites is now possible. In this paper, the Combinatorial Clustering Of Residue Position Subsets (ccorps) method is introduced and used to synthesize kinome-wide structure datasets with a kinome-wide inhibitor affinity screening dataset consisting of 38 kinase inhibitors. ccorps identifies structural features of the kinase binding site that are correlated with an inhibitor binding and uses these features to predict if this inhibitor will be capable of binding to uncharacterized kinases. This paper demonstrates the ability of ccorps to accurately predict inhibitor binding and identify features of the kinase binding site that are unique to kinases capable of binding a given inhibitor.
| The protein kinases constitute the largest enzyme family encoded by the human genome, with currently 518 known sequences, making up 1.7% of all human genes [1], [2]. Because these protein kinases are intimately involved in cellular communication and regulation networks, the loss of normal kinase regulation has been implicated in a wide variety of pathological conditions. The large number of disease states found to be associated with kinase dysregulation has motivated the development of kinase-specific inhibitor compounds and research to discover protein kinase inhibitors has come to constitute 20–30% of the drug development programs at many companies [1].
The bulk of this effort has been directed at identifying inhibitors that bind at the atp binding site. However, due to the large number of existing protein kinase domains and the high degree of (atp) binding site similarity among them, designing highly selective inhibitors has proven difficult. For example, type I kinase inhibitors that only target the atp site have typically been found to have low selectivity across the kinome [3]. To increase inhibitor selectivity, type II inhibitors bind both the atp site and the immediately adjacent allosteric site. By also binding to the allosteric site, type II inhibitors are able to make additional highly specific interactions, thereby allowing them to be more selective [3].
Identifying highly specific structural features that can be uniquely targeted by inhibitors can be facilitated by comparative analysis of multiple kinase structures [4]. Comparative analysis of multiple structures allows for the identification of kinase structural features that are available for inhibitor targeting as well as insight into the effect of activation conformation dynamics, such as structural features that are only available for targeting in the inactive, DFG-out conformation [3]–[6]. Furthermore, combining structure and sequence is important when analyzing the kinases holistically due to the large degree of sequence divergence among the protein kinases [7]. A specific example of the insight derived from the comparative analysis of kinase structural features follows.
Many of the effective inhibitor selectivity strategies involve exploiting the differences in the size of the atp binding site and targeting residue variability at a few key positions [3], [8]. These structure-based comparison approaches have proven more useful than sequence-only measures of overall kinase similarity in evaluating the potential selectivity profile of inhibitors [8]. For example, the size of the gatekeeper residue directly moderates the availability of a hydrophobic pocket. Inhibitors having larger functional groups that bind this hydrophobic pocket may be able to select for the roughly 20% of protein kinases that have a relatively small gatekeeper residue (e.g., Gly, Val, Ala or Thr). This is because kinases with a larger gatekeeper residue (e.g., Phe) do not have a large enough hydrophobic pocket to accommodate the inhibitor [8]. However, in order to select for an even more specific subset of the human kinome, it has proven necessary to take advantage of multiple structural features of the kinase binding site (both atp and allosteric sites) simultaneously [3], [8].
A review of related work is given below. Recent work has illustrated that local structural similarity exists among phylogenetically diverse groups of kinases [5], [9] and has highlighted the importance of large-scale, multiple-structure analysis of structure-affinity relationships among the kinases [9], [10].
The PharmMap method [10] incorporates an aligned set of receptor-ligand co-crystals in order to identify pharmacophores common to a set of inhibitors. It has been developed to identify kinase inhibitor pharmacophores useful for selecting molecules for kinase screening panels.
Huang et al. have utilized a knowledge-based approach to constructing a minimal binding site “fingerprint” that captures only a pre-specified set of well-studied, structurally selective features, such as the size and hydrogen-bonding ability of the gatekeeper residue [8]. The per-kinase fingerprint utilizes nine binding site features (e.g., residue type at gatekeeper position) that have been shown to encode for selectivity among type I inhibitors. Anecdotally, kinases with similar fingerprints were shown to also have similar inhibitor selectivity profiles [8], illustrating the utility of structural features in predicting and understanding kinase selectivity.
Rather than relying upon pre-specified structural features, the recently developed Pocketfeature method decomposes a binding site into all possible “micro-environments” [11]. Pairs of kinase binding sites with highly similar sets of micro-environments were anecdotally shown to share a common inhibitor in 9 out of the top 50 most similar (as calculated by Pocketfeature) kinase binding site pairs. The CavBase [12] cavity matching approach has been used to cluster kinase atp binding cavities from multiple families across the kinome [5]. The kinase binding cavity clusterings derived from CavBase have been shown to generally agree with the sequence-derived kinase families and sub-families [5], demonstrating that the overall kinase cavity is well-conserved within families.
Recent work by Jackson et al. demonstrated a related structural clustering approach to predicting kinase inhibitor binding affinities [9]. Their geometric hashing approach to whole-site comparison of the atp binding pocket was demonstrated to be effective at identifying possible instances of inhibitor cross-reactivity and further emphasized the importance of taking into account subtle conformational changes in the binding site.
However, despite the successes of existing approaches, several outstanding problems to identifying structural features of the kinase binding site that are predictive of inhibitor selectivity remain. The reliance upon a single, representative structure precludes the ability of existing methods to identify features common only to active conformations if an inactive structure is chosen as representative (and vice versa). Additionally, choosing one representative structure disregards the role that binding site flexibility and plasticity may play in inhibitor interactions. Furthermore, the availability of multiple structures for individual kinases, exhibiting a variety of binding site conformations and bound ligands, provides a vast quantity of structure data that remains unexploited by existing approaches. Much of the difficulty in incorporating multiple conformations per individual kinase sequence into existing analyses stems from the non-uniform distribution of available kinase structures, with kinases such as CDK2 having more than a hundred available crystallographic structures while other kinases have only a single (or no) available structure. Finally, the availability of multiple kinase structures known to bind a given inhibitor and other kinase binding sites known not to bind that same inhibitor provides a rich set of structural examples and counter-examples beyond a single instance of pairwise similarity. Existing receptor-based methods focus on identifying meaningful pairwise similarity to a characterized kinase known to bind a given inhibitor. These methods currently do not account for the similarity of a given kinase binding site to other kinase sites that have been characterized to not bind the inhibitor in question.
To this end we have developed the Combinatorial Clustering Of Residue Position Subsets (ccorps) method. ccorps solves the following problem: given a set of sequence-aligned kinase domains (each having available PDB structures) and a per-sequence inhibitor binding label (either binds, does-not-bind or unknown), predict whether a given kinase domain binds the given inhibitor. Taking a set of kinase binding site residue positions as input, ccorps identifies clusters of kinases that share structurally and chemically similar subsets of residue positions. Given a particular kinase with unknown ability to bind a particular inhibitor, ccorps identifies kinase binding sites that share similar residue positions that are both known to bind and not to bind the inhibitor (i.e., it finds evidence both for and against binding a particular inhibitor). Finally, ccorps aggregates the residue position subset similarities for all possible -position subsets of the kinase binding site and predicts whether or not the given inhibitor will bind the given uncharacterized kinase binding site.
In addition to the development of ccorps, three major results from the analysis of the human kinome are presented here. First, the identification of structural features within the kinase atp binding site that are correlated with the ability of certain kinases to bind specific inhibitors is demonstrated. Second, the existence of affinity-correlated structural features that are shared among kinases from distinct families of the kinome are enumerated, shown to be not rare and also to differ depending upon the inhibitor being analyzed. Third, the ability of ccorps to predict the affinity for kinases lacking affinity annotations is quantified and compared to a recent structural binding site analysis approach [9].
ccorps is demonstrated to make perfect or near-perfect predictions for the binding ability of 8 of the 38 kinase inhibitors studied, while only having overall poor predictive ability for 1 of the 38 compounds. The performance of ccorps for predicting inhibitor binding is compared to the method of Jackson et al. [9] and shown to meet or exceed the predictive ability for the subset of the 38 inhibitors also tested by Jackson et al. We also compare ccorps against a sequence-based approach and show that they have complementary strengths. Finally, ccorps is shown to identify shared structural features across phylogenetically diverse groups of kinases that are correlated with binding affinity for particular inhibitors; such instances of structural similarity among phylogenetically diverse kinases are also shown to not be rare among kinases. These function-specific structural features may serve as potential starting points for the development of highly specific kinase inhibitors and provide a basis for understanding patterns of inhibition by compounds such as sunitinib that target multiple kinases [13].
In contrast to existing pairwise binding site comparison approaches, ccorps provides an automated way to incorporate the similarity of an uncharacterized binding site to all characterized binding site structures, without the need to manually select a reference binding site. ccorps also accounts for the similarity of an uncharacterized binding site to both kinases that bind and those that do not bind a particular inhibitor.
The high degree of atp binding site similarity shared across the protein kinases has made them a difficult target for which to design highly selective inhibitors. However, by identifying the patterns of local structural similarity among binding sites at the kinome scale, potential off-target interactions may be identifiable at earlier stages of pharmaceutical development and compensated for by further inhibitor modification. This would allow researchers to make predictions of binding affinity for a given ligand across the kinome with less experimental data. Furthermore, the emergence of kinase inhibitor resistance due to binding site position mutations may be better understood through the identification of kinases having similar structural features at the mutated positions. Structural features that are found to be unique to one or a small number of chosen kinases may provide the starting point for designing highly specific inhibitor interactions and therefore highly selective protein kinase inhibitors.
The Combinatorial Clustering Of Residue Position Subsets (ccorps) method is designed to solve a very general semi-supervised learning problem:
While in the Results section we focus on the specific problem of predicting ligand binding affinity across the human kinome, we will first describe ccorps in its most general form. To solve the general semi-supervised learning problem stated above, ccorps requires the following interface:
Input: an aligned set of protein substructures, where a substructure is defined as a collection of residues not necessarily contiguous in sequence but grouped together in 3D.
Input: annotation labels for some of the protein substructures.
Output: predicted annotation labels for the unlabeled protein substructures.
The per-substructure annotation labels may derived from a wide range of sources [14]–[25]. While this paper focuses on the application of ccorps for the prediction of inhibitor binding affinity annotations for the human kinome, ccorps generalizes to a variety of annotation prediction problems. The ability of ccorps to also identify specificity-determining enzymatic substructures for the prediction of ec class annotations for 48 different protein families is outlined in Text S2 and summarized at the end of this section.
In order to identify locally similar features among substructures, all -sized combinations of the residue positions (i.e., combinations) are generated. For example, given and , all 3-position subsets (1140 subsets) are generated. Then, each of these position subsets are examined one-by-one. Continuing the example, given the position subset , all the protein structures are compared by examining the pairwise similarity of only positions 7, 13, and 14 in isolation (i.e., disregarding the other 17 positions). 3-position subsets are used in this work because they allow for a unique 3-dimensional lrmsd superposition and are more computationally tractable than subsets of size , while still allowing for binding site position decomposition.
The dissimilarity between a pair of substructures is quantified by a combination of structural distance and chemical feature dissimilarity introduced in [26]. Specifically, the distance between any two substructures and is expressed as:The term is the pairwise-aligned side chain centroid lrmsd between the substructures. The remaining terms account for differences in the amino acid properties between the substructures and as quantified by the pharmacophore feature dissimilarity matrix as defined in [26].
For a given set of residue positions, we can calculate a matrix of pairwise distances between substructures using the distance measure defined above. Each row can be thought of as a feature vector that represents how different a protein is with respect to all other proteins in terms of the selected residues. The distance matrix is highly redundant. We use Principal Component Analysis (pca) to obtain a low-dimensional embedding. Our previous work [27] showed that this dimensionality reduction typically results in negligible information loss. Some technical details on how we correct for overrepresentation are described in Text S3. The dimensionality-reduced feature vectors are then clustered to identify sub-groups that share strong structural similarity. The number of clusters is not known beforehand and the number of clusters will vary depending on the set of positions being compared. The Gaussian Mixture Model (gmm) clustering method implemented in the mclust package [28] was used to identify both the number of clusters present and the cluster memberships for each of the feature vectors.
The above feature vector computation and clustering steps are repeated for each possible 3-position subset in order to compare all possible local structural features across all proteins. Structural variation in most subsets is not expected to be informative, either because no significant variation is present, or because spurious patterns can occur due to chance. However, functionally relevant structural variation can be detected with many different subsets and therefore distinguished from random patterns, as will be shown below.
A cluster that is dominated by one annotation label can be used to predict the label for other structures in that cluster whose annotation is unknown. We therefore call such clusters “highly predictive” (HPCs). Identification of HPCs is performed by selecting a minimum threshold for the label purity of clusters, and then selecting all clusters with equal or greater label purity than this minimum as HPCs; we used the strictest purity threshold possible (1.0 or 100% purity) in this work (see Fig. 1). In general, purity is calculated for a multiset of labels, , as where is the multiplicity function of a label within the multiset and is the most frequent label within . As with the dimensionality reduction mentioned above, we need to correct for overrepresentation bias, the details of which are described in Text S3. Purity alone does not account for the distinctness of the proteins in the cluster relative to the remainder of the dataset. For example, an hpc for label that partially overlaps a second hpc for label is less likely to be informative than an cluster greatly separated from the remainder of the dataset. The “degree of separation” or “distinctness” of a cluster was quantified by calculating the cluster silhouette score [29]. The mean silhouette score for a cluster was then used as a further selection criteria for identifying HPCs by removing potential HPCs with negative average silhouette scores (malformed clusters).
Each time an unlabeled protein falls within an hpc, that protein receives a single vote in favor of the majority label associated with the hpc. Because a protein can be a member of at most one cluster per -position subset, the maximum number of votes any protein can receive is equal to the number of possible -position subsets. For any given -position subset, it is possible that all clusters are HPCs or that no clusters are HPCs, depending on how the labels are distributed among the clusters. It is also possible that a protein may never fall within any hpc and therefore would receive zero votes for any label; such proteins are excluded from further analysis after the voting step. In the experiments described below this case rarely occurred. After tallying the label votes across all -position subsets, the label predicted for a given structure is determined by an SVM-derived decision boundary as described below.
Given a set of label votes that have been determined for an unlabeled structure, the threshold(s) used to decide which of the two or more label classes to assign to the structure requires the definition of a decision boundary procedure. For example, given a set of annotation labels containing the label classes (e.g., indicating whether a kinase is known to bind to a given ligand), a simple decision rule may be that given a structure with true vote, predict the true label for that structure. However, determining a single threshold for deciding the number of label votes required to classify a structure into one of several classes is difficult to generalize.
Because ccorps is a semi-supervised approach, the labels for the training structures are known and can be used to empirically estimate a vote count decision boundary. For example, given structure with known label, the number of times that appeared in a false hpc or a true hpc, across all -position subsets, can be calculated using the same approach as for unlabeled structures. The structure is then represented by an -dimensional vote vector, where each of the dimensions corresponds to the number of votes received for label ( for the case of kinase binding affinity, since we only have false and true labels). Application of this procedure to all labeled structures in the dataset provides an empirical basis for calculating a decision boundary in the vote space given the vote distribution for labeled structures. For example, the blue and red points shown in the scatter plot of Fig. 2 denote the vote vectors for training set substructures with known true and false labels, respectively.
Given the vote vectors calculated for all labeled training set substructures in the dataset, it is then possible to train any number of classifiers in order to determine a decision boundary. To compute a decision boundary in the vote space for classifying unlabeled proteins, svms were selected in this paper. First, an svm (linear kernel) is trained using the vote vectors of labeled training set substructures. For example, the decision boundary determined by training an svm on vote vectors is shown in Fig. 2 as the bold, black line. Next, for an unlabeled substructure with a given vote vector, the label for the substructure can be predicted by determining which side of the svm decision hyperplane the unlabeled structure falls within. As illustrated in Fig. 2, test vote vectors falling within the blue region will be predicted as having the true label and those falling within the red region, the false label. For training svms and calculating the -values of predictions made by those svms, libSVM [30] was used in this work.
To validate the predictive ability of the structural features identified by ccorps an extensive dataset of 48 families was automatically constructed using the Pfam database [17] as a source of well-curated protein alignments. The annotation labels analyzed in the validation set were per-structure Enzyme Commission (EC) number classifications. Cross-validation was performed in order to evaluate the predictive power of ccorps and the utility of the distinguishing structural features identified. The overall classification accuracy of ccorps (Text S2) when applied to the validation dataset demonstrates the ability of ccorps to identify structural features that distinguish functionally different protein homologs and the ability of ccorps to generalize to non-kinase protein families.
First, we will introduce the components of the kinome structure and affinity datasets used in this work. Next, structural features of the kinase binding site that are identified by ccorps to be predictive of inhibitor binding ability are presented. Then, cases of these predictive structural features that are common to phylogenetically diverse sets of kinases are highlighted. Finally, the performance of ccorps for predicting the binding ability of inhibitors across the kinome is quantified and compared to the related approach of Jackson et al. [9] as well as a sequence-based approach.
In order to enable the kinome-scale analysis of the protein kinase atp binding site presented here, a dataset of protein kinase binding site structures was assembled and then mapped to the affinity dataset of Karaman et al. [31]. Karaman et al. studied the affinity of 38 kinase inhibitors across 317 kinases and was one of the most comprehensive studies of kinase inhibitor selectivity at that time. Mapping a structure-affinity-phylogeny dataset by further incorporating the kinome family labeling of Manning et al. [2] has enabled the incorporation of all available crystallographic structures of the atp binding site and the analysis of shared structural features between major kinase families that is presented later in this paper.
The process by which ccorps recognizes structural features that are associated with kinase binding affinity is through the identification of Highly Predictive Clusters (HPCs). Given the 27-position binding site (Fig. 3), ccorps computes a clustering for each of the unique 3-position subsets. For example, consider the 3-residue substructure shown in Fig. 4A. The 3 residues shown correspond to 3 positions in the full kinome alignment and the corresponding residues for each structure in the kinome dataset are structurally compared to compute the substructure clustering shown in Fig. 4B. Each of the 1958 substructures within the kinase structure dataset is shown in Fig. 4B as a single point. The color of each point in Fig. 4B corresponds to the cluster assignment as computed by ccorps.
Several informative observations regarding kinase structural diversity and its association to inhibitor binding affinity can be made by further examination of the substructure clustering shown in Fig. 4B. Immediately upon examination of the substructure clustering it can be noted that multiple distinct clusters of kinases exist. This observation alone indicates that the 3-position substructure that resulted in this clustering is highly diverse among kinase binding sites. Conversely, the presence of a single large cluster would indicate that the 3-position substructure was structurally conserved, exhibiting little variance across the kinome; indeed instances of clusterings with a single dominating cluster were also observed for some 3-position subsets. As demonstrated in Fig. 4C, where one randomly selected representative substructure is shown for each of the 21 clusters identified by ccorps, both the geometry and residue types vary significantly for this 3-position subset.
By incorporating the affinity annotation labels for a particular inhibitor, further observations can be made about the association between the 3-position substructure shown in Fig. 4A and the kinases capable of binding that inhibitor. For example, mapping the affinity annotation labels for the inhibitor flavopiridol onto the substructure clustering (Fig. 4D) reveals that some of the clusters consist of only a single annotation label while others are a mixture of labels. In Fig. 4D, kinases capable of binding flavopiridol are colored red (true label), kinases incapable of binding flavopiridol are colored black (false label) and kinases lacking affinity annotation are colored white (undefined label). As shown in Fig. 4D, multiple clusters of purely true labels exist and are considered to be HPCs by ccorps.
The existence of true-only clusters indicates that the 3-positions shown in Fig. 4A are a distinguishing structural feature for identifying kinases that bind flavopiridol. More interestingly, however, is the fact that multiple, structurally distinct versions of the same 3-position substructure exist for different kinases that all are capable of binding flavopiridol. This result is significant because it indicates that across the kinome there are different structural motifs that are associated with binding flavopiridol, as opposed to a single, shared structural motif across all flavopiridol-binding kinases. The ability to identify multiple structural motifs that can each be associated with inhibitor binding is a strength of ccorps.
Furthermore, the existence of clusters containing only kinases incapable of binding flavopiridol can also be observed in Fig. 4D. These HPCs are also informative because they identify particular structural versions of the 3-position substructure in Fig. 4A that are all incapable of binding flavopiridol. Finally, clusters consisting of a mixture of kinases that are both capable and incapable of binding flavopiridol can be identified in Fig. 4D. For kinases in these clusters, the 3-position substructure is not a distinguishing feature of flavopiridol-binding ability.
Finally, while flavopiridol is discussed in detail here for illustration, the same analysis was computed by ccorps for each of the 38 different inhibitors within the affinity dataset. For each of the inhibitors, the affinity labels can be mapped separately onto the same substructure clustering as shown in Fig. 5. However, it should be noted that no information is shared between the results for different inhibitors in this work; that is, each inhibitor is computed in a fully separate ccorps computation (the substructure clusterings do not vary, just the annotation labels).
Examination of the affinity-annotated substructure clusterings shown in Fig. 5 reveals that the set of clusters which are HPCs varies greatly depending on the inhibitor considered. While the flavopiridol-annotated substructure clustering contains multiple HPCs for both true and false labels, the correspondingly annotated clustering for other inhibitors, such as VX-745, PI-103 and imatinib, contain only false HPCs. This result demonstrates that the substructures that are informative of inhibitor binding are inherently inhibitor-specific. That is, a subset of binding site positions that are predictive for one inhibitor are not necessarily predictive for another inhibitor.
It is important to note that Fig. 4 and Fig. 5 represent the same clustering for just one 3-residue substructure. However, all 2925 clusterings are computed and all HPCs detected in these clusterings are used to predict binding affinity. The particular three-residue subset shown in Fig. 4A was chosen because the resulting clustering exhibits a number of illustrative features. First, the clustering is relative “clean” with well-separated clusters. Second, it contains highly predictive clusters for both binding and not binding to flavopiridol (the ocher cluster in the top-left and the red cluster in the bottom right of figure Fig. 4B, respectively). None of these features are essential for predicting binding affinity; all automatically selected HPCs in all clusterings are used to predict affinity, each casting one “vote.”
Numerous instances of cross-family affinity for both type I and II kinase inhibitors have been identified, as was clearly illustrated by the kinome affinity maps created by Karaman et al. [31]. It is important to identify structural features shared among phylogenetically diverse kinases that share affinity for a particular inhibitor, because they provide a basis for reasoning about inhibitor cross-reactivity when overall sequence similarity will be low. Furthermore, by identifying these shared structural features, it may be possible to rationally re-engineer the specificity of inhibitors by avoiding the targeting of these features, since they are not unique to the intended kinase target. In order to identify the number of instances of cross-family structural features that can be associated with specific inhibitor binding, the distribution of substructure clusters across all 3-position subsets was analyzed.
Each individual cluster, across all 2925 clusterings and all 38 inhibitors, was evaluated to calculate the purity of both affinity labels and family-level phylogenetic labels. For example, a cluster containing 3 distinct kinase sequences with affinity labels and family labels {agc, camk, tk} would have an affinity purity of and a phylogenetic purity of 0.33. By plotting the affinity and phylogenetic purity scores of each cluster (separately for each inhibitor) as shown in Fig. 6, the distribution of clusters across the spectrum of possible scores can be evaluated. Note that only the clusters having a true label majority are plotted in Fig. 7 (i.e., a true label majority is purity in the true label). Additionally, Table 2 lists per inhibitor statistics for cluster distributions shown in Fig. 7.
In order to build intuition for interpreting the cluster distributions, the cluster distribution for VX-680 (Fig. 7) is examined in more detail because it is representative of the distribution for many of the other inhibitors. As listed in Table 2, 23,495 clusters were identified by ccorps that have purity in the true label for VX-680 (hereafter referred to as true-majority clusters). Only these true-majority clusters are plotted in the cluster distribution shown in Fig. 7, meaning the minimum “affinity purity” displayed in Fig. 7 is 0.5 by definition (because only 2 different affinity labels exist, true and false).
As can be seen in Fig. 7, the vast majority of clusters identified by ccorps have low affinity purity as well as low phylogenetic purity. This is to be expected because highly conserved portions of the kinase atp binding site are known to exist. Structural features that consist of conserved residue positions will be common to many kinases from different families due to the fact that these positions are so heavily conserved, which explains the low phylogenetic purity of these clusters. Furthermore, these conserved features are unlikely to be correlated with the affinity for a particular inhibitor because most inhibitors have been engineered to not have broad cross-reactivity across the kinome. Staurosporine is an exception as it is a very non-selective inhibitor due to its interaction with highly conserved binding site features; the cluster distribution corresponding to staurosporine (Fig. 6) is markedly different from the other inhibitors with most clusters having high affinity purity across a range of phylogenetic purity values.
Examination of the extremes of the VX-680 cluster distribution reveals further insights into the frequency of structural similar features among kinases with different degrees of sequence similarity. Clusters having a phylogenetic purity of 1.0 (i.e., all proteins belong to the same family) but having low affinity purity exist, and for VX-680 276 such clusters were identified by ccorps. This observation is interesting because it illustrates that kinases sharing sequence similarity (relative to kinases outside the family) have multiple common structural features that are not informative of the ability of these kinases to bind VX-680 and are therefore unlikely to be good features to target in design studies. Because ccorps only incorporates clusters with high affinity purity (i.e., HPCs), these conserved structural features that are not indicative of VX-680 binding are ignored by ccorps when predicting affinity for unannotated kinases. This observation can also be made for each of the other inhibitors as shown in Fig. 6.
Another interesting extreme of the VX-680 cluster distribution to examine is the existence of HPCs that are phylogenetically diverse. The HPCs selected by ccorps correspond to the right-most column of points in Fig. 7; these clusters all have an affinity purity of 1.0 for VX-680 and therefore contain only structures with known VX-680 affinity. As can be noted in Fig. 7, HPCs exist at a range of phylogenetic purity values. ccorps identified a total of 2707 HPCs for VX-680, and 1786 (66%) of these HPCs contain proteins belonging to two or more distinct kinase families. This result demonstrates that ccorps is capable of identifying cross-family structural features that are associated with VX-680 binding. Furthermore, this result is not unique to VX-680. As shown in Fig. 6 and tabulated in Table 2, cross-family structural features associated with inhibitor binding were identified for all of the inhibitors tested with the exception of GW-2580, for which no true-majority clusters were identified.
Examination of the cluster distributions across each of the inhibitors reveals a wide range of observations. While many inhibitors have a cluster distribution similar to that of VX-680, for some inhibitors ccorps identified relatively fewer true-majority clusters. For example, only 133 clusters with affinity purity 0.5 were identified by ccorps for SB-431542 and all of these happen to be HPCs. However, even among this relatively low number of HPCs, 69 (52%) of the clusters contain kinases from two or more families. As demonstrated by the corresponding distributions for all 38 inhibitors in Fig. 6, such shared structural similarity is not rare.
The approach used by ccorps to classify an unlabeled kinase is to identify the cluster to which the unlabeled kinase belongs. If the associated cluster is an hpc, the label for the hpc is transferred to the unlabeled kinase. Non-informative clusters containing a mix of labels (non-HPCs) do not contribute to the label prediction process. This “co-clustering” analysis approach is repeated for all of the 2925 substructure clusterings and the final label prediction for an unlabeled kinase is then selected as detailed in Methods.
The ability of ccorps to predict the binding of each inhibitor for proteins within the annotated structural dataset was assessed using the cross-fold validation approach described in the following section. For each of the 38 inhibitor annotation label sets, an independent evaluation of ccorps was performed. No information was shared among the evaluations in order to validate the predictive ability of ccorps to identify structural features predictive of the binding ability of each inhibitor independently.
Identifying structural features of the kinase binding site that directly or indirectly mediate the binding ability of inhibitors is a significant component in developing and optimizing kinase inhibitors. Given the increasingly large number of available kinase structures, kinome-wide comparative binding site analysis is now possible as has been demonstrated here. By combining available structure data with large-scale inhibitor affinity data, it becomes possible to automatically learn the features of the kinase binding site that predict the binding ability of a given inhibitor. This is useful for predicting whether kinases whose binding affinity is unknown will bind to a given drug, but, perhaps more importantly, knowing the structural basis for binding to a particular drug can be exploited in the design of analogs that bind more strongly and have fewer off-target interactions. This information could further improve well-established structure-based computer-aided drug design methods, where it is challenging to develop reliable models for the contributions of individual interactions or groups of interactions between inhibitor and protein to binding affinity.
ccorps has been demonstrated here to be capable of learning the features of the kinase binding site that are informative of inhibitor binding across a set of 38 inhibitors. Furthermore, the binding site features selected by ccorps as informative of inhibitor activity/inactivity have been shown to be interesting in and of themselves, for example, the existence of residue triad clusters that are unique in kinases capable of binding a given inhibitor but that exist within kinases from different major branches of the kinase family tree. The identification of such shared binding site features among sequence-diverse kinases is an important contribution for structure-based methods because of the relative difficulty of identifying small subsets of sequence non-contiguous but spatially compact positions that are correlated with a given indicator, such as inhibitor binding ability. The complete set of 41,964 true-majority HPCs that contain kinases from two or more of the kinome families as defined by Manning et al. [2] is provided as Dataset S1 to facilitate further analysis of these phylogenetically diverse structural features that distinguish kinases binding each of the 38 inhibitors.
As was demonstrated here, ccorps is capable of incorporating all of the available protein kinase structure data, so as to operate at the kinome scale, and then using this data to construct highly accurate predictors of kinase affinity for a variety of different small molecule inhibitors. While ccorps relies upon the aggregation of structural similarity that coincides with affinity similarity to build predictors, the individual instances may be informative in and of themselves. Further analysis of the vast number of structurally similar features shared among phylogenetically distant kinases may provide additional insights into the structural mechanisms of inhibitor recognition occurring across the kinome.
The existence of affinity datasets containing structurally similar inhibitors, that differ by only one or a small number of chemical substitutions, provides the opportunity to associate specific structural features identified by ccorps with specific inhibitor pharmacophores. A recent approach by Milletti and Hermann [6] has been demonstrated to identify specific chemical transformations that can be associated with selectivity differences. In future work we will seek to further incorporate this cross-inhibitor level of analysis and broaden the scale of the structure dataset by further incorporating newly available kinase crystallographic structures.
Several potential optimizations of ccorps may increase its inhibitor binding prediction performance on broad spectrum inhibitors. For the 38 inhibitor dataset analyzed in this paper, the number of HPCs identified was well correlated with the number of kinases inhibited (). That is, ccorps tended to perform less well on inhibitors for which large numbers of HPCs were identified. Developing an approach to weighting and ranking the large number of HPCs generated by broad spectrum inhibitors may aid in increasing the predictive performance of ccorps for these inhibitors. For example, ranking HPCs by the mean within-cluster affinity () would more heavily weight structural features correlated with strong binders and decrease the impact of structural features only correlated weak binders. Such an approach would help to increase the signal-to-noise ratio of HPCs when the number of HPCs identified grows large. As our results showed, there are cases where ccorps significantly outperforms a sequence-based method, but there also cases where the reverse is true. While this paper focused on quantifying the extent at which structure alone can be used to predict binding affinity, for practical usage we envision that structure- and sequence-based methods are used in tandem.
A major advantage of the work presented is the generality of ccorps to detect structurally distinguishing features for a wide variety of applications beyond the kinase inhibitor affinity analysis presented here. No assumptions regarding the nature of the annotation labels nor of the alignment type are made at any point by ccorps. ccorps provides a general framework for automatically learning structural features that distinguish proteins having different annotation labels. This allows the incorporation of purely structure-based alignments, such as those available in databases like homstrad [33] or even local structure alignments such as those identified by motif/template search algorithms (e.g., soippa [34], and LabelHash, [35]). Other sources of annotation labels, including Gene Ontology (GO, [14]) terms, binding affinity for a given molecule and ligation state can be incorporated as-is with ccorps without modification to the method.
|
10.1371/journal.pcbi.1005072 | Determining Physical Mechanisms of Gene Expression Regulation from Single Cell Gene Expression Data | Many genes are expressed in bursts, which can contribute to cell-to-cell heterogeneity. It is now possible to measure this heterogeneity with high throughput single cell gene expression assays (single cell qPCR and RNA-seq). These experimental approaches generate gene expression distributions which can be used to estimate the kinetic parameters of gene expression bursting, namely the rate that genes turn on, the rate that genes turn off, and the rate of transcription. We construct a complete pipeline for the analysis of single cell qPCR data that uses the mathematics behind bursty expression to develop more accurate and robust algorithms for analyzing the origin of heterogeneity in experimental samples, specifically an algorithm for clustering cells by their bursting behavior (Simulated Annealing for Bursty Expression Clustering, SABEC) and a statistical tool for comparing the kinetic parameters of bursty expression across populations of cells (Estimation of Parameter changes in Kinetics, EPiK). We applied these methods to hematopoiesis, including a new single cell dataset in which transcription factors (TFs) involved in the earliest branchpoint of blood differentiation were individually up- and down-regulated. We could identify two unique sub-populations within a seemingly homogenous group of hematopoietic stem cells. In addition, we could predict regulatory mechanisms controlling the expression levels of eighteen key hematopoietic transcription factors throughout differentiation. Detailed information about gene regulatory mechanisms can therefore be obtained simply from high throughput single cell gene expression data, which should be widely applicable given the rapid expansion of single cell genomics.
| Many genes are expressed in bursts, which can contribute to cell-to-cell variability. We construct a pipeline for analyzing single cell gene expression data that uses the mathematics behind bursty expression. This pipeline includes one algorithm for clustering cells (Simulated Annealing for Bursty Expression Clustering, SABEC) and a statistical tool for comparing the kinetic parameters of bursty expression across populations of cells (Estimation of Parameter changes in Kinetics, EPiK). We applied these methods to blood development, including a new single cell dataset in which TFs involved in the earliest branchpoint of blood differentiation were individually up- and down-regulated.
| Many genes are expressed in stochastic bursts: there are time periods where many transcripts are quickly produced, interspersed randomly with gaps of little or no transcriptional activity. Bursting gene expression was initially proposed as a mechanism to explain why cells in a seemingly uniform cell culture responded heterogeneously to steroids [1]. Two decades later, new live imaging technologies enabled researchers to observe transcriptional and translational bursting in real-time, finally confirming that bursting gene expression is a widespread phenomenon [2–4]. In fact, Dar et al. [5] tested 8,000 human genes and found that all of them were expressed in episodic bursts.
Ko et al. [6] described bursting gene expression using a two-state model of gene expression, depicted in Fig 1A. In this model, each gene can either be in an on or an off state, and the gene stochastically transitions between these states, with transcription only taking place when the gene is on. The distribution of mRNA across a population of cells is determined by the following three kinetic parameters: the rate the gene turns on (Kon), the rate the gene turns off (Koff) and the rate of transcription when the gene is on (Kt), all normalized to the rate of mRNA degradation [7]. The values of these three kinetic parameters determine the distribution of mRNA transcripts within a population of cells (Fig 1B).
Predicting the rate at which genes turn on, turn off, and transcribe mRNA can provide insights into how genes are regulated. For instance, TFs which are necessary for transcription would control the rate at which the genes turn on (Kon). One example of such a TF would be a pioneering factor, which opens up the chromatin to allow transcription. Brown et al. [8] found that switching between active/inactive states corresponded to distinct chromatin changes in PHO5. This is consistent with the results in Dadiani et al. [9], which show that manipulating nucleosome-disfavoring sequences in yeast can influence the burst frequency. On the other hand, TFs that control Kt are responsible for modulating the levels of gene expression of genes that are already on [10]. For instance, they may be involved in polymerase II (PolII) recruitment or transcriptional elongation. Therefore, estimating these kinetic parameters could help generate hypotheses for gene regulation mechanisms.
Until now, the study of transcriptional bursting has been limited by the available experimental approaches. The most common high-throughput strategies (conventional RNA-seq or qPCR) for measuring gene expression require biological material from thousands of cells. These bulk strategies only measure the average levels of gene expression in populations of cells, data that cannot be used to make functional predictions about the bursting dynamics of transcription. While transcriptional bursting can be visualized in real-time in single cells, this is a low-throughput approach which can only measure expression for a single gene per cell [3, 4]. Recently, there has been an emergence of single cell resolution RNA-seq and qPCR technologies, which can observe the full profile of gene expression in a population of cells. However, these are snapshot methods, which can only measure gene expression at a single point in time, because they involve lysing the cells.
Nevertheless, preliminary studies have shown that it is possible to utilize the shape of the distribution of gene expression at a single point in time to estimate the kinetic parameters of the two-state model. Raj et al. [11] used a variation of the mathematical analysis done by Peccoud and Ycart [7] to estimate the kinetic parameters in fluorescent in situ hybridization (FISH)-based expression studies, and Kim and Marioni [12] developed a strategy to estimate kinetic parameters in RNA-seq data. In addition, Teles et al. [13] applied a similar approach to a single cell qPCR dataset in the context of a hematopoietic developmental system.
As it is now accepted that bursting dynamics can in principle be resolved from single cell snapshot datasets, we can take advantage of this type of analysis to develop new algorithms that can answer a wide range of biological questions. In this paper, we apply our understanding of bursting gene expression to develop a more effective clustering algorithm for single cell gene expression data, which we call Simulated Annealing for Bursty Expression Clustering (SABEC). This can help researchers identify previously uncharacterized sub-populations of cells within their single cell data. Secondly, we develop a statistical tool for identifying whether the burst frequency (Kon or Koff) or burst magnitude (Kt) are the source of gene expression variations between experimental samples. Instead of simply observing whether gene expression increases or decreases across two populations of cells, this Estimation of Parameter changes in Kinetics (EPiK) toolkit allows researchers to make inferences about the underlying regulatory mechanisms affecting gene expression. We provide a complete pipeline in R for analyzing single cell qPCR data, including data normalization steps, SABEC clustering and EPiK cluster comparisons.
This pipeline can be utilized in a wide range of biological contexts. We apply it to study how some of the key transcription factors in hematopoiesis are being regulated, discovering that the hematopoietic stem cells studied in Moignard et al. [14] formed two distinct sub-populations, distinguishable by Tel and Gata1 expression levels. Then, by manipulating the expression levels of two of the TFs involved in early differentiation (Gfi1 and Gata2), we predict that the primary mechanisms by which these TFs influence their downstream targets is by manipulating Kon. Finally, we compare the mechanisms by which cell surface markers are regulated in healthy and leukemic cells based on data previously reported in Guo et al. [15].
We have developed a pipeline for analyzing single cell qPCR data in order to determine the mechanism by which TFs were regulated during differentiation (Fig 2).
The input data consists of single cell qPCR data for a set of genes, where the cells have been sorted into their expected cell populations, using fluorescence-activated cell sorting (FACS). There are two main outputs of the pipeline. Firstly, cells are divided into populations (clusters) based on their bursting behaviour– each gene has the similar rates of activation (Kon), rates of inactivation (Koff) and rates of transcription (Kt) across all cells within each cluster. Secondly, the pipeline compares between each pair of cell populations and predicts the mechanism by which each gene changes its expression– by changing the rate the gene turned on and off or by changing the rate of transcription once the gene is already active.
A key component in the pipeline is our strategy for estimating the kinetic parameters. Previous approaches for estimating the kinetic parameters were slow to compute, because they involved iterative algorithms, such as simulated annealing [13], expectation maximisation [11], or Gibbs Sampling [12]. Meanwhile, we have a pre-computed look-up table for P(x|Kon, Koff, Kt), the probability of seeing x mRNA transcripts, given a set of kinetic parameters (Fig 2A and Eq 1). This table can be used to compute the most likely set of kinetic parameters using a single matrix multiplication– a very efficient computation (Eq 2).
In the pipeline, first we normalize the measurements from the qPCR and transform them into estimated mRNA counts (See Eq 6). Next, these measurements are input into a newly developed unsupervised clustering algorithm (SABEC) which identifies populations of cells with uniform bursting kinetics. SABEC is an iterative algorithm that starts by randomly assigning cells to clusters; next, the algorithm alternates between estimating the kinetic parameters for each gene within each cluster (using Eq 2) and probabilistically re-assigning cells to clusters based on their likelihood of belonging to each cluster (using Eq 9), until the algorithm converges– i.e. when fewer than 5% of cells swap clusters. SABEC is run fifty times using different initial sets of randomly assigned clusters, and the results are summarised as a consensus matrix– a matrix that records the number of times two cells were grouped together. The number of clusters can be determined by one of the following methods: the proportion ambiguously clustered (PAC), Variable Information (VI) or the corrected Rand score (See Fig 2B).
Finally, each pair of populations is compared using a statistical tool we developed called EPiK, in order to identify whether each gene is regulated by modifying the rate the gene turns on (burst frequency) or the level of transcription once the gene is active (burst magnitude). EPiK is a compilation of three different methods: the Bayesian Information Criteria (BIC), a Marginal Probability (MP) Score, and a subsampling method. Since for each gene there can be between 0 and 3 parameters that can differ across a pair of populations, BIC identifies the set of parameters most likely to differ after penalising parameter sets that are larger (See Eq 12). The MP score separately calculates the likelihood of each parameter changing, independent of whether the other parameters change or stay the same (See Eq 15). In the subsampling method, the kinetic parameters are calculated for small random subsets of cells, then the distributions of estimated kinetic parameters for the subsets are compared by the Kolmagorov-Smirnov statistic(Fig 2C).
All together, our R data processing scripts, SABEC, and EPiK come together to form a pipeline that utilizes the mathematics of bursting gene expression to determine how genes are differentially regulated across populations of cells (Fig 2D).
These methods were all tested in silico using data sets that have similar structures to the experimental datasets. Each of the three methods introduced in this paper (ML approach for kinetic parameter estimation, SABEC and EPiK) required a different set of simulated data.
The ML method for kinetic parameter estimation was benchmarked populations of cells with known kinetic parameters to allow us to quantify the accuracy of the method. 3,000 parameter sets were randomly selected (uniformly distributed) with Kon between 0 and 5, Koff between 0 and 20, and Kt between 0 and 600. We randomly sampled 10% of the transcripts from each of the simulated cell, to represent the technical noise caused by the loss of 90% of the starting material during sample preparation. We included a stochastic loss of mRNA transcripts to account for material loss during cDNA library construction– Islam et al. [33] estimates that only 48% of transcripts are reverse transcribed into cDNA, and Wu et al. [34] could capture 42% of the total unique transcripts that were identified in bulk RNA-seq.
The simulated datasets for testing the SABEC method were chosen to be as similar to the experimental datasets as possible. 100 simulation sets were generated, with each completely parallel to the Moignard dataset; Each simulation set consisted of five populations of 124 cells with 18 genes each, with their kinetic parameters equal to those estimated by the ML method for the experimental data.
Finally, to test the final method of comparing kinetic parameters between two populations of cells, 1600 simulated datasets were generated, each one consisted of a pair of populations of 124 cells with 1 gene each. There are eight combinations of kinetic parameters that can change (none, all, three ways one parameter can change and three ways two parameters can change). Two hundred pairs of populations were selected for each of these eight scenarios, with 100 simulations with 0 < Koff < 5 and 100 simulations with 5 < Koff < 10. The range of the other parameters were 0 < Kon < 5 and 0 < Kt < 600. In these simulations we also simulated the random loss of 90% of the mRNAs.
Even after simulating 90% loss of biological material, the predicted parameters correlated well with their real values, although Koff was the most difficult parameter to predict (See Fig 3A–3C), especially at wider parameter ranges (S1 and S2 Figs), but it still performed better than the Kim and Marioni Method (S3 Fig). These in silico results emphasize that the cDNA library efficiency can have a large impact on the absolute values of predicted parameters, which suggests that raw parameter values are not comparable across different experiments. SABEC also accurately predicted clusters of simulated datasets, with similar population structures and parameter ranges to the experimental dataset (Fig 3D–3F). Finally, we found that the EPiK method was very conservative, with approximately a 0.05% false positive rate when the methods were intersected (Fig 3G and 3H). Further details about the validations and comparisons with alternative methods are described in depth in the Methods. Next, we applied these methods to experimental datasets.
The dataset we use to test our pipeline comes from Moignard et al. [14], which is a high quality single cell qPCR dataset that includes approximately 124 cells each from five different populations of cells during hematopoeisis. Hematopoiesis is the process by which hematopoietic stem cells (HSC) in the bone marrow differentiate into different types of red and white blood cell types. This process can be depicted as a differentiation tree, in which each cell must make multiple “decisions” at each branching point in the tree that will determine its final cell fate. Moignard et al. [14] includes two key branching points: i) HSC cells can become lymphoid-primed multipotential progenitors (LMPPs) or premegakaryocytes (PreMegEs, also referred to as PreMs in figures) and ii) LMPP cells can become granulocyte-macrophage progenitors (GMPs) or common lymphoid progenitors (CLP).
The focus of this study is on the densely interconnected TF regulatory network that has been shown to contribute to cell fate decisions. There is strong evidence that at least seven of these TFs directly interact with one another, potentially forming a SCL/Lyl1/Gata2/Runx1/Lmo2/Fli-1/Erg heptad, with some TFs directly binding to the DNA (such as Gata2 and SCL) and other TFs serving a bridge between the DNA bound components (such as LMO2) [16]. Other key TFs that were profiled in Moignard et al. [14] were PU.1, Meis1, Hhex, Tel, Nfe2, Eto2, Mitf and Ldb1.
There are a number of open questions about the differentiation of HSC into the various progenitor cell populations. Firstly, although the cells were assigned to their populations via FACS, it is unclear whether these sub-populations are truly uniform. For instance, some HSC cells may be biased towards self-renewal or producing cells in the lymphoid or myeloid lineages.
Secondly, the specific mechanisms by which these TFs are being regulated are as yet uncharacterized. Influencing the rate at which genes turns on (Kon) could increase the proportion of cells with an active copy of a gene; whereas manipulating the transcription rate (Kt) would result in there being higher concentrations of mRNA, while maintaining the same proportion of cells with an active gene. In hematopoiesis, cellular TF concentrations help determine which branch the cell will take as it differentiates towards its final cell fate. Therefore, choosing to manipulate Kon instead of Kt could influence the proportion of cells that enter a certain differentiation trajectory. In addition, since the expression of these TFs is tightly linked, one gene’s bursting dynamics could have repercussions on the dynamics of the entire network.
Although the cells that we study have been sorted into distinct subpopulations using FACS, this does not guarantee that the populations are indeed transcriptionally uniform. Some cells may be misclassified by FACS (expected misclassification rate is 1% for the Moignard dataset), and some of the known populations may be composed of as yet unidentified subpopulations. In addition, extrinsic variability, such as having cells in different stages of the cell cycle, could cause the populations to be heterogeneous.
It can be difficult to accurately identify homogenous subpopulations using standard clustering approaches like K-means or hierarchical clustering (S5 Fig). For instance, K-means is most effective when each cluster is normally distributed and has similar variances. Due to bursting, gene expression is unlikely to come from such a distribution; gene expression distributions can look like a Poisson distribution, a negative binomial distribution, or even a bimodal distribution, depending on Kon, Koff, and Kt. Therefore, we developed a simulated annealing strategy (Simulated Annealing for Bursty Expression Clustering, SABEC), which takes into account bursting gene expression, in order to have more robust clustering. SABEC was rigorously tested against simulated datasets that were designed to be as close as possible to the structure of the experimental data (S6 Fig). Additionally, the algorithm was tested on a wide range of simulated datasets, to see if this method could perform well under varied conditions, such as on data with different numbers of genes measured in each cell and different numbers of cells in each population (S7 Fig).
Next, SABEC was applied to the experimental single cell qPCR data from Moignard et al. [14]. The final clustering is depicted in Fig 4A. Although the predicted hematopoiesis differentiation tree is expected to look like Fig 4Bi, we found that HSC is divided into two distinct subpopulations. One of these populations had 12.8% of cells expressing Gata1, while the other population had none. Since Gata1 is uncommon in HSC cells, but often found in multi-potent progenitor populations (MPP, the earliest progenitor population formed by HSCs [17]), we hypothesized that the populations of cells without Gata1 may be self-renewal HSCs and the other may be HSCs poised for differentiation.
Our clustering with SABEC was based solely on the expression of 18 TFs. However, Moignard et al. [14] also profiled cell surface protein tyrosine kinase c-Kit. It is known that low levels of c-Kit correlate to greater self-renewal potential in HSCs [18]. Even though all of the HSC cells were positive for the c-Kit protein according to FACS, not all the cells had high levels of c-Kit transcripts. Our hypothesized self-renewal HSC population had significantly lower levels of c-Kit than our poised-to-divide population (p-value 9.117e-06 with Kolmogorov-Smirnov test). Even though c-Kit was not one of the genes used to cluster the cells, there was substantial difference in expression levels, suggesting the tree topology depicted in Fig 4Bii. Further evidence for this arrangement is in S9 Fig.
The SABEC algorithm has provided us with subpopulations that appear to have uniform bursting kinetics. In the next section we will identify which kinetic parameters were most likely adjusted across each pair of cell populations, using EPiK. EPiK works best when cell populations have uniform bursting dynamics; S11 Fig shows how having mixed cell types can influence estimates of kinetic parameters. Therefore, we remove cells that do not cluster well with other cells of the same type, with cutoff thresholds designated by the vertical lines in Fig 4C, which is determined by the region of the curve with the steepest slope. However, this pruning protocol could bias parameter estimation if it were to eliminate cells that are falsely identified as outliers (S10 Fig). Therefore, we run EPiK both on the pruned and unpruned datasets, and only consider parameters that are consistently found to have changed under both conditions. It is important to note that the proposed outlier cells may be of biological interest– for instance, they may be rare cell types or cells in the process of differentiation. The pruned dataset is only for use to boost the accuracy of predictions with EPiK, but all cells should be observed for other types of analysis.
EPiK incorporates three different metrics for evaluating whether kinetic parameter changes are significant. By taking the intersection of these three prediction methods, the false positive rate decreases without a significant drop in the true positive rate. This gives us a very conservative list of probable kinetic parameter changes (a 0.05% false positive rate with our simulated dataset).
The first method is the Bayesian Information Criterion (BIC) (S12 Fig), and the second method is the marginal probability (MP), which is the log likelihood of a certain parameter being varied, independent of whether the other two parameters vary or stay the same (S13 Fig). In the third method, we repeatedly subsample cells from each population and estimate the kinetic parameters for each subset, comparing the maximum distance between the cumulative density functions of the distributions (S14 Fig).
These three methods were applied to the experimental data from Moignard et al. [14]. Each of the methods is based on slightly different assumptions, so they each have different distributions of parameters being varied (See S15 Fig). For instance, BIC predicts that Koff is adjusted in a few cases, but this is not deemed significant by either the MP or subset methods. In addition, there are more significant parameter changes in the case of the pruned populations compared to the complete populations, because these populations are more distinct from one another.
We can now take a closer look at a few examples of TFs that have predicted kinetic parameter changes during differentiation (Fig 5). The predictions from each method are drawn along the branches of the hematopoiesis differentiation tree (Fig 5A–5E). To demonstrate the magnitude and direction of these kinetic parameter changes, Fig 5E–5H illustrates the kinetic parameter estimates for each population of cells.
In Fig 5A and 5F, four out of the six methods predict that Eto2 is up-regulated by increasing Kon (red) during the HSC to PreM transition, which is consistent with Eto2 having a role in increasing the proportion of cells that differentiate into PreM [19].
One striking feature of Fig 5A–5E is that Eto2, Mitf, and PU.1 are regulated by different kinetic parameters in different stages of blood differentiation. PU.1 has known cell-type-specific enhancer elements [20], and our results may suggest that each of these may regulate PU.1 through different mechanisms. On the other hand, Nfe2 is consistently regulated by Kt throughout differentiation.
In some instances, all six methods come to a consensus as to which kinetic parameter was the source of gene expression variability for a particular TF, and these are shown in Fig 6A. Throughout hematopoiesis, most of the differences in gene expression come from changes of Kon, but Lmo2, Nfe2 and Meis1 are regulated by Kt in the transition from LMPP to GMP. Recall that in the previous section, we identified that HSC forms two distinct sub-populations. We compared the two HSC sub populations with their child populations (LMPP and PreM) (see Fig 6). Between HSC1 and HSC2 only Tel seemed to consistently change its kinetic parameter (Kon). As expected, the cell population with higher Gata1 and c-Kit expression has more in common with LMPP and PreM cells than the other HSC subpopulation.
In summary, our methods have allowed us to not only identify which genes have changed their expression, but also what physical mechanism caused that change.
In the previous section, we identified TFs that were regulated by Kon or Kt during blood differentiation, but it is unclear which TFs were controlling these changes. It may be possible to discover the specific mechanistic role of a TF by manipulating its expression experimentally and then calculating the change in kinetic parameters of its downstream targets.
We decided to focus on the Gfi1-Gfi1b-Gata2 subnetwork that was identified as being important at the first branching point of HSCs to LMPP and PreMegEs [14]. Primary HSCs are difficult to isolate in large quantities, culture and manipulate, so we turned to HPC7 cells, a model cell line for hematopoietic stem and progenitor cells which has some differentiation potential towards more mature blood cells [16, 21]. Similarly to HSCs, HPC7 cells express Gata2 and Gfi1b, but little or no Gfi1. We therefore up-regulated Gfi1 expression and down-regulated Gata2 expression and performed single cell gene expression analysis for the gene set described by Moignard et al. [14], as well as some additional genes involved in HSC differentiation. We analysed 81 cells expressing an shRNA against Gata2 and 77 cells control cells expressing an shRNA against Luciferase, and 72 cells overexpressing Gfi1 and 45 control cells expressing an empty vector.
There are a number of strategies by which a TF could reduce gene expression: by decreasing the rate a gene turns on, by increasing the rate a gene turns off, or by decreasing the rate of transcription of an active gene, and each of these strategies would result in different temporal dynamics of gene expression bursting. All six methods came to a consensus that up-regulating Gfi1 seemed to significantly alter Erg, Gfi1b, Hhex and Mpl, by lowering Kon. Previous research suggests that Gfi1 is usually a repressor that either keeps the chromatin in a condensed state or actively competes for binding with activators [22]. Both of these mechanisms of action are consistent with Gfi1 down regulating its targets by lowering Kon. Based on ChIP-seq experiments from Sanchez-Castillo et al. [23], Gfi1 binds in or near all four of these potential targets (see S1 Table).
In the other experiment, Gata2 was down-regulated; however, this was not a complete knockdown, with only a slight overall decrease in expression (S20 Fig). All six methods suggest that Gfi1b expression was decreased via a change in Kon as Gata2 levels decreased. In the pruned population of cells, Procr (also known as EPCR, a known target of Gata2) had higher Kon after the knockdown [24].
When Gata2 and Gfi1 were down- and up-regulated, our methods could only detect changes in Kon. Therefore, we can hypothesize that this could be the mechanism of action of these two TFs.
Our pipeline for the identification of differential kinetic parameter values can also be applied to compare healthy and diseased cell populations. In particular, we apply it to four of the cell populations isolated in Guo et al. [15]: two cell types from a healthy mouse (GMP, Lin+) and two from a leukemic mouse (LGMP, LLin+), whose leukemic cells came from a MLL-AF9 fusion protein [25]. The most significant kinetic parameter changes are shown in Fig 7. Some of the genes profiled by Guo et al. [15] were found in fewer than 10 cells in one or more cell population, and these were excluded from kinetic parameter comparisons.
Some of the genes are enriched in a single cell population; for instance, TLR9, a factor whose expression influences the prognosis of leukemia [26], is found predominantly in the LLin+ cells, and appears to be regulated by Kon. Most of the identified kinetic parameter changes were in Kon or a combination of Kon and Kt. However, MUC13 and SIGLEC5 were predicted to have been regulated by only Kt and IL3RA was predicted to be regulated by Koff. Interestingly, IL3RA is the only example where all methods predict changes in Koff, which suggests that this is a promising gene to focus on in future research.
In this study, we exploit established mathematical models of bursting gene expression to develop a new pipeline for analyzing single cell qPCR data to more robustly cluster biological samples and provide insight into the mechanics of gene regulation. We apply these methods to study gene regulation in hematopoietic stem cell and progenitor populations. Even though single cell qPCR data can only provide snapshots of gene expression in a population of cells across different time points, we can infer the temporal dynamics of gene expression in these cells, and use this information to infer the population substructure (via SABEC) or regulatory mechanisms (via EPiK). This pipeline can be applied to study how genes are regulated during the natural process of differentiation or as cells progress into a diseased state. In addition, by manipulating the expression of a TF within its cell culture, we can infer its specific regulatory role. These algorithms perform well in simulated datasets (S1, S2, S6, S7, S8, S12, S13 and S14 Figs), performing better than similar computational tools (Figs 8, S3 and S5).
Instead of simply observing how much gene expression heterogeneity there is in a sample, it is now possible to predict the specific regulatory mechanisms that contributed to heterogeneity. Most crucially, this type of analysis can be done in a high-throughput manner.
In addition, commonplace clustering algorithms like K-means and hierarchical clustering are not meant to cluster data drawn from Poisson, negative binomal, and bimodal distributions, as is the case for single cell gene expression data. For instance, K-means performs best on data that comes from normal distributions with similar standard deviations. For this reason, it is critical to use a clustering algorithm that incorporates information about the shape of the gene expression distributions when analyzing single cell resolution datasets.
However, it is unclear how the cell cycle could influence our results, so future experiments ought to include cell cycle markers as controls [27]. In addition, our approach assumes that the gene expression distributions are close to equilibrium. Fortunately, Peccoud and Ycart [7] demonstrated that the two state model approaches the equilibrium distribution very rapidly (at an exponential rate), so this assumption likely holds. Other researchers have attempted to fit a multi-state model to data, instead of a two state model [28], but unless there is strong evidence for a more complex model, it is wise to use the simplest approach to avoid over-fitting the data. In the future, it may be possible to modify the model to detect cells that are in the process of transitioning between cell populations– for instance, these may be cells that have estimated kinetic parameters that are between two other populations. In addition, Teles et. al. [13] estimated the probability of transition between cell populations using machine learning methods, but this is beyond the scope of this paper.
One particularly useful application of this pipeline is the validation of assumptions used to model specific sub-networks of genes important in differentiation. Often, these models assume certain mechanisms by which the TFs influence one another. For instance, Narula et al. [29] constructed a mathematical model of a hematopoietic sub-network under the assumption that Koff was the parameter that is biologically regulated, in the absence of any experimental data. Instead of arbitrarily selecting a modeling strategy, we can now choose one that fits the data best. In addition, we have discovered that different TFs are regulated through different mechanisms in each stage of differentiation, implying that a single model of a gene network might not universally apply. Therefore, these strategies would allow us to make more biologically plausible models of gene subnetworks. Having a better understanding of gene regulation processes is important in order to learn how these are perturbed in disease, and also to develop protocols that produce desired cell types for cell therapy.
A modified version of this pipeline for RNA-seq data would be an important future development. The kinetic parameter estimation strategy and EPiK may be applied to single cell RNA-seq, as long as there is sufficient sequencing depth to capture the population-wide distribution of gene expression. Both these methods scale approximately linearly with the number of cells and genes under study, and could also be run in parallel for very large datasets. However, SABEC would not scale well with the large number of genes analysed in RNA-seq experiments, so alternative clustering approaches must be tested in an RNA-seq context.
The transcription process is a multi-step chemical reaction: chromatin must enter the correct state, TFs must bind in the right places, the general transcription machinery must be recruited and initiated, etc. It is currently impossible to distinguish the effects of all of these mechanisms in a high-throughput way. Our pipeline provides a first attempt to understand how the kinetic parameters underlying complex transcriptional processes influence heterogeneity within and across cell populations, through the analysis of single cell gene expression data.
HPC7 cells [21] were grown in suspension in Iscove’s Modified Eagle’s Medium (IMDM, Gibco) with 10% FCS, 10% stem cell factor-conditioned medium, 1% penicillin/streptomycin (Sigma) and 1.5 × 10−4 M monothioglycerol (MTG) at 37°C and 5% CO2. Cells were passaged every two days to maintain a concentration of 0.5 − 2 × 106 cells/ml.
For knockdown experiments, shRNA fragments were cloned into pMSCV/LTRmiR30-PIG (pLMP, Open Biosystems): Luciferase (control, 5’ CACGTACGCGGAATACTTCGAA 3’, [30]), Gata2 (5’ CGCCGCCATTACTGTGAATATT 3’, [31]). For overexpression experiments, the mouse Gfi1 cDNA was inserted into pMSCV-ires-GFP (Addgene plasmid 20672), with the empty vector used as a control. Retrovirus was produced using the pCL-Eco Retrovirus Packaging Vector (Imgenex) in 293T cells.
HPC7 cells were infected with retrovirus by centrifugation at 800 xg at 32°C for 1.5 hours with 4 μg/ml polybrene (Sigma), after which the retroviral supernatant was replaced with fresh media and cells were cultured as normal. Transduction efficiency was monitored by flow cytometry for GFP.
Single GFP+ cells were sorted by FACS into individual wells of 96 well plates and single cell RT-qPCR was carried out as described previously [14]. Cells were captured 48 hours after retroviral transduction for Gfi1 overexpression and 72 hours after transduction for Gata2 knockdown (S1 Table).
Bursting gene expression can lead to a number of different distributions of gene expression in a population of cells (See Fig 1B), ranging from a bimodal distribution (when Kon and Koff are low) to a Poisson distribution (when Kon is much higher than Koff) to an exponential decay-like distribution (when Koff is much higher than Kon). The probability of having x mRNA molecules in a cell with kinetic parameters Kon, Koff and Kt is given by the analytical solution developed by [11]:
P ( x | K o n , K o f f , K t ) = Γ ( K o n + x ) Γ ( K o n + K o f f ) K t x Γ ( x + 1 ) Γ ( K o n + K o f f + x ) Γ ( K o n ) 1 F 1 ( K o n + x , K o n + K o f f + x , - K t ) (1)
In this equation, 1F1 represents the confluent hypergeometric function of the first kind, a summation over an infinite series that is time intensive to compute. To improve our runtimes, we precomputed an extensive lookup table of values of P(x|Kon, Koff, Kt). Let us say that we have n cells, each with an mRNA molecule count (for a particular gene) of xi, and let X = {x0, x1, …xn}. Given this list of mRNA counts from a semi-uniform population, we can assess the log likelihood of each possible set of parameters:
L ( K o n , K o f f , K t | X ) = ∑ X ( ln ( P ( x i | K o n , K o f f , K t ) ) ) (2)
The set of kinetic parameters that has the maximum likelihood is chosen. However, it is important to note that some areas of the parameter space are more sensitive to parameter changes than others. For instance, [11] notes that at large values of Koff the equation of P(x|Kon, Koff, Kt) approaches:
P ′ ( x | K o n , K o f f , K t ) = 1 + K t K o f f - K o n Γ ( K o n + x ) Γ ( K o n ) Γ ( x + 1 ) K t K o f f 1 + K t K o f f x (3)
This equation depends of the ratio of Kt/Koff rather than Kt and Koff separately, which means that Koff and Kt are more difficult to distinguish as Koff increases. The practical implication of this observation is illustrated in Fig 9A and 9B, which depict the log likelihoods at different kinetic parameter values for GFI1b in HSC cells, with regions that have log likelihoods close to the peak value coloured in (specifically, within 0.5 of the maximum likelihood). Although there is only a narrow range of possible Kon values (A), there is a range of values where Koff and Kt can partially compensate for one another (B). A specific example is illustrated with simulated data in Fig 9C: while the original distribution of gene expression (grey) varies visibly when Koff (blue) or Kt (red) are varied, it is possible to change both variables (purple) and almost recover the original distribution. Furthermore, this analysis suggests that it would be difficult to incorporate any additional parameters into the system and find unique estimates for each of them.
In S1 Fig, we test the performance of this strategy of kinetic parameter estimation on simulated data, including artificial technical noise. Although Koff cannot be accurately identified when Koff > 5, the ratio of Koff to Kt is accurately predicted (S2 Fig). Our method also performs better than the Gibbs Sampling based approach developed by Kim and Marioni, 2012 (S3 Fig).
Next, we applied this maximum likelihood approach for kinetic parameter estimation on hematopoietic stem cell and progenitor populations (CLP, GMP, CLP, LMPP and PreM cells) from [14]. The values for Kon and ln(Koff/Kt) for the Moignard data are shown in Fig 9F. Note that many of the TFs were particularly chosen due to their importance in early differentiation of HSCs, so one would expect a lower level of gene expression (and therefore a higher Koff/Kt ratio) in later stage progenitor populations such as CLP and GMP. There is a wide range of estimated kinetic parameter values across the TFs in each cell population; however, we need to ensure that these kinetic parameter estimates are not being skewed by mis-classified cells before we can evaluate whether these differences are statistically significant.
It is crucial to select a table of appropriate range and point density for applying to the experimental data. The criteria for selecting this table were: i) fewer than 5% of the experimental points were at the maximum parameter value for the Kon and Kt parameters. ii) Koff had to be high enough in order to include the parameter range in which only the ratio of Kt to Koff matter ii) the density of points was sufficient to minimise artifacts arising from the discretization of the parameter space.
Given these constraints, the range of parameters was chosen to be 0 < Kon < 5, 0 < Koff < 20 and 0 < Kt < 200 and possible mRNA counts as 0 < x < 200. The sampling density for Kon was every 0.1, for Koff every 0.4 and for Kt every 5 for a total of 395000 possible parameter sets.
We deposited the look-up table, the code used to generate the table in Mathematica and the code for calculating the maximum likelihood in R on Github: https://github.com/ezer/SingleCellPipelineOverview.
The main source of data analysed in this paper comes from Moignard et al. [14], and contains single cell qPCR data from five populations of hematopoietic stem cell and progenitor cells (CLP, GMP, HSC, LMPP, PreM), as determined by FACS, with approximately 124 cells in each population. The genes profiled by the qPCR include 18 TFs with crucial roles in cell fate, which were normalised to two “housekeeping” genes (PolII and Ubc), as described in Moignard et al. [14].
The outcome of a PCR experiment is a normalised Ct value, which relates to mRNA molecules (x) as follows:
x = i n t ( b · 2 a - C t ) (4)
where a and b are constants. For each TF, we chose b:
b T F = x max 2 a - min ( X T F ) (5)
where xmax = 200 is the maximum number of mRNA pre-calculated in our lookup table and XTF is the set of mRNA counts for the particular TF. This choice of bTF stretches the values of x to have as wide a range as possible. It also removes the need to set an a parameter, since the equation for calculating x can be simplified to:
x = i n t ( x m a x · 2 min ( X T F ) - C t ) (6)
It is important to note that while a different b parameter was chosen for each TF, this b factor is consistent across all five populations of cells, which is crucial for our later attempts to compare kinetic parameter values between cell populations. Even though the choice of b changes the absolute value of the kinetic parameters that are estimated, it has minimal effects on the strength of the linear correlation between the known and estimated values.
SABEC begins by assigning each cell to a random population 1 to K. Note that the total number of clusters K must be set at the start of the algorithm. Next, SABEC iteratively calculates the kinetic parameters of each of the K populations, and the cells are reassigned to new populations probabilistically.
In a standard Expectation Maximisation clustering algorithm, the probability of assigning a cell to a population is proportional to the ratio of the likelihoods of the cell coming from each population. Let the kinetic parameter set for a single population be Si(g) = (Kon(g), Koff(g), Kt(g)), where g is the gene (between 1 and G) and i labels the population (between 1 and K), and let x(g) be the vector of mRNA counts for each gene, for a particular cell. The log likelihood for a certain population can be calculated as follows:
L ( S i | x ) = ∑ g = 1 G ( ln ( P ( x ( g ) | S i ( g ) ) ) (7)
If the sum of the likelihoods for each population is Ltot, this value can be scaled as such:
L ′ ( S i | x ) = L ( S i | x ) L t o t (8)
Since it would take a long time for this algorithm to converge if the clusters are close to one another, we add a temperature parameter, so that initially it is easy for cells to be assigned to different clusters, but it becomes harder and harder to swap clusters over time:
L ″ ( S i | x ) = L ′ ( S i | x ) τ t (9)
where τ is the temperature parameter and t is the iteration number of the algorithm (a counter that increments each time the cells are reassigned to new clusters). A cell is probabilistically assigned to a new cluster based on the relative values of L″(Si|x) for each population. The algorithm terminates when fewer than 5% of the cells swap clusters in an iteration, or after 100 iterations. S4 Fig shows how the accuracy of the algorithm depends of the temperature parameter, τ.
Since this algorithm is randomized and since it is possible for certain runs of the algorithm to fall into local optima, we run this algorithm 50 times and conduct a secondary consensus clustering step. In this step, a consensus matrix is produced, in which each cell of the matrix represents the number of times that two cells are found in the same cluster. A plot of the cumulative density function of the values of this matrix can help visualise the robustness of the clustering (See Fig 8A). For comparison and to show the necessity for SABEC, consensus clustering of K-means was also conducted (See Fig 8B).
Fig 8 shows the cumulative density functions for the number of times two cells cluster together for SABEC (A) and K-means (B). [32] determined that the most robust metric for comparing the consistency of consensus clustering methods is the Proportion Ambiguously Clustered (PAC) score, which is defined as the proportion of cell pairs that cluster together in 10% and 90% of the repeated runs of the algorithm. This corresponds to the proportion of cell pairs that lie between the vertical lines in the cumulative distribution functions in Fig 8(A) and 8(B). The PAC scores for SABEC and K-means are compared in C, illustrating that SABEC usually has more consistent outcomes than K-means, with the fewest ambiguously clustered cells at K = 7.
In addition, to estimate the accuracy of our method, we can assume that the expected cluster assignment (as determined by FACS) is our gold standard. By comparing the results of our clustering approach with the gold standard, we can estimate the accuracy of our method on the experimental data. To do this, we cluster each of our consensus matrices into five clusters using partitioning around medoids (PAM), a clustering approach similar to K-means (but more consistent since it uses data points as centres). We can then compare our results to the gold standard labels using metrics such as VI (See Fig 8D) and the corrected Rand index (See Fig 8E). SABEC performs better than K-means by these two metrics, with the estimated number of clusters equal to 6. The one exception is that the corrected Rand index suggests that K-means performs slightly better than SABEC when K = 5; however, SABEC provides substantially better outcomes at K = 6. Based on simulated datasets with values similar to the experimental data, we determined that these latter two methods provide more accurate estimates of the number of clusters than the PAC method, which can sometimes overestimate the appropriate number of clusters (S8 Fig).
Figs 8F and 2G compare the consensus matrices for K = 6 for SABEC (F) and K-means (G), with the cells sorted by their FACS-determined labels. SABEC results appear more consistent, with cells frequently clustering with other cells of the same type. In addition, any cells that are “misclassified” tend to cluster instead with cell populations of their parent or children populations. For instance, some HSC cells cluster with their child populations (PreM and LMPP), and some GMP and CLP cells cluster with their parent population (LMPP).
The R script for SABEC is available in Github: https://github.com/ezer/SingleCellPipelineOverview, including a sample input file to run in parallel on a Condor cluster, and appropriately merge the outputs.
The SABEC method was compared to hierarchical clustering and K-means approaches. The hierarchical clustering approach used was the default one associated with the heatmap function in R (Euclidean distance metric and complete clustering). The K-means approach used the default algorithm in R (Hartigan and Wong), but the maximum iterations was increased to 100 in order to be more comparable to the SABEC approach. K-means was repeated 50 times and an additional consensus clustering step was taken, in order to provide a fairer comparison to SABEC. Note that the input to both the hierarchical clustering and K-means algorithms were the normalised Ct values, while the input to SABEC is the scaled mRNA counts. These algorithm and distance metrics were chosen since they are the most commonly used. Other variations of hierarchical clustering and K-means were tested, but none of the results were significantly better or different than the ones shown.
The R script for determining which kinetic parameters vary across populations is available in Github: https://github.com/ezer/SingleCellPipelineOverview.
Three methods were tested on simulated datasets of genes with randomly selected kinetic parameters. We conducted these simulation tests for two different ranges of the kinetic parameters to illustrate that there is different sensitivity to parameter changes in different regions of the parameter space (S12, S13 and S14 Figs). These results suggest that changes in Koff cannot be accurately detected when Koff > 5. In addition, the methods often performs better when fewer kinetic parameters change at once (S12 and S17 Figs). The correctly identified simulated cases were those that have the largest magnitude of kinetic parameter change (S18 Fig) and had average kinetic parameter values closer to the origin, where the kinetic parameters are more accurately estimated (S19 Fig).
Two of the methods (MP and Subset methods), have continuous-valued outputs, and so a threshold must be set for determining whether or not a change in a kinetic parameter is likely significant. The thresholds were set to have approximately a 2% false positive rate, based on the in silico validation tests. The threshold values for the MP method when Koff < 5 are −6.3, −8.5 and −6.8 for Kon, Koff and Kt, and −4.9 and −6.0 for Kon and Kt when Koff > 5. For the subset method, the thresholds are 0.77, 0.91 and 0.86 (Kon, Koff and Kt, respectively) when Koff < 5, and 0.681 and 0.870 (Kon and Kt) when Koff > 5.
|
10.1371/journal.ppat.1004242 | G3BP1, G3BP2 and CAPRIN1 Are Required for Translation of Interferon Stimulated mRNAs and Are Targeted by a Dengue Virus Non-coding RNA | Viral RNA-host protein interactions are critical for replication of flaviviruses, a genus of positive-strand RNA viruses comprising major vector-borne human pathogens including dengue viruses (DENV). We examined three conserved host RNA-binding proteins (RBPs) G3BP1, G3BP2 and CAPRIN1 in dengue virus (DENV-2) infection and found them to be novel regulators of the interferon (IFN) response against DENV-2. The three RBPs were required for the accumulation of the protein products of several interferon stimulated genes (ISGs), and for efficient translation of PKR and IFITM2 mRNAs. This identifies G3BP1, G3BP2 and CAPRIN1 as novel regulators of the antiviral state. Their antiviral activity was antagonized by the abundant DENV-2 non-coding subgenomic flaviviral RNA (sfRNA), which bound to G3BP1, G3BP2 and CAPRIN1, inhibited their activity and lead to profound inhibition of ISG mRNA translation. This work describes a new and unexpected level of regulation for interferon stimulated gene expression and presents the first mechanism of action for an sfRNA as a molecular sponge of anti-viral effectors in human cells.
| Dengue virus is the most prevalent arbovirus in the world and an increasingly significant public health problem. Development of vaccines and therapeutics has been slowed by poor understanding of viral pathogenesis. Especially, how the virus subverts the host interferon response, a powerful branch of the innate immune system remains the subject of debate and great interest. Dengue virus produces large quantities of a non-coding, highly structured viral RNA, termed sfRNA, whose function in viral replication is elusive but has been linked in related viruses to inhibition of the interferon response. Nonetheless the mechanisms involved are yet to be characterized. Here, we show that dengue virus 2 sfRNA targets and antagonizes a set of host RNA-binding proteins G3BP1, G3BP2 and CAPRIN1, to interfere with translation of antiviral interferon-stimulated mRNAs. This activity impairs establishment of the antiviral state, allowing the virus to replicate and evade the interferon response. While this particular mechanism was not conserved among other flaviviruses, we believe it is highly relevant for dengue virus 2 replication and pathogenesis. Taken together, our results highlight both new layers of complexity in the regulation of the innate immune response, as well as the diversity of strategies flaviviruses employ to counteract it.
| The critical roles of type I interferon (IFN) in detecting and clearing a wide range of viral infections have been well established [1]. IFNs are produced and released into the extracellular space by virtually all cell types upon recognition of pathogen-associated molecular patterns. Secreted IFNs act on the producing and neighboring cells to induce transcriptional activation of hundreds of antiviral IFN-stimulated genes (ISGs), establishing an antiviral state that rapidly targets viruses at various steps of their life cycle. While the transcriptional regulation of ISGs has been long defined, post-transcriptional events have recently emerged as critical regulators of the amplitude and specificity of the response. Regulation of mRNA stability [2], [3], translation [4], [5] or ubiquitination [6], [7] were shown to be critical for IFN-mediated antiviral effects. Nevertheless, the relative contribution of these post-transcriptional regulators and how they fine-tune the IFN system remain poorly understood.
Like other cytoplasmic RNA viruses, flaviviruses are highly sensitive to the antiviral effects of IFNs and as a result have evolved a wide array of countermeasures to avoid their action [8]. Described mechanisms include concealing double-stranded RNA replication intermediates in virally-induced ER membranes to decrease activation of innate immune sensors [9], cap methylation to mimic cellular mRNAs [10], degradation of regulators of IFN activation by the viral protease NS2B/3 [11], [12], or destabilization of transcription factor STAT2 by viral NS5 protein to dampen transcriptional activation of ISGs [13] More recently a ∼0.5 kb, abundant non-coding RNA derived from incomplete degradation of the viral 3′ untranslated region (3′UTR) by the cellular 5′-3′ exonuclease XRN1 and produced by all flaviviruses (termed sfRNA for subgenomic flaviviral RNA) was reported to be required for viral pathogenicity in a mouse model of the attenuated Kunjin strain (KUNV) of West Nile virus (WNV) [14]. Follow-up studies determined that KUNV sfRNA counteracted IFN antiviral activity [14], [15]. Strikingly, while the majority of genomes synthesized during infection are processed into sfRNA, it is dispensable for RNA replication in IFN-incompetent cells, arguing for an important, conserved role in antagonizing immune defenses. The only possible mechanism for the anti-IFN action of the sfRNA was suggested by a recent report that suggests the Japanese encephalitis virus (JEV) sfRNA inhibits IFN production by blocking the phosphorylation of IRF-3 [16]. The observation suggesting a decrease in IFN production by transfecting JEV sfRNA was not properly controlled by the use of other RNAs and thus we believe that to date the anti-IFN mechanism of sfRNA remains unknown.
As the flaviviral positive-strand 11 kb RNA genome encodes only 10 viral proteins, it is not surprising that host proteins, especially RNA binding proteins (RBPs), play a critical role in viral replication and pathogenicity. In a screen for host proteins interacting with dengue virus 2 (DENV-2) RNA, we identified ubiquitous, multifunctional RBPs, G3BP1, G3BP2 and CAPRIN1 [17]. G3BP1 and CAPRIN1 had been reported as proviral factors in vaccinia virus (VACV) and respiratory syncytial virus (RSV) infection [18], [19]. On the other hand, G3BP1 and G3BP2 had antiviral activity against poliovirus (PV) and alphaviruses [20], [21], suggesting a variety of possible mechanisms of action in viral infections. In this study, we investigated their role in DENV-2 infection. We found that these proteins have a potent antiviral action against flaviviruses, linked to a previously unknown role in regulating translation of ISG mRNAs. We further demonstrate that this activity is targeted by DENV-2 sfRNA, which binds G3BP1, G3BP2 and CAPRIN1 and prevents their function, protecting viral replication against IFN-mediated antiviral effects.
G3BP1, G3BP2 and CAPRIN1 were initially discovered as interacting with DENV-2 3′UTR in an RNA affinity chromatography screen performed in our laboratory [17]. The variety of cellular functions described for these proteins in control of mRNA translation and stability, regulation of cell signaling pathways, and in the integrated stress response [22]–[26] prompted us to examine their role in DENV-2 infection.
RNAi-mediated knockdown and overexpression studies indicated that G3BP1, G3BP2 and CAPRIN1 had a modest but significant antiviral activity against DENV-2 NGC in HuH-7 cells (Figure S1A–D). Importantly, transfection of control and specific siRNAs did not induce IFN mRNA accumulation, which could have influence DENV-2 replication as an off target effect, [27] (Figure S1E). The antiviral effect was not restricted to the laboratory adapted DENV-2 NGC as G3BP1, G3BP2 and CAPRIN1 had antiviral activity against the clinical DENV-2 isolate PR1940 as well as the related yellow fever vaccine strain (YFV-17D) (Figure S1F). Nonetheless, G3BP1, G3BP2 and CAPRIN1 depletion had no effect on the DENV-2 strain PR6913, which exhibited low replication levels (Figure S1F). Since the type I IFN response has been long established as a critical mediator of anti-flaviviral innate immunity [8] and low levels of flaviviral replication have been shown to correlate with lower induction of type I IFN response [28] the data above suggested that G3BP1, G3BP2 and CAPRIN1 could play a previously unidentified role in the innate immune response.
In order to investigate a link between IFN-mediated antiviral activity and that of G3BP1, G3BP2 and CAPRIN1, cells depleted of these three proteins (Figure 1A) were pretreated with IFN-β before infection with DENV-2 NGC. As in previous studies [29], 100 UI/ml IFN-β, which is comparable to that observed in sera of DENV infected patients [30], abrogated formation of viral replication complexes (Figure 1B). Strikingly, the IFN effect was dramatically diminished in G3BP1, G3BP2 and CAPRIN1-depleted cells (Figures 1B), indicating that G3BP1, G3BP2 and CAPRIN1 are required for IFN-β antiviral effects.
Next, DENV-2 infectivity was measured over a range of IFN concentrations, revealing that treatment with two independent sets of siRNAs targeting G3BP1, G3BP2 and CAPRIN1 (siG12C#1 and siG12C#2) resulted in a 4- to 5-fold increase in IFN-β IC50 (Figure 1C). This effect was even more pronounced at the level of infectious progeny virus formation and accumulation of viral RNAs (Figures 1D and 1E), consistent with IFN-β targeting various steps of the viral life cycle. Notably, the magnitude of the G3BP1, G3BP2 and CAPRIN1-mediated antiviral activity in the presence of IFN-β was much greater than its ∼2-fold effect in the absence of IFN-β, (Figure S1B), suggesting that the main antiviral role of these proteins occurs through IFN action. It should be noted that HuH-7 cells secrete low levels of IFN-β upon DENV-2 infection ([31] and see below), and therefore conditions without exogenously added IFN-β should be consider to have low levels of IFN. Interestingly, G3BP1, G3BP2 and CAPRIN1 antiviral activity was redundant since depletion of all three proteins was required to observe an effect on DENV replication (Figure S2), which was consistent with previous studies on other functions of these RBPs [18], [21], [25], [32]. The requirement of G3BP1, G3BP2 and CAPRIN1 in IFN-mediated antiviral activity was observed in YFV-17D infection (Figure 1F). Taken together, these data indicate that G3BP1, G3BP2 and CAPRIN1 have an unexpected and important role in mediating the anti-flaviviral activity of IFNs.
We next investigated the mechanism of action of this unexpected role of G3BP1, G3BP2 and CAPRIN1 in the IFN response. Since G3BP1, G3BP2 and CAPRIN1 had previously been determined to be critical regulators of stress granules (SG) assembly, another aspect of the cellular response to infection, we examined SG formation in infected cells and upon IFN treatment. Indeed, a recent study of SG dynamics in HCV infection revealed that treatment of HCV-infected cells with IFN-α triggered potent SG formation [33], suggesting that SG could mediate IFN antiviral effects. We monitored SG formation during DENV-2 infection in the presence or absence of IFN-β and found no evidence of increased SG formation in DENV-2 infected cells, even following treatment with IFN-β that effectively reduced viral replication (Figure S3). These data indicate that bona fide SG formation, which is defined microscopically by the appearance of cytoplasmic granules containing a set of protein markers, was not required to mediate the IFN anti-DENV-2 effects. Our data do not exclude an important role for SG and SG-associated proteins in DENV-2 infection, but they suggested that G3BP1, G3BP2 and CAPRIN1 could play an alternative role in the IFN response to DENV-2 infection.
To gain insight into such alternative modes of action, we examined the integrity of the pathway leading to the establishment of the IFN induced antiviral state. Binding of IFNs to their receptor on the cell surface activates the JAK-STAT signaling cascade, leading to the transcriptional activation of hundreds of IFN-stimulated genes (ISGs), which have specific antiviral activities [34]. We selected a representative panel of ISGs and measured their induction in response to IFN-β in control versus G3BP1, G3BP2 and CAPRIN1 depleted HuH-7 cells. IFN-inducible IFITM2, RIG-I/DDX58 ISG15 and STAT1 have been reported to have anti-DENV-2 activity; however, the dsRNA-activated kinase PKR/EIF2AK2 and MX1 do not affect DENV-2 replication [35]–[38]. Quantitative real-time RT-PCR analysis showed no significant decrease of control mRNA levels (GAPDH or ACTINB) or ISG mRNA induction in response to increasing IFN-β concentration in G3BP1, G3BP2 and CAPRIN1-depleted cells (Figures 2A to 2D and S4A to S4D). Strikingly, expression of all six ISG proteins, however, was significantly reduced in G3BP1, G3BP2 and CAPRIN1-depleted cells (Figures 2E, 2F and S4E).
PKR protein accumulation, as measured by densitometry analysis of the western blot data shown in Figure 2E, was reduced 4.05 and 3.21 fold compared to siGFP control in siG12C#1 and siG12C#2-treated cells, respectively. In this experiment, RIG-I accumulation was reduced 1.86 and 1.80 fold, and ISG15 2.15 and 1.60 fold respectively. STAT1 protein levels were reduced 4.50 and 4.98 fold, and MX1 2.46 and 11.2-fold respectively in cells depleted of G3BP1, G3BP2 and CAPRIN1 (Figure S4). Because determination of protein levels from HRP-based western data is semi-quantitative, we performed analysis of IFITM2 levels using fluorophore-conjugated secondary antibodies in the Licor Odyssey system. Quantification of western blots by fluorescence intensity in three independent experiments revealed a 3- to 5-fold reduction in IFITM2 levels, normalized to ACTINB levels, after stimulation with 100 or 1000 UI/ml IFN-β (Figure 2F). These data indicate a general and robust effect of G3BP1, G3BP2 and CAPRIN1 on establishment of the antiviral state through post-transcriptional control of multiple ISGs. We propose that many more ISGs will be affected, and the cumulative effect would likely explain the dramatic drop in IFN antiviral potency observed in the previous experiments.
Given our data above it was possible for G3BP1, G3BP2 and CAPRIN1 to influence ISG mRNA splicing, transport or translation, or ISG protein stability. Because of previous reports on these proteins [23], [24], [26], [39], [40], we first addressed their action on ISG mRNA translation. We focused on ISGs IFITM2 and PKR to delineate the role of G3BP1, G3BP2 and CAPRIN1. Using 35S metabolic labeling, we established that depletion of G3BP1, G3BP2 and CAPRIN1 depletion did not downregulate global cellular translation (Figure 3A). Consistent with a lack of a profound global effect on translation, polyribosome fractionation revealed no significant difference in rRNA profiles in G3BP1, G3BP2 and CAPRIN1-depleted cells (Figure 3B). Polyribosome association profiles of several cellular mRNAs (ELF2, GAPDH and BIP/GRP78 mRNAs) showed minimal or no change upon G3BP1, G3BP2 and CAPRIN1 depletion and IFN-β treatment (Figures 3C–E). Indeed, although GAPDH and BIP mRNA slightly shifted to lighter fractions, the fraction containing the majority of the mRNA remained unchanged, indicating that association of these mRNAs with polyribosomes was not dramatically affected. Polyribosome-association of IFITM2 and PKR mRNAs, however, was strongly impaired in the absence of G3BP1, G3BP2 and CAPRIN1 (Figures 3F–G), suggesting that the strong effect of G3BP1, G3BP2 and CAPRIN1 depletion was specific for ISG mRNA translation. Importantly, depletion of G3BP1, G3BP2 and CAPRIN1 did not affect polyribosome-association of DENV-2 genomic RNA (Figure 3H), supporting the previous hypothesis that G3BP1, G3BP2 and CAPRIN1 antiviral action is principally mediated by the IFN system rather than by a direct role in viral translation.
In order to confirm the specificity of G3BP1, G3BP2 and CAPRIN1 in regulating ISG mRNA translation, we established stable cell lines expressing firefly luciferase reporters under the transcriptional control of a minimal promoter and an ISRE to provide IFN induction, and including the ELF2, GAPDH, IFITM2 or PKR UTRs with the first and last 30 nucleotides of the coding sequence (ELF2-Fluc, GAPDH-Fluc, IFITM2-Fluc and PKR-Fluc Figure 4A). We observed that IFN-induction of GAPDH-Fluc, IFITM2-Fluc and PKR-Fluc mRNAs was modestly reduced, although the effect was not significant for IFITM2-Fluc and PKR-Fluc. The induction of ELF2-Fluc mRNA was robustly and significantly reduced in the absence of G3BP1, G3BP2 and CAPRIN1 (Figures 4B–E). However, the significance of these observations is not clear since depletion of G3BP1, G3BP2 and CAPRIN1 did not affect absolute levels of endogenous GAPDH, IFITM2 or PKR mRNAs (see Figures 2 and 3). Importantly, IFN induction of ELF2-FLuc luciferase activity was not inhibited by knockdown of G3BP1, G3BP2 and CAPRIN1, however, IFN induction of IFITM2-Fluc and PKR-Fluc activity was robustly inhibited (Figure 4F, H and I). While induction of GAPDH-Fluc activity was significantly reduced in these conditions, the effect could be fully explained by the aforementioned effect on GAPDH-Fluc mRNA levels (compare Figures 4C and G). Indeed, a calculation of the relative translation efficiency, the ratio of protein induction (derived from luciferase activity) relative to mRNA induction, clearly revealed that both ELF2 reporter translation was increased 6.3-fold in G3BP1, G3BP2 and CAPRIN1-depleted cells compared to control siGFP, while the relative translation efficiency of IFITM2-Fluc and PKR-Fluc was reduced 2.54- and 1.3-fold, respectively (Table 1). In the case of GAPDH-Fluc, the relative translation efficiency was slightly reduced (1.07-fold), which correlates with the modest shift observed in GAPDH mRNA distribution in polyribosomes fractions in the absence of G3BP1, G3BP2 and CAPRIN1 (see Figure 3E).
Taken together, these results show that G3BP1, G3BP2 and CAPRIN1 differentially affect reporter mRNA translation and that elements in the IFITM2 and PKR mRNA UTRs and/or the first and last 30 nucleotides of their coding sequence render translation of these messengers specifically dependent on G3BP1, G3BP2 and CAPRIN1. This suggest that G3BP1, G3BP2 and CAPRIN1 can, as previously described in the literature, play various roles in cellular mRNA metabolism, but are specifically required for translation of ISG mRNAs. While the precise mechanisms of translational regulation and how these proteins achieve selectivity remain to be investigated, several hypotheses will be proposed in the discussion.
Flaviviruses, like other viruses, have been reported to interfere with the host IFN response by hijacking a large variety of cellular factors required for establishment of the antiviral state. In the case of DENV-2, all previously described evasion strategies affect signaling pathways upstream of ISG transcriptional activation [11], [13], [41], [42] However, these mechanisms are not completely efficient since ISG mRNA upregulation is observed widely in response to DENV-2 infection [43]. Data presented above suggested that ISG mRNA translation could be targeted by DENV-2 and this would not have been detected in previous studies measuring ISG mRNA induction as a surrogate for efficient IFN response.
In DENV-2 infected cells, viral RNA replication resulted in a 24-fold increase in IFN-β mRNA between 24 and 48 h post infection, which was accompanied by a 7-fold increase in IFITM2 mRNA (Figures 5A to 5C). No induction of IFITM2 was detected up to 72 h post infection (Figures 5D and 5E), indicating that ISG expression was indeed controlled at a post-transcriptional level during DENV-2 infection. Importantly, polyribosome fractionation analysis showed that in DENV-2 infected cells and in G3BP1, G3BP2 and CAPRIN1-depleted cells, IFITM2 and PKR mRNA translation was strongly impaired (Figures 5F to 5I). Interestingly, DENV-2 infection inhibited IFN induction of both PKR mRNA and protein (Figure S5), indicating that this virus can regulate some ISGs via multiple mechanisms to keep the IFN response under check. Most importantly the data indicate that DENV-2 interfered with IFITM2 and PKR mRNA translation, which phenocopied G3BP1, G3BP2 and CAPRIN1 depletion, and suggested that DENV-2 gene product(s) target these RBPs.
Previously G3BP1 had been reported to be antagonized in poliovirus infection, where it is cleaved by a viral protease [20]. DENV-2 infection however, did not decrease the levels of G3BP1, G3BP2 or CAPRIN1 (Figure S5), suggesting a mechanism other than proteolytic degradation. We have shown previously that G3BP1, G3BP2 and CAPRIN1 each interact with the 3′UTR of DENV-2 RNA [17], a region included in the 3′UTR-derived non-coding sfRNA. These interactions, together with the fact that the sfRNA from a related flavivirus, Kunjin virus (KUNV), interferes with the IFN response [15], made DENV-2 sfRNA an ideal candidate for targeting G3BP1, G3BP2 and CAPRIN1. Therefore, we hypothesized that DENV-2 sfRNA would bind G3BP1, G3BP2 and CAPRIN1 and inactivate their antiviral effect.
In order to determine whether DENV-2 sfRNA interacts with G3BP1, G3BP2 and CAPRIN1, we first performed co-localization experiments using in situ hybridization for DENV-2 RNAs and immunofluorescence for G3BP1 in infected cells. In situ probes detecting the viral 5′UTR, which interrogate only the gRNA, and 3′UTR, which detect both gRNA and sfRNA, were both found to colocalize with G3BP1 during infection (Figure S6). To test and quantify an interaction between viral RNAs and the three RBPs, we used RNA-immunoprecipitation and a real-time PCR strategy designed to discriminate between gRNA and sfRNA, which is identical to the last 428 nucleotides of the genome [44] (Figures 6A and S7A–E). As suggested previously [44], we found that DENV-2 sfRNA was 5–10 times more abundant than the gRNA during infection of HuH-7 cells (Figure S7G). Both DENV-2 gRNA and sfRNA were found enriched in G3BP1-immunoprecipitates from infected cells (3- and 6-fold relative to GAPDH RNA, respectively), but were not found to interact with KSRP, an unrelated host RBP (Figures 6B and 5C). DENV-2 gRNA and sfRNA were also enriched in G3BP2 and CAPRIN1 immunoprecipitates, while c-Myc mRNA was not enriched in these (Figure S8), confirming the specificity of the interaction. Taken together, these data indicate that G3BP1, G3BP2 and CAPRIN1 interact with DENV-2 gRNA and sfRNA during infection.
Having established that G3BP1, G3BP2 and CAPRIN1 interacted with DENV-2 sfRNA in infected cells, we sought to examine which sequence or structural elements were required for this interaction. The DENV-2 3′UTR contains a series of highly conserved secondary structures (Figure 6A), which have been proposed to serve as platforms of interaction for host RBPs [17] [45]. We designed sfRNA variants containing various deletions and point mutations (Figure 6D) and tested their ability to interact with G3BP1, G3BP2 and CAPRIN1 by RNA affinity chromatography. We found that stemloop II (SL-II), but not the predicted pseudoknot PKSL-II, was required for G3BP1, G3BP2 and CAPRIN1 binding to DENV-2 3′UTR (Figure 6E). We also tested the ability of the 3′UTR of related flaviviruses to interact with these RBPs and observed that only 3′UTRs of clinical isolates from DENV-2, but not DENV-3, the attenuated WNV subtype KUNV or the YFV vaccine strain 17D were able to pull-down G3BP1, G3BP2 and CAPRIN1 (Figure S9). Notably, in these mutants and isolates, the three proteins shared the same binding requirements, suggesting that these proteins interact as a complex.
In order to determine whether DENV-2 sfRNA was able to inhibit G3BP1, G3BP2 and CAPRIN1 activity and impair ISG expression, we transfected increasing amounts of DENV-2 3′UTR RNAs, as sfRNA-mimics, into cells and measured ISG mRNA and protein expression upon IFN-β treatment. To control for effects mediated by functions of the sfRNA unrelated to G3BP1, G3BP2 and CAPRIN1 binding, we constructed a mutant unable to bind these RBPs but containing all other sequence elements of DENV-2 sfRNA. Since the structure required for binding, SL-II, has been implicated in formation and stability of flaviviral sfRNAs [46], we sought to minimize the effects of its deletion by replacing it with the equivalent structure from YFV-17D, SLE (Figures 7A and 7B), which did not interact with G3BP1, G3BP2 and CAPRIN1 in vitro (Figure S9 and Ward et al, unpublished data). This hybrid mutant, DENV-2 3′UTR YFSLE, exhibited 5-fold decreased binding to G3BP1 (Figure 7C), and additional point mutations in DENV-2 SL-IV, whose secondary structure resembles SL-II (indicated on Figure 7B), further decreased the interaction to background levels (DENV-2 3′UTR YFSLE-ST4, Figure 7C).
When increasing amounts of in vitro transcribed RNAs were transfected into cells followed by treatment with 100 UI/ml IFN-β, we observed that DENV-2 3′UTR, but not the control DENV-2 3′UTR YFSLE RNA, was able to decrease in a dose-dependent manner expression of ISGs IFITM2 (Figures 7D to 7G) and PKR (Figure S10). Both RNAs accumulated to similar levels and had no effect on ISG mRNA induction levels (Figures 7D, 7E and S10), indicating that DENV-2 sfRNA is able to post-transcriptionally interfere with ISG expression and that this activity depended on G3BP1, G3BP2 and CAPRIN1 binding. As observed for G3BP1, G3BP2 and CAPRIN1 depletion, ectopic expression of DENV-2 3′UTR interfered with ISG mRNA association with polyribosomes, while GAPDH and ELF2 mRNA were minimally or not affected in the same conditions (Figure S11). Taken together, these data show that ectopic expression of DENV-2 sfRNA mimics inhibited IFITM2 and PKR mRNA translation through G3BP1, G3BP2 and CAPRIN1 binding.
We showed that interaction of DENV-2 sfRNA with G3BP1, G3BP2 and CAPRIN1 was able to downregulate expression of ISGs, which is consistent with the sfRNA acting as a decoy for these host RBPs. To further test this hypothesis and determine the importance of this mechanism during infection and for viral evasion of the IFN response, we constructed mutant DENV-2 replicons unable to sequester G3BP1, G3BP2 and CAPRIN1. We used the established DENV-2 replicon system [47], whose biphasic reporter activity examines translation of input RNAs and subsequent replication and translation steps independently, to evaluate the effect of G3BP1, G3BP2 and CAPRIN1 binding on translation, replication and sensitivity to inhibition by IFN-β. We modified the DENV-2 replicon 3′UTR deleting the SL-II and introducing point mutations in SL-IV described before (D2Rep-dSLII-ST4, Figure 8A, S12A) and confirmed that replicon RNAs bearing these mutations had reduced binding to G3BP1 (Figure 8B). While SLII was reported to be required for sfRNA formation in some flaviviruses, we did not measure a decrease in sfRNA formation in dSLII-ST4 mutants (Figure S12B), ruling out the possibility that the effect of the mutation could be linked to SL-II functions mediated by other regions of the sfRNA. Indeed this finding is consistent with in vivo results in the recent report by Liu et al [48]. We electroporated these reporters into HuH-7.5 cells, which were derived from HuH-7 cells and harbor a point mutation in RIG-I that renders them deficient in IFN production through this pathway [49] and the parental HuH-7 cells. Importantly, we detected no difference in luciferase activity between D2Rep-WT and D2Rep-dSLII/ST4 at any time after electroporation (Figure 8C), indicating that reduced binding to G3BP1, G3BP2 and CAPRIN1 had no effect on replicon translation or replication. In HuH-7 cells the dSLII-ST4 mutation did not alter very early luciferase activity, a measure of translation of input RNAs (Figure S12C), however luciferase activity of the D2Rep-dSLII/ST4 was reduced by an average of 4.4-fold compared to the WT replicon at 72 h post-electroporation (Figure 8D). The different effects in HuH-7.5 and HuH-7 cells suggested that G3BP1, G3BP2 and CAPRIN1 binding to viral RNAs was required for viral replication in the context of a functional innate immune response.
To examine the effect of adding exogenous IFN on D2Rep-WT and D2Rep-dSLII-ST4 activity we electroporated these in HuH-7 cells and treated these with 50 UI/ml IFN-β at 4 h post-electroporation. The modest deleterious effect of the dSLII/ST4 mutation was strikingly enhanced by IFN treatment, with luciferase activity reduced 16.9-fold compared to the D2Rep-WT at 72 hr post-electroporation (Figure 8E). Overall, addition of exogenous IFN-β inhibited D2Rep-WT activity by 3.2-fold at 72 h post-electroporation while D2Rep-dSLII-ST4 was inhibited 12.3-fold (Figure 8F), indicating that the dSLII/ST4 mutation renders replicons more sensitive to the antiviral effects of IFNs. Finally, we analyzed the rates of accumulation of luciferase reporter between 48 and 72 h post electroporation. We observed no significant difference between the rates of D2Rep-WT and D2Rep-dSLII/ST4 in the absence of exogenously added IFN-β; however the D2Rep-dSLII/ST4 was significantly impaired in the presence of exogenously added IFN (Figure 8G). On the one hand, in the presence of low levels of endogenous IFN, which we expect with HuH-7 but not HuH-7.5 cells, after an initial delay the mutant replicon is still able to surmount IFN-mediated inhibition. On the other hand in the presence of higher levels of IFN the D2Rep-dSLII-ST4 is persistently inhibited. The results above convincingly argue that anti-DENV-2 action of G3BP1, G3BP2 and CAPRIN1 is mediated by their pro-IFN activity and support the hypothesis that the DENV-2 sfRNA antagonizes the IFN response in part by sequestering these host RBPs.
In this study we make two new and important observations in the understanding of host innate antiviral measures and their inhibition by viral countermeasures. First, we identified G3BP1, G3BP2 and CAPRIN1, three conserved, multifunctional RNA-binding proteins, as critical positive regulators of the antiviral IFN response. This unexpected role was mediated through the specific activation of antiviral ISG mRNA translation. Second, we described the DENV-2 sfRNA as an antagonist to their antiviral effect, providing the first mechanism of action for this abundant, non-coding flaviviral RNA (Figure 9).
Although G3BP1, G3BP2 and CAPRIN1 have been shown to have a large variety of cellular functions, this report associates them for the first time with innate immunity. We show that these three RBPs were required for an antiviral IFN response against several isolates of DENV-2 and YFV-17D. Our data indicate that G3BP1, G3BP2 and CAPRIN1 regulate the expression of ISGs known to have broad antiviral activity: PKR, RIG-I, IFITM2, ISG15, STAT1 and MX1 [34]. Therefore, while the full spectrum of ISG targets and the individual contributions of the three RBPs remain to be determined, we posit their antiviral activity will be conserved against a wide array of viruses. Indeed previous evidence suggested this: the poliovirus (PV) protease degrades G3BP1 [20]; the core protein of Japanese encephalitis virus (JEV), another flavivirus, was identified as an important CAPRIN1 antagonist [50], and the nsP3 protein of Chikungunya virus (CHIKV), an alphavirus, as G3BP1 and G3BP2 opponent [51]. All these interactions were shown to be required for optimal viral replication, supporting our conclusion that G3BP1, G3BP2 and CAPRIN1 are major regulators of the cellular immune response.
The fact that these RBPs were not previously identified as IFN-related antiviral factors can be explained by two reasons. First, previous studies usually focused on SG formation and were performed in absence of exogenously added IFN. Second, the role of these proteins was examined independently, in ways that would not unearth their redundant functions in the IFN system. Interestingly, the direct antiviral role of SG formation is intuitive but has not been formally demonstrated given the challenges in differentiating the role of the granules themselves from the role of their numerous individual components. While the relative contributions and connections of these two branches of the innate immune response remain to be determined, our study suggests that activity of G3BP1, G3BP2 and CAPRIN1 against DENV-2 is primarily through the IFN system.
Perhaps our most unexpected finding was that G3BP1, G3BP2 and CAPRIN1 are critical for ISG mRNAs translation, a step previously understudied in the IFN response. Although the dogma is that establishment of the IFN-mediated antiviral state is primarily controlled by transcriptional activation, recent evidence suggests that additional layers of control regulate the amplitude and specificity of the response. For instance, a screen for host proteins implicated in IFN-mediated inhibition of hepatitis-C virus identified a large number of splicing factors [52]. This result implicates post-transcriptional mechanisms, which could regulate splicing or the proteins could moonlight in other aspects of RNA metabolism. Control of the stability of IFN-β mRNA by KSRP and STAT mRNA by PCBP2 were equally able to modulate IFN-mediated inhibition of viral replication [2], [3]. A recent study implicates, but does not directly address, the importance of ISG translational regulation [53] in antiviral signaling and underscores the importance of our findings.
The precise mode of action of G3BP1, G3BP2 and CAPRIN1 in ISG translational regulation and especially how specificity for ISG mRNAs is achieved remain to be elucidated. Several hypotheses could be considered. G3BP1, G3BP2 and CAPRIN1 could bind to ISG mRNA UTRs and recruit translation initiation factors, recruit ISG mRNAs to subcellular localizations where translation is more efficient in conditions of stress, or relieve miRNA-mediated inhibition of ISG mRNA translation. The RBPs could also act indirectly either activating or repressing mRNAs coding for positive or negative regulators of ISG mRNA translation. Alternatively, G3BP1, G3BP2 and CAPRIN1 could modulate signaling events leading to translational activation in the IFN response, such as the PI3K/Akt or Mnk pathways that are required for ISG mRNA translation [5], [54]. Finally, the RBPs could be involved in a stress response induced by IFNs that while not inducing bona fide SG would generally repress many mRNAs and by mass action enhance the translation of ISG mRNAs. In all above scenarios though, the cis-acting elements in the ISG mRNA UTRs conferring dependency on G3BP1, G3BP2 and CAPRIN1 will be a critical feature to determine.
In the second part of our study, we show that the DENV-2 abundant, non-coding sfRNA interacts with G3BP1, G3BP2 and CAPRIN1, inactivates them and thus mediates inhibition of ISG expression. The sfRNA - G3BP1, G3BP2 and CAPRIN1 interaction that we propose as a decoy mechanism was conserved for DENV-2 clinical isolates indicating its potential relevance for DENV-2 pathogenicity. While the antagonism of the immune response by viral non-coding RNAs has been well described, few mechanisms of action have been uncovered. Sequestration of host proteins has been widely hypothesized but only in a few instances was it formally demonstrated. The adenovirus VA RNAs was shown to bind and antagonize PKR and a similar role was proposed for Epstein Barr virus EBER RNAs [55]–[57]; the Sendai virus trailer RNA was hypothesized to sequester the RBP TIAR to subvert apoptosis [58]; the interaction between Kaposi's sarcoma-associated herpesvirus (KSHV) PAN RNA and PABP was suggested to participate in the host translational shutoff effect [59], [60]. Here we demonstrate a role for DENV-2 sfRNA as a molecular sponge or decoy for G3BP1, G3BP2 and CAPRIN1 resulting in a crippled IFN response.
While the sfRNA-G3BP1, G3BP2 and CAPRIN1 interaction was conserved for all DENV-2 viruses tested, no binding was detected for DENV-3, KUNV or YFV-17D 3′UTR. This suggests that although the IFN antagonist action of the sfRNA is conserved among flaviviruses, the precise mechanisms diverge for different viruses [15]. This is not unexpected since specific tactics for viral evasion of the IFN response have been shown to vary widely between related viruses, even strains of the same virus [8], [34]. For instance, NS4B proteins from some DENV-2 clinical isolates, but not from others, were able to interfere with IFN signaling [61]. Furthermore the sfRNA includes the so-called variable region (VR), which, while generally conserved in RNA secondary and tertiary structure, diverges significantly in primary sequence among flaviviruses. The VR is therefore a propitious platform for rapid evolution of new host RBP binding sites providing this viral genus with a wide array of tactical solutions to counter host innate defenses. It is thus conceivable that KUNV sfRNA, although not binding to G3BP1, G3BP2 and CAPRIN1, could target different subsets of host RBPs to prevent establishment of the antiviral state.
To conclude, it is widely accepted that host immune measures and pathogen countermeasures evolve rapidly, leading to remarkable diversity on both sides. Here, we propose that RBPs such as G3BP1, G3BP2 and CAPRIN1 are critical mediators of the antiviral state and that antagonizing them is a strategy employed by many viruses, including DENV-2. Equally, we believe that targeting different subsets of host RBPs is a pan-flaviviral anti-IFN strategy, for which many targets remain to be uncovered.
HuH-7 hepatocellular carcinoma cells were maintained in DMEM supplemented with 10% FBS. BHK-21 cells, which were used for virus titration, were maintained in RPMI supplemented with 10% FBS. DENV-2 strain NGC and YFV-17D were propagated in Aedes albopictus C6-36 cells. All infections were carried at a multiplicity of infection of 1 (MOI = 1) for 24 h unless otherwise indicated. Infectivity was measured using indirect immunofluorescence detection of viral antigens or dsRNA in infected cells, quantitative real-time RT-PCR analysis of viral genomes, or quantification of infectious particles released by focus forming assay, as previously reported [17].
25 nM of the indicated siRNA duplexes or 75 nM control siRNA (see supplementary materials) were transfected into cells at 50% confluency twice at 48 hr intervals (day 1 and 3) with Lipofectamine RNAiMax (Invitrogen) following manufacturer's instructions. Human IFN-β (PBL Interferon Source) was added to cells 24 hrs post-transfection and incubated for 16 h prior harvesting or infection with DENV-2. Lysates were collected 24 hrs post infection (day 6) and analyzed by western blotting for knockdown efficiency and ISG expression, and quantitative real-time RT-PCR for RNA levels (see supplementary methods).
Polyribosome fractionation was performed as previously described [62] with minor modifications: cells were harvested by trypsinization and 50 µg/ml cycloheximide was added into polyribosome lysis buffer. Individual mRNA levels in each fraction were measured by quantitative real-time RT-PCR and expressed as percentage of total for this mRNA in all the gradient fractions.
pcDNA3.1 constructs containing the firefly luciferase open reading frame flanked by IFITM2, PKR, GAPDH, or ELF2 5′ and 3′UTRs and driven by an ISRE promoter (for cloning details refer to supplementary materials) were transfected in HuH-7 cells and selected for stable expression in DMEM supplemented with 1500 µg/ml G418 (Gibco). siRNA-mediated knockdown was performed as described and cells stimulated with 1000 UI/ml IFN-β for 10 h. Firefly luciferase activity was assessed using the Dual luciferase reporter assay system (Promega). Firefly luciferase mRNA levels were measured by quantitative real-time RT-PCR.
RNA-immunoprecipitations were performed using the MAGNA-RIP kit (Millipore) following manufacturer's recommendations. The levels of RNA in IP were determined by quantitative real-time RT-PCR and normalized to GAPDH mRNA levels and control rabbit IgG IP following the formula:Tobramycin RNA affinity chromatography was carried out as described previously [17].
DENV-2 gRNA and sfRNA levels were quantified using a differential quantitative real-time RT-PCR assay designed based on the sfRNA mapping in Liu et al [44] (see Figure 5 and S7). One primer, annealing upstream of the stop codon in which one pair of primer recognizes specifically gRNA while a second pair of primers amplifies sequences shared between gRNA and sfRNA. Briefly, RNA extracted from experimental samples was reverse transcribed and parallel reactions set-up. Primer QG-FOR (5′ CCATGAAAAGATTCAGAAG 3′, annealing upstream of the stop codon) was used to detect gRNA only while primer QGSF-FOR (5′ GTG AGC CCC GTC CAA GG 3′, annealing downstream of the start of sfRNA) detected both gRNA and sfRNA. The reverse primer QGSF-REV (5′ GCTGCGATTTGTAAGGG 3′ annealing downstream of DB2) was shared, leading to products of 309 and 184 bp, respectively. In order to determine the relative sfRNA/gRNA ratio in a given sample, 1–2 µg of total cellular RNA (1–10 ng of in-vitro transcribed RNA) were incubated at 70°C for 5 min and reverse transcribed using the ImPromII kit (Promega) following manufacturer's recommendation. Triplicate wells containing 100–200 ng of cDNA, 300 pmol of each primer (QG-For or QGSF-For and QGSF-Rev) and Biorad SYBR Green reagent following manufacturer's recommendation were set up in a total of 25 µl. Reactions were run on a Biorad CFX96 quantitative real-time RT-PCR with the following parameters: 90°C 5 min, 40 repeats of 90°C for 30 s, 55°C for 30 s and 72°C for 30 s. Fluorescence detection was performed during the 72°C elongation step at each cycle. For each reaction the molar amount of template (n(G) and n(GSF)) was calculated from the CT value using a standard curve generated from serial dilutions of reverse transcribed purified full-length D2Rep RNA. sfRNA levels were inferred by subtracting the molar amount n(GSF) – n(G). A similar strategy was designed for analysis of YFV-17D gRNA and sfRNA levels, in this case primers were based on sfRNA mapping in Silva et al [63]. (YFV-G-For 5′ GGATGGAGAACCGGACTCC 3′, YFV-GSF-For 5′ GCTAAGCTGTGAGGCAGTGC 3′, YFV-GSF-Rev 5′ CGTCTTTCTACCACCACGTG 3′).
Templates for synthesis of control DENV-2 3′UTR YFSLE, in which the DENV-2 SL-II sequence (DENV-2 nt 10306–10348) was replaced by YFV 17D SLE sequence (YFV17D nt 10530–10611), were custom synthesized by GenScript. DENV-2 3′UTR and 3′UTR YFSLE templates were PCR amplified from stock plasmids to add a T7 promoter immediately upstream of the DENV-2 stop codon (T7-VR-For), and in vitro transcribed using the MegaScript kit (Ambion). 10, 100 or 1000 ng/ml RNA were transfected in cells at 50% confluency using Lipofectamine RNAiMax for 4 h. Cells were washed and incubated with complete medium containing 100 UI/ml IFN-β for 4 h before analysis of protein and mRNA contents.
The DENV-2 reporter replicon system (D2Rep), based on DENV-2 strain 16681 (U87411.1), has been described before [47]. Detailed experimental procedures are available in supplementary materials.
All results are presented as mean ± SEM of at least 3 independent experiments, unless otherwise indicated. Data were analyzed using unpaired, two-tailed Student's t-test and considered significant if p<0.05 (*p<0.05; **p<0.01; ***p<0.005).
The following reference sequences were used to design oligonucleotides throughout the study: DENV-2 NGC (AF038463.1); DENV-2 PR1940 (GQ398308.1); DENV-2 PR5344 (GQ398283.1); DENV-2 EDEN 05K3295 (EU081177.1); DENV-3 EDEN 05K802 (EU81184.1); DENV-3 EDEN 05K4454 (EU081222.1); YFV-17D (X03700.1); G3BP1 (NM_005754.2); G3BP2 (NM_203505.2); CAPRIN1 (NM_005898.4); IFITM2 (NM_006435.2); ISG15 (NM_005101.3); MX1 (NM_01144925.2); DDX58/RIG-I (NM_014314.3); EIF2AK2/PKR (NM_002759.3); STAT1 (NM_007315.3); GAPDH (NM_002046.3); ELF2 (NM_201999.2); GRP78/BIP (NM_005347.4).
|
10.1371/journal.pgen.1003671 | Integrated Model of De Novo and Inherited Genetic Variants Yields Greater Power to Identify Risk Genes | De novo mutations affect risk for many diseases and disorders, especially those with early-onset. An example is autism spectrum disorders (ASD). Four recent whole-exome sequencing (WES) studies of ASD families revealed a handful of novel risk genes, based on independent de novo loss-of-function (LoF) mutations falling in the same gene, and found that de novo LoF mutations occurred at a twofold higher rate than expected by chance. However successful these studies were, they used only a small fraction of the data, excluding other types of de novo mutations and inherited rare variants. Moreover, such analyses cannot readily incorporate data from case-control studies. An important research challenge in gene discovery, therefore, is to develop statistical methods that accommodate a broader class of rare variation. We develop methods that can incorporate WES data regarding de novo mutations, inherited variants present, and variants identified within cases and controls. TADA, for Transmission And De novo Association, integrates these data by a gene-based likelihood model involving parameters for allele frequencies and gene-specific penetrances. Inference is based on a Hierarchical Bayes strategy that borrows information across all genes to infer parameters that would be difficult to estimate for individual genes. In addition to theoretical development we validated TADA using realistic simulations mimicking rare, large-effect mutations affecting risk for ASD and show it has dramatically better power than other common methods of analysis. Thus TADA's integration of various kinds of WES data can be a highly effective means of identifying novel risk genes. Indeed, application of TADA to WES data from subjects with ASD and their families, as well as from a study of ASD subjects and controls, revealed several novel and promising ASD candidate genes with strong statistical support.
| The genetic underpinnings of autism spectrum disorder (ASD) have proven difficult to determine, despite a wealth of evidence for genetic causes and ongoing effort to identify genes. Recently investigators sequenced the coding regions of the genomes from ASD children along with their unaffected parents (ASD trios) and identified numerous new candidate genes by pinpointing spontaneously occurring (de novo) mutations in the affected offspring. A gene with a severe (de novo) mutation observed in more than one individual is immediately implicated in ASD; however, the majority of severe mutations are observed only once per gene. These genes create a short list of candidates, and our results suggest about 50% are true risk genes. To strengthen our inferences, we develop a novel statistical method (TADA) that utilizes inherited variation transmitted to affected offspring in conjunction with (de novo) mutations to identify risk genes. Through simulations we show that TADA dramatically increases power. We apply this approach to nearly 1000 ASD trios and 2000 subjects from a case-control study and identify several promising genes. Through simulations and application we show that TADA's integration of sequencing data can be a highly effective means of identifying risk genes.
| The genetic architecture of autism spectrum disorders (ASD) is complex and thought to involve the action of at least hundreds of genes. Yet, despite this complexity, four recent studies [1]–[4] identified five novel genes affecting the risk for ASD from whole-exome sequencing (WES) of 932 ASD probands. The studies made these discoveries by also sequencing the parents of the probands and thereby discovering a multiplicity of independent Loss-of-Function (LoF) mutations in each of these five genes. The multiplicity is key: due to the rarity of de novo LoF events, two or more independent recurrent events in a sample of this size generate more evidence for association than would two LoF variants found in a comparable case and control sample. Thus, even though de novo events are rare, these observations provide an excellent signal-to-noise ratio, have proven valuable in the pursuit of reliable signals for genes affecting the ASD risk, and are likely to form the foundation for many studies targeting gene discovery in the future [5].
Note, however, that the multiplicity test is using only a small fraction of all the information collected by a WES study. Many other de novo events occur, beyond LoF, and these are ignored. Moreover it completely ignores inherited rare variants within families. And, of course, delineation of rare variants into inherited and de novo is challenging or impossible for case-control studies. We conjecture that the distribution of variation, whether inherited, de novo and from case-control, can be leveraged, in combination with the de novo mutations, to maximize the statistical power to detect risk genes.
We propose an integrated model of de novo mutations and transmitted variation to address these challenges. We demonstrate that both the number of de novo mutations and the numbers of different types of transmitted variations in family trios (father, mother and an affected child), follow simple distributions dependent on a set of common parameters: mutation rates, relative risks of mutations and population frequency of the variants. This model readily incorporates additional data from case-control studies. The statistical framework of our model enables us to rigorously analyze the genetic architecture of a complex disease, conduct power and sample size analysis, and identify risk genes with higher sensitivity. Through simulations we show that the power of our novel statistical test, called TADA for “transmission and de novo association”, is substantially higher than competing tests. Our simulations also provide guidance in planning future studies targeting discovery of genes involved in the risks of complex diseases, henceforth, risk genes.
We demonstrate the benefits of TADA through an extensive study of ASD using published WES data from 932 ASD trios as well as nearly 1000 ASD subjects and matched control subjects from the ARRA Autism Sequencing Consortium (AASC) study [6], [7]. Using the model underlying TADA, we estimate there are approximately 1000 genes that play a role in risk for ASD, with an average relative risk of approximately 20 due to LoF in one of these genes. Finally, we identify several potential novel ASD risk genes (genes whose mutations affect the risk of ASD) using TADA and the ASD data.
For concreteness we start by reviewing the multiplicity test to detect risk genes by evaluating the independent recurrence of de novo mutations in the same gene. The multiplicity test classifies a gene as affecting risk if it sustains or more recurrent de novo LoF mutations in a sample of families. Based on computations of expected rates of de novo events as a function of a gene's exonic length and base pair composition [2], a recent study [1] found that LoF events for is significant evidence to declare a gene as a risk gene (, genomewide). Applying this threshold to data from four ASD family studies [1]–[4] led to the discovery of five novel genes affecting ASD risk.
A weakness of the multiplicity test is that it produces a single threshold for the entire genome, regardless of the heterogeneity amongst genes in their sizes and base pair composition, and its threshold is a function of sample size, so that the threshold for is inadequate when the sample increases to . To illustrate the power of the Multiplicity Test and its properties, we performed some simulations using genetic parameters that are described and estimated in the next section.
As demonstrated previously [1], the power for detecting a gene increases monotonically with increasing sample size and it depends strongly on the gene's mutation rate (Figure 1A). Although the per gene power is relatively low, for a disorder like ASD, more than 60 genes are expected to contain at least two LoF mutations with families (Figure 1B). The corresponding false discovery rate (FDR) is less than 5% for and well below 10% for as large as 5,000; switching to a threshold of to diminish false discoveries leads to a significant loss in power (Figure 1B).
The original treatment of the multiplicity test as requiring a single threshold is simple to adjust. Instead one can compute the p-value for each gene using a Poisson model for the probability of observing or more recurrent de novo events based on the gene's mutation rate. We will call such a test the De Novo Test. This test automatically incorporates the number of families and a gene specific mutation rate to determine the likelihood of recurrent de novo events.
TADA model is formulated for sequence data from individual genes. Data for the model can come from sequences of trios (unaffected parents and an affected child) and from cases and controls. Given the information from a gene, namely the pattern of de novo mutations and inherited, damaging variants in the affected progeny, the goal is to relate the data with the underlying genetic parameters such as the relative risk of the mutations. In the model, we restrict the class of variation to rare and deleterious mutations acting dominantly and assume subjects can be classified as carrying one of two “alleles”, those with a deleterious mutation of this type () and those without (). We put alleles in quotes because, for example, we treat all LoF events in the same gene as a single LoF “allele”. Because severe mutations are generally present at very low frequencies in the population (typically ), there are effectively two possible genotypes per gene, and . If we let denote the allele frequency of , then the frequencies of the genotypes and in the population are approximately and , respectively.
For a trio consisting of unaffected parents and an affected child, there are four likely genotype combinations (Figure 2), of which only three are informative: if both parents are homozygous, a heterozygous child results from a de novo mutation; and if one parent is heterozygous, the allele is either transmitted or not. Based on the de novo and transmitted alleles, we formulate a likelihood model for the observed data. Let denote the rate of mutation for the gene being analyzed per generation and chromosome; let denote the genotype relative risk for the genotype ; and let and denote the penetrance of and , respectively. Let , and be the counts of each of the three outcomes (de novo, transmitted and nontransmitted, respectively), from a sample consisting of families. These counts approximately follow Poisson distributions (see Text S1 for derivation): , , and .
For case-control data, counts of genotype in cases and controls follow a Poisson distribution with approximate rate parameters and , respectively (see Text S1). From this structure it is apparent that the transmitted counts can be viewed as a type of case-control data with sample size . Combining data, let be the total number of in the controls plus the number of transmitted variants, and let be the total number of in the cases plus the number of transmitted variants. It follows that(1)for which and . The resulting probability model has three parameters () per gene. For each gene, the mutation rate per gene () can be estimated from its exonic length and nucleotide content [1] and hence this quantity can be treated as known. The statistical problem for each gene is to estimate and then test if .
We conjecture that a more powerful strategy to discover risk genes from family data is to combine the information on de novo and inherited mutations into an unified statistical framework, such as the one we just proposed, which forms the basis for TADA. TADA tests the hypothesis against the alternative . A traditional likelihood ratio test will not work well in this setting because one or more of the counts will be zero for many genes, leading to poor maximum likelihood estimates for and . To circumvent this problem we cast TADA in a Hierarchical Bayes (HB) framework, thereby improving estimates of and by pooling information across all genes, but still modeling rates as gene-specific. The underlying assumption is that LoF and severe missense mutations are rare in all genes and hence we can learn about the frequency distribution in a given gene by looking at the distribution across all genes. Likewise, we can learn about how mutations in one gene affect risk by examining the range and distribution of risks across all disease-related genes.
The HB model assumes a fraction of the genes are associated with the disorder (model ); the remaining fraction follow the null model (model ). Under , the relative risk is constrained (), but under , is assumed to follow a distribution across risk genes. For both models, the frequency of severe mutations per gene, , is assumed to vary by gene, with some commonality across the genome. The distributions of and under both models are specified by prior parameters, and we estimate the values of these parameters by maximizing the marginal likelihood of the data (this is known as the Empirical Bayes method, see Methods). Once the prior parameters are estimated, we compute the evidence for and for each gene. Specifically, for the i-th gene, let be its data, the evidence for is defined as:(2)where is given by Equation 1, and represent the prior distributions. Unlike the likelihood-based test, the evidence for is not based on point estimates of and ; instead it integrates out the two parameters. The model evidence of can be defined similarly, except that is fixed at 1. The Bayes factor of any gene is the ratio of to . The statistical significance of the Bayes factor is given by its p-value, determined empirically by simulating data under the model assuming (see Text S1).
Some insights into the relationship to a likelihood-ratio test (LRT) can be gained by examining an approximation of , the Bayes factor:(3)where the parameters are estimated by Bayesian mean posterior estimators. These parameter estimates are a weighted average of the maximum-likelihood estimate for the i-th gene and the mean of the prior distributions. For example, is interpolated between the allele frequency derived from all genes and the gene-specific estimate (Figure S1). Thus the Bayes factor is similar to the LRT except that we utilize a refined estimator of the allele frequency.
The model just described is designed for a single type of mutation (say LoF), but it can incorporate multiple types. For different types of mutations, such as LoF and damaging missense mutations, the distributions of and are likely to be different, so we model each type of mutation and estimate the prior parameters separately using the HB framework. Then the total Bayes factor of a gene is the product of the Bayes factor from each type of mutation, and the p-value can be computed similarly from simulations. In practice, we note that the damaging missense mutations predicted by bioinformatic tools likely contain a number of mutations having no effect on the gene function, thus we introduce an additional model to account for this feature, downweighting the evidence from missense mutations (see Methods).
The TADA method we described can also be used for de novo data alone. Basically, we ignore inherited and standing variants, but allow multiple types of de novo mutations. The details are not repeated here, but are provided in our supporting Website (see Methods). We call this simplified model, TADA-Denovo, and it is particularly useful for genes with multiple de novo events in different categories (e.g. some nonsense and some missense mutations).
We use the proposed model to estimate the number of ASD risk genes (), their average relative risk (), and the distribution of the population frequency of the mutations. These estimates yield insight into the genetics of ASD and pave the way for realistic simulations to study the power of statistical tests. Our overall strategy is first to use de novo mutations to estimate an approximate range of the parameter values, then use the HB method to refine these estimates using both family and the case-control data.
Consider the de novo LoF mutations in families [1]–[4]. These data reveal a total of de novo LoF mutations across all genes, and multiple-hit genes (at least 2 independent de novo LoF events per gene). Our goal is to find values of and that best predict the observed counts and (Text S1). We assume that the relative risk of an ASD risk gene varies across genes, with the average relative risk of the LoF mutations equal to . The mathematics of TADA reveal there is an inverse relationship between and (Figure 3A, see Equation 27 in Text S1). For an alternative and more intuitive explanation of why these parameters have an inverse relationship, see the arguments in [2]. For any given value of , we can compute the expected number of multiple-hit genes; matching the expected with the observed value of , we estimate the the number of ASD risk genes is between 550 to 1000 (Figure 3B). In the next step, we use the HB model to estimate the most likely value of within this range, and the result is ASD risk genes, with the corresponding relative risk (see Text S1). These estimates are similar to published results using somewhat different methods [1], [2].
We examine evidence for the hypothesis that the population frequency of LoF mutations for ASD risk genes () is lower than that for non-risk genes () because mutations in ASD risk genes are under stronger negative selection than the average gene. These frequencies are of interest because they have a major influence on the power of association test [8]. We estimate based on the number of LoF variants in the case-control data from the AASC [7] and the transmitted/nontransmitted data from 641 families (the transmission data are only available for a subset of the 932 families). To obtain the empirical distribution of across all genes we first count the frequency of the LoF mutations in each gene (Figure 3C); we find a substantial number of genes with 0 LoFs. We next estimate the prior distributions of under the null and alternative models, respectively, using the HB model and find they provide a good fit to the observed data (Figure 3C, Figure S1). From these analyses the mean of under , i.e. the average for ASD risk genes, is about , significantly smaller than that of non-risk genes, (see Text S1 for a description of how the HB model uses a mixture model to permit estimation of parameters specific to ASD risk genes without actually classifying genes as such.) Notably, while the empirical estimate of for most genes is 0 (thus not useful for inference), the value of from the HB model is never equal to 0 due to smoothing.
Using the same procedures we also estimated these parameters for missense mutations that are probably damaging according to the PolyPhen prediction [9] (denoted as Mis3 mutations). Estimates reveal lower risk for these mutations, as expected, and lower for ASD risk genes compared with non-ASD genes (Table S1).
Equipped with estimates of the genetic parameters, we can simulate genetic data under the model and assess the performance of statistical methods. We compare performance of three tests: De Novo, as described in Section 2.1; TADA, described in Section 2.3; and a “Meta test”, which combines two tests, one based on de novo events and the other on inherited variants, via meta analysis. For the meta test we compute the p-value from data on inherited variants using a Fisher exact test, treating transmitted/untransmitted events as case-control data; and compute a p-value for de novo events using the De Novo test. Then these p-values are combined using Fisher's method. In all the simulations, different parameters are used to generate the data, yet TADA always uses the same set of parameters derived from the real data, as described previously. Thus these results establish the robustness of TADA under different parameter settings and thus, to some extent, how it should behave for real data.
Because TADA is a novel method, data were first simulated under the null hypothesis of no association to obtain the distribution of the TADA test statistic and its associated p-values. The results show that the test is well calibrated and type I error is properly controlled (Figure S2).
Next, data were simulated under the alternative model, using different sample sizes and different combinations of the parameters and , within the range of plausible values estimated in the previous section. This comprehensive simulation showed TADA has superior power relative to the other two tests (Figure S3). In Figure 4, we show a selected portion of the simulation results under the most likely scenarios, reflecting the trade-off between relative risks and allele frequencies, i.e. mutations with high risks are likely to exist in lower frequencies in the population. For a gene with typical parameter values (Figure 4B), the power of the TADA test, at , was about fivefold larger than that of the other two tests.
To assess the performance of the tests from a genome-wide analysis, we generated realistic simulated counts based on the estimated genetic parameters for ASD, namely average relative risk of 20 and risk genes, among a total of 18,000 genes sequenced. We focus on false discovery rate (FDR), calibrating the empirical FDR to control at 10%, and estimated power as the number of true discoveries. Results confirmed the advantage of TADA (Figure S4A). For example, at , TADA identified more than 200 ASD risk genes at FDR below 10%, while the De Novo and Meta tests identify about 50 and 70 genes at this level of FDR, respectively (cf Figure 1). We performed additional simulations with somewhat different procedures to demonstrate the robustness of these findings. In one experiment, we simulated data under the average relative risk of 10, instead of 20, while TADA still uses the relative risk of 20. The power of all methods was significantly reduced, as expected, yet TADA still performed better than both de novo test and the simple meta-analysis (Figure S4B). In another experiment, the simulation procedure incorporated the possible dependency between the LoF frequency of a gene () and its relative risk (), based on simple mutation-selection balance: the two were not sampled independently, but rather the frequency was inversely proportional to the risk (see Methods). Despite this change of simulation model, the results were virtually identical to those from earlier simulations (Figure S4C).
The data we used were all reported de novo mutations from 932 ASD families [1]–[4]; transmitted mutations from 641 of these families; and case-control data from the AASC, consisting of 935 ASD subjects and 870 controls [7]. Each missense mutation was classified into a category of damage to the protein based on its predicted effect on the coding sequence using PolyPhen2 [9]: benign (Mis1); possibly damaging (Mis2); and probably damaging (Mis3). Note that de novo LoF mutations occurred at about two-fold enriched rate in the probands relative to the unaffected siblings (Figure 5A, Table S2). The rate for de novo Mis3 was also higher in probands than siblings, but the difference was not as striking. There is essentially no difference in probands and siblings for other types of mutations. We thus applied the TADA method to the LoF and Mis3 mutations.
The overall inflation of the results due to population stratification is negligible: a modified [7] genomic control factor [10] (see Text S1). There is significant enrichment of genes with low p-values compared with random expectation (Figure 5B): 244 genes have , 64 more than expected under the null model. There is an intriguing coincidence in the excess of small p-values - namely that it is very similar to the excess number of genes with single-hit de novo LoF events in ASD subjects compared to their unaffected siblings [1]. Notably the large tail in the QQ plot is largely driven by the de novo LoF events, and appears to reflect true signal instead of inflation.
We control for the multiple hypothesis testing using the Benjamini-Hochberg procedure [11]. Fifteen genes meet the criteria of a False Discovery Rate less than 20% (Table 1, see Table S3 for the complete results). The list includes all five genes with two de novo LoF mutations, as well as several novel genes that are promising candidates for ASD based on existing evidence. For the novel predictions, the p-values from the de novo data alone are far from achieving genome-wide significance (the column in Table 1) and would be impossible to identify without combining the de novo, transmitted and case-control data.
The results of TADA generally depend on the estimates of the mutation rates of the genes, as well as the Bayesian prior parameters of the model. We perform additional analyses to study how sensitive the results are to these parameters. Based on our findings, we choose several genes from Table 1 for this investigation. Although the error of mutation rate estimation is likely small [1], we vary the mutation rate of each gene: from 1/2 of the estimated rate to twice the rate. As expected, the p-value increases as the mutation rate increases, although overall the impact is modest (Figure S5A). Next we vary the Bayesian prior parameter, , which represents the average relative risk over all risk genes, from 10 to 20. The p-values from TADA are even less sensitive to this parameter (Figure S5B).
For disorders like ASD, recent results show that detection of de novo LoF events can be a powerful means of discovering novel risk genes [1]–[4]. Yet de novo events are relatively rare, roughly one per exome, and de novo LoF events even more so, and thus many families must be assessed to identify multiple de novo LoF events in the same gene. To make the most of this experimental design, we develop a new statistical approach, TADA, that utilizes both transmitted and de novo variants from nuclear families and case-control data to determine genetic association. TADA builds on the simple multiplicity test, which relies on recurrent de novo events, but it creates a full analytical framework to incorporate all of the information on the distribution of rare variation. The result is a test with greater power. Our test achieves its good performance properties by providing an analytic framework that links the observed pattern of de novo mutations with the underlying genetic parameters, such as the relative risk conveyed by such mutations. In addition to analyzing data for novel gene discovery, this framework can be used to analyze the power of a test and predict the required sample size to attain sufficient power for future investigations. Moreover, by using empirical Bayes methods, TADA refines estimates of allele frequencies of the damaging mutations by using the full genome to estimate these quantities. This approach increases the information in the transmitted variants in each gene considerably and yet maintains good control of false discoveries.
Association studies evaluating cases and controls have been a common design for identifying variation affecting risk for complex diseases. It has proven successful for identifying common variation affecting risk, after sufficient samples had been amassed to ensure variation having modest impact on risk could be detected [12]. Common variants surely play a role in ASD [13], [14], but the effect sizes are small [15] and it will be challenging to detect individually-significant SNPs. Indeed virtually every discovery for ASD risk genes traces to rare and de novo variants [1]–[4], [16]–[20].
As the cost of sequencing drops, genetic research increasingly focused on the role of rare variants in complex diseases such as ASD, but the sample size has been limited and so has the yield of such studies. For a sample of nearly 1000 ASD case and well matched controls the ARRA ASD sequencing consortium (AASC) found no significant associations [7], except for variation acting recessively [6]. These results comport with studies of other disorders and suggest that large sample sizes will be required to achieve good power in rare variant association studies [21]. Arguably a fundamental difficulty is that most of the mutations with large effects tend to be under strong negative selection, existing at very low frequencies in the population [22]. Variants that occur with greater frequency often have smaller effect on the phenotype, reducing the power of gene-based test statistics.
Our analysis provides insight into some advantages of de novo over case-control studies, especially for LoF events. The de novo test gains power because the mutation rate for genes can be estimated accurately from supplementary sources, and need not be estimated as part of the statistical procedure. Because of the low mutation rate, the number of de novo LoF events expected by chance is very small, and thus we could attach high statistical significance to any gene with more than one independent LoF mutation. While a single de novo LoF event is certainly not definitive evidence, it can put a gene on the short list as a risk gene – for ASD, it is more likely than not an ASD risk gene. In contrast, for case-control data, we require an estimate of the allele frequency under the null hypothesis. When the mutant allele is very rare (as for ASD risk genes), a very large sample is required to ensure that this frequency is indeed small.
Another feature of observed de novo mutations is that they have not been subject to the force of purifying selection, which plays a key role in shaping the pattern of standing variation. Therefore it is likely that de novo mutations, especially LoF mutations, have stochastically larger effect sizes than rare variation transmitted for generations, because selection tends to drive down allele frequencies of variants having large effects on reproductive success. Moreover, allele frequency is inversely tied to power, critical for any experimental design. Therefore studies utilizing de novo variation can have distinct advantages, in terms of power, relative to those that do not.
By simulations we demonstrate that the power of TADA is higher than tests based solely on de novo events or standard meta-analysis that combines p-values from de novo and inherited data (transmission or case/control). There are two explanations for this gain of power. First, TADA's hierarchical model uses the information in the case-control (or transmission) data more efficiently than the standard hypergeometric or trend test. One important property of LoF mutations, compared to less severe functional variants, is their rarity in the population (Figure 3C). TADA, which is similar in spirit to a Poisson test of rare events, is able to exploit the rarity of these damaging events by estimating the distribution of LoF alleles across the exome (see Figure S1B), whereas the other methods cannot. Second, because damaging de novo mutations are rare, most genes will not harbor them even when thousands of cases have been sequenced. For such genes, using Fisher's method to combine the de novo p-value, which will be close to 1, with the p-value from the case-control data penalizes the overall test statistic. In contrast, the Bayesian approach uses de novo events when they are informative and disregards the de novo data when they are uninformative; the Bayes factor from de novo in such cases would be close to 1, making little contribution to the gene's total Bayes factor.
We estimate that there are about 1,000 ASD risk genes with average relative risk about 20. In a recent paper using the same de novo data, the number of ASD risk genes () was estimated at 370 [4]. In that paper, the expected number of genes with recurrent LoF events was derived as a function of , and equating it to 5 (the observed number), produced the solution that . The analysis made the implicit assumption that all ASD risk genes are equally likely to sustain multiple de novo LoF events. In Text S1 we show, using Jensen's Inequality, that the non-uniform distribution of the mutation rates and the relative risks among the ASD risk genes leads to a significant under-estimation of , explaining the discrepancy between our results and those of Iossifov et al. [4].
When applied to ASD data, TADA predicts a number of novel ASD risk genes (Table 1), as well as supporting results for known ASD risk genes. For some of the newly implicated genes it is straightforward to garner other supporting evidence for their role in ASD. S100G is a downstream target of CHD8, a key transcriptional regulator often disrupted in ASD subjects [23]. CUL3 plays a critical role in neurodevelopment [24], [25] and in particular regulates synaptic functions [26]. A recent study identified an additional de novo protein-changing mutation in CUL3 in ASD probands [27], replicating our finding here. COL25A1, a brain-specific collagen, was implicated in risks for Alzheimer's disease [28] and antisocial personality disorder [29].
Inspection of other genes slightly below our chosen FDR threshold reveals several more interesting genes that likely play some role in ASD (all ranked among the top 25, see Table S3). TBR1, a transcription factor critical in brain development, regulates several known ASD risk genes [30]. A recent study has identified recurrent de novo disruptive mutations in TBR1 in ASD subjects [23]. MED13L, a component of the Mediator Complex, is intriguing because of its role in Rb/E2F control of cell growth [31] and the fact that RB/E2F plays a key role in neurogenesis [32] and neuronal migration [33]. Recently MED13L has been associated with risk for schizophrenia [34]. NFIA is a member of the NFI transcription factor family, thought to have a neuroprotective role [35], and NFIA-knockout mice display profound defects in brain development [36].
Genotyping/sequencing errors can introduce biases in data analyses, especially those for family data [37], [38] and for combining data across multiple heterogeneous studies [39]. Our analyses are likely robust to these possible biases because the variant calls were all carefully evaluated: (i) all de novo mutations described previously [1]–[4] and analyzed here, a total of 122 LoF and 314 damaging missense mutations, have been validated by previous studies; (ii) the case-control data have been carefully harmonized to minimize batch effects by using stringent quality control filters [7]; and (iii) for the case-control data, all variant calls in two genes (CHD8 and SCN2A) have been evaluated by Sanger sequencing and 20 out of 20 validate, further supporting the quality of the variant calls in the case-control data. When the sensitivity of calling minor variants is low (under-calling), this may create an under-transmission bias in family-based test statistics; however, TADA is effectively a one-sided test of the adverse effect of the minor allele. As such, TADA is only powered to detect risk variants that are over-transmitted and thus bias due to under-transmission is not a significant concern. Nonetheless, data quality is always an important concern, and can change over time in subtle ways [37], [38], making high-quality filters and validation of de novo events critical for good data analyses. It is possible that TADA would benefit by modeling measurement errors and this will be a topic for future research, when the error structure in the data is better understood.
While much of our focus has been on ASD data and the genetic architecture of ASD, TADA has utility beyond the genetics of ASD. For example, we would expect TADA to be useful for gene discovery by the analysis of data from any genetic disorder or disease for which de novo mutations play a substantive role in risk. Early onset diseases and disorders are obvious candidates for the use of TADA, as are disorders such as schizophrenia and congenital heart disease. Indeed there are a plethora of human diseases for which de novo mutations account for at least a small fraction of risk, even diseases that onset in mid-life such as cardiovascular disease. Because TADA is based on a general theoretical framework for combining rare variation found in exons of genes, we predict that its logic can have even broader applications than simply the analysis of single genes for their association with disease.
We combined exome sequence data from four recent studies of ASD, covering 932 families [1]–[4]. Detailed information about study design, including family structure (simplex versus multiplex), ascertainment, and DNA source (blood versus cell line), can be found in the Supplements of these papers. The de novo mutations, including both single nucleotide variants (SNVs) and indels, were identified as described in the original papers. The transmitted and non-transmitted variants were extracted from 641 of these families (see Text S1 for details on data processing). We excluded all common variants from the analysis, defined as those present at population frequency in the Exome Sequencing Project (ESP) controls and/or the 1,282 parents [40]. Only SNVs were called for the transmission data, indels were not identified. We also included case/control data from the ARRA ASD Sequencing Consortium, consisting of 935 ASD subjects of European ancestry and 870 controls of ancestry similar to cases, selected from the NIMH repository (see complete information on study design in the supplement of Liu et al [7]). The SNVs and indels in the case/control were called as described in [7]. Each mutation/variant in the combined data was classified into different categories, based on its predicted effect on the protein function, according to the program PolyPhen2 [9]. In this study we focused on (1) LoF mutations, defined as nonsense mutations, mutations in splice sites or frameshift indels; and (2) mutations classified as “probably damaging” to protein function by PolyPhen2 (Mis3). We also removed all genes with more than 10 LoF events in the control samples (166 genes in total) from the analysis, as these genes are unlikely to be related to ASD.
For each gene, the total rate of base pair substitutions was estimated using a probability model taking the gene length and base content into account [1]. To estimate the rate of a specific type of mutation (LoF or Mis3) of a gene, we multiplied the gene-level mutation rate and the proportion of that type of mutation. The proportion of LoF or Mis3 mutations was estimated from the data of unaffected siblings in the ASD families (Table S2). In these siblings, there were 461 single-nucleotide variants (SNVs) and 34 LoF variants, thus the LoF fraction was . Similarly the Mis3 fraction was calculated as .
Two hypotheses were compared, versus , for each gene. For most genes, the number of LoF mutations either transmitted or not (or in cases and controls) was generally very small and often 0, leading to a naive estimate of and creating a challenge for a likelihood-based test. To refine inference we took an Empirical Bayes approach and developed a hierarchical Bayes model for the data (Figure 6). We estimated the prior parameters in the model by maximizing the marginal likelihood. The hierarchical model assumed a fraction of the genes was associated with the disorder (model ) and the remaining fraction followed the null model (). Under , we assumed for all genes and followed a Gamma distribution (we parameterized the distribution so that its mean was ). The scaling parameter of the Gamma distribution () played the role of a precision parameter or pseudo count; the bigger the more similar was estimated to be across genes. Under , we assumed of the i-th risk gene follows a distribution and follows a distribution.
Let be the prior parameters of , and be those of (they are also called hyperparameters). The counts for the i-th gene, , follow Poisson distributions parameterized by (1 for non-risk genes) and , as defined in Equation 1.
The marginal likelihood of the i-th gene under either model, and , is given by:(4)(5)The marginal likelihood of all the data, as a function of the hyperparameters , is(6)We assume the proportion of risk genes, , is known (in our analysis of ASD data, this is obtained by the estimated value of , the number of ASD risk genes, see Section 2.4). The hyperparameters can then be found by maximizing this marginal likelihood function. Once we have the estimated values of and , we compute the Bayes factor of any gene:(7)The p-values of the observed Bayes factors are calculated by sampling the null distribution according to Equation 1 (see Text S1).
When analyzing multiple types of mutations (LoF and Mis3 in our analysis of ASD data), we assumed the data for each type of mutation were independent of each other, and hence we estimated the prior parameters for each type of mutation separately. The Bayes factor of a gene is defined as the product of the Bayes factor for each type of mutation. For these ASD data, the Mis3 mutations are likely to be a mixture of those causing protein-damaging changes and those having no real effects on the protein function. We thus computed the joint Bayes factor of the gene using this equation:(8)we used in our ASD analysis (see Text S1).
Our simulation procedure generated data using the estimated genetic parameters of the LoF mutations of the ASD risk genes (Text S1). For our initial simulations, we compared the power of several statistical tests, at the single gene level, under various combinations of parameter values. We set the mutation rate as the mean mutation rate of the LoF mutations of all human genes, . The parameters and were chosen according to their estimated mean values: , and . We compared the power of the three tests under type I error 0.001.
For the second set of simulations, we assessed the performance of the three tests in the genomewide setting. Specifically, from among 18,000 genes in the human genome, we first randomly sampled risk genes and the rest were assumed to be unrelated to disease (we used the estimated mutation rates of all genes to make this simulation realistic). For a risk gene and a LoF mutation, the effect size parameter was sampled from the distribution . Its population frequency parameter was sampled from the distribution, . For a non-causal gene, its relative risk , and the frequency parameter was sampled from the distribution . The simulation procedure then generated, for each gene, the number of de novo mutations (), the number of transmitted variants () and the number of nontransmitted variants (), according to Equation 1.
We ran the three statistical tests, as described in the text, on the simulated data from all genes. At various significance levels, we calculated the number of true discoveries (), i.e. the number of diseases genes whose test statistic reached significance level . We chose the value of so that FDR is less than 0.1, and reported at this value of (see Text S1 for our procedure for controlling FDR in the simulations.)
In additional simulations, we varied the basic procedure just described. In one setting, the average relative risk was set to 10 instead of 20, i.e., of a risk gene was sampled from the distribution . In another setting, instead of sampling and of each risk gene independently, we modeled the two as dependent. Specifically, for the i-th risk gene, let and be the relative risk and the LoF frequency, respectively. First sample from , then determine according to a simple mutation-selection balance: , in which is the mutation rate and is a constant. To determine the value of , we plugged in the mean values of , and in the above equation and solve .
TADA software is available as an R package at http://wpicr.wpic.pitt.edu/WPICCompGen/. The package also includes TADA-Denovo, the simplified version of TADA, that analyzes only de novo data.
|
10.1371/journal.ppat.1004197 | The Calcium-Dependent Protein Kinase 3 of Toxoplasma Influences Basal Calcium Levels and Functions beyond Egress as Revealed by Quantitative Phosphoproteome Analysis | Calcium-dependent protein kinases (CDPKs) are conserved in plants and apicomplexan parasites. In Toxoplasma gondii, TgCDPK3 regulates parasite egress from the host cell in the presence of a calcium-ionophore. The targets and the pathways that the kinase controls, however, are not known. To identify pathways regulated by TgCDPK3, we measured relative phosphorylation site usage in wild type and TgCDPK3 mutant and knock-out parasites by quantitative mass-spectrometry using stable isotope-labeling with amino acids in cell culture (SILAC). This revealed known and novel phosphorylation events on proteins predicted to play a role in host-cell egress, but also a novel function of TgCDPK3 as an upstream regulator of other calcium-dependent signaling pathways, as we also identified proteins that are differentially phosphorylated prior to egress, including proteins important for ion-homeostasis and metabolism. This observation is supported by the observation that basal calcium levels are increased in parasites where TgCDPK3 has been inactivated. Most of the differential phosphorylation observed in CDPK3 mutants is rescued by complementation of the mutants with a wild type copy of TgCDPK3. Lastly, the TgCDPK3 mutants showed hyperphosphorylation of two targets of a related calcium-dependent kinase (TgCDPK1), as well as TgCDPK1 itself, indicating that this latter kinase appears to play a role downstream of TgCDPK3 function. Overexpression of TgCDPK1 partially rescues the egress phenotype of the TgCDPK3 mutants, reinforcing this conclusion. These results show that TgCDPK3 plays a pivotal role in regulating tachyzoite functions including, but not limited to, egress.
| Calcium-dependent protein kinases are plant-like enzymes of apicomplexan parasites that regulate a variety of biological processes including stage-conversion, post-translational repression and egress from the host cell. In this study, we analyzed Toxoplasma CDPK3, which has recently been shown to regulate rapid egress from the host cell. The specific pathways that TgCDPK3 regulates, however, have not previously been known and so we used a quantitative phosphoproteome approach to determine phosphorylation site usage in wild type and TgCDPK3 mutant parasites before, during and after egress. This revealed >150 novel phosphorylation sites that are differentially phosphorylated between WT and TgCDPK3 mutant parasites. Some of these sites are on proteins predicted to play a role in parasite egress. However, we also identified many phosphorylation sites on proteins not thought to be involved in egress, as well as many proteins of unknown function. We confirm that basal calcium levels are affected by CDPK3 inactivation and observed a link between TgCDPK3 and another calcium-dependent kinase (TgCDPK1). Known targets of TgCDPK1 were hyperphosphorylated in the TgCDPK3 mutants, and overexpression of TgCDPK1 partially rescued the observed egress phenotype of TgCDPK3 mutants.
| Apicomplexan parasites like Toxoplasma gondii and Plasmodium species contain a number of plant-like calcium-dependent protein kinases (CDPKs) [1]. These have been shown to be druggable targets that are distinct from their mammalian hosts. For that reason and for their suggested role in regulating calcium-dependent processes in apicomplexan parasites, CDPKs have been the object of intense study. For example, several reports have used either direct, conditional or chemical knock-out strategies to investigate the function of CDPKs in T. gondii and Plasmodium [2]–[11].
In T. gondii, two CDPKs have been investigated in detail. TgCDPK1 (TGGT1_059880) has been shown to be essential for the secretion of micronemes [5] and in three independent studies, TgCDPK3 (TGGT1_041610) has recently been shown to be important for regulating the rapid egress from the host cell upon treatment with the calcium ionophore A23187 [6]–[8], [12]. TgCDPK3 mutants show retarded microneme-secretion and ionophore-induced egress although they still extrude the conoid, a complex structure in the apical part of the parasites, with the usual kinetics [13] [6], [7]. These results show that the mutant parasites can sense the calcium signal induced by the ionophore to some extent but fail to transduce that signal in the normal way. The fact that egress does still occur in the TgCDPK3 mutants, albeit with a delay, indicates that other signaling pathways can operate in the absence of this enzyme [6].
A second phenotype that is associated with TgCDPK3 malfunction is resistance to ionophore-induced death (IID) [13]. IID describes the sensitivity of extracellular tachyzoites to prolonged treatment with a calcium-ionophore. All parasite lines identified harboring mutations in the TgCDPK3 gene show resistance to IID, but the biochemical basis for this is not yet known. TgCDPK3 also appears to be necessary for producing latent stages in mice [14], but the mechanism is also not yet understood.
To understand the role of TgCDPK3 during both normal conditions and ionophore-induced egress, we performed a quantitative phosphoproteome and proteome study of wild type and TgCDPK3 mutants using stable isotope labeling with amino acids in cell culture (SILAC). Our study revealed clues to the role played by this enzyme in normal physiology and induced egress.
To identify the pathways controlled by TgCDPK3, we infected human foreskin fibroblasts (HFFs) with wild type (WT) or TgCDPK3-mutant tachyzoites previously grown in either “heavy” (H) or “light” (L) conditions. “Heavy” indicates the presence of 13C,15N-Lys in place of the naturally-occurring (12C or 14N) amino acids in “light” media [15], [16]). Approximately 24 h after continued growth in either “heavy” or “light” media, the infected cells were incubated for 30 seconds in the presence of 1 µM calcium-ionophore or DMSO as a control (Figure 1A). These samples are called “intracellular”. We also compared WT and mutant parasites grown under “heavy” or “light” conditions and then exposed to ionophore after being released from the host cells by syringe lysis. These samples, called “extracellular,” were used to examine the role of TgCDPK3 in ionophore-induced death and gliding motility, a process by which the parasites use their own motor-proteins to glide over a surface.
In total, we generated 6 datasets (experiments 1–6, see Supplemental Figure S1 and Supplemental Table S1). These compared WT and mutant parasites under each of three conditions in technical duplicate: 1) intracellular parasites without ionophore (“IC/ION−”); 2) intracellular parasites with ionophore (“IC/ION+”); and 3) extracellular parasites with ionophore (“EC/ION+”). We measured each condition with forward and reverse labeling; that is, we labeled WT parasites with “light” amino acids and the TgCDPK3-mutant parasites with “heavy” in one experiment and reversed this labeling for the replicate experiment (Figure 1A and Supplemental Figure S1). This strategy ensured that host-cell peptides, which are abundantly present in the “intracellular” samples, did not introduce quantification errors; reverse labeling effectively removes false positives resulting from misidentified host-derived peptides, since they will show the same heavy/light ratio, irrespective of the labeling. True, parasite-derived quantifications will have reciprocal log2 H/L ratios for the reverse-labeled experiment.
Knock-out mutants of TgCDPK3 (RH:Δcdpk3 [8]), but not the point mutants resulting from chemical mutagenesis (e.g., MBE1.1 [13]), show impaired growth relative to WT parasites. This phenotypic difference led us to use Δcdpk3 and MBE1.1 in different experiments (see Supplemental Figure S1) in our study as phosphorylation sites that differ between WT and both Δcdpk3 and MBE1.1 are unlikely to be caused by either the growth defect observed for Δcdpk3 or a mutation in the chemical mutants outside of the TgCDPK3 gene. To ensure that any change in phosphorylation site abundance is not simply a result of a change in the general abundance of that protein, we also measured non-phosphorylated peptides (for the remainder of the manuscript called the “proteome”) from two experiments.
We analyzed 192 LC-MS/MS runs from 6 different experiments under three different conditions to identify signaling pathways that are differentially regulated between WT and TgCDPK3 mutant parasites: 72 phosphoproteome samples, and 24 proteome samples from two experiments (“IC/ION−” and “EC/ION+”), all analyzed in two independent runs (i.e., in “technical duplicate”; Supplemental Table S1).
We identified differences in protein levels and phosphorylation site usage between WT and TgCDPK3 mutant strains using a phosphopeptide-enrichment strategy [17] which we previously applied to Toxoplasma parasites [18]. All datasets were initially filtered to a false discovery rate (FDR) of <1% on the peptide and <3% on the protein level. In total, we identified 32,147 phosphorylation sites (Figure 1C), 69.4% of which we previously identified in phosphoproteomic studies of WT tachyzoites [18]. Primary analysis of the data obtained revealed up to 50% FDR of “decoy” hits (obtained by searching a fictional decoy database consisting of reversed protein sequences [19]) in phosphorylation sites with a very high or low H/L SILAC ratio even though the total set FDR was well below 1% across all datasets. To largely eliminate false positives from these “tails” (high or low SILAC ratios), we further filtered all quantifications for the signal to noise ratio, spectral counts and MS1-elution parameters of the SILAC pairs (see materials and methods). After such filtering, 19,257 sites remained which were considered quantified with a site FDR of 0.41% and a protein FDR of 1.61% (see Supplemental Table S1). Note that the FDR refers to the estimated number of incorrect identifications of phosphopeptides in the dataset as a whole and does not estimate the correctness of all quantifications, so additional filtering, as described above, is required for highly reliable quantifications. We include all quantifications after the above-mentioned filtering in this manuscript (Table S1) because these represent a resource that can inform the interpretation of complementary approaches to identify targets of TgCDPK3. 78.5% of all quantified sites identified have an ASCORE >19, indicating a 99% probability of being correctly localized to an S, T or Y residue within the peptide they were found [20]. The median-centered SILAC ratios of all quantified sites show a normal Gaussian distribution with a median standard deviation of ∼0.5 for each experiment (Figure 1D), showing that the presence of mostly unlabeled host-cell material in our samples did not skew the expected distribution of SILAC ratios at a detectable level.
We identified a subset of phosphorylation sites as reliably different between WT and mutant parasites by applying two additional criteria: 1) the site was identified in at least two datasets without conflicting ratios for different experiments (e.g., they must have a positive log2 H/L ratio in a forward and a negative log2 H/L ratio in the reverse experiment); 2) the log2 H/L ratio was >0.75 or <−0.75, (∼1.5 times the average standard deviation of any given experiment); sites were designated as “not different” if the log2 ratios were between 0.5 and −0.5. Sites were considered dependent on TgCDPK3 if they differed between WT and mutant parasites under one or more of the three experimental conditions, “IC/ION−”, “IC/ION+” or “EC/ION+”.
Given that treatment with ionophore was for 30 s followed by a wash, minor differences in the timing of the wash could contribute to considerable phosphorylation state variation between datasets. Thus, we allowed one unexpected value in any of the conditions. Using this filtering, we obtained a list of 156 phosphorylation sites that are different between WT parasites and TgCDPK3 mutants with a final FDR of 0.5% (Supplementary Table S2).
The criteria above yielded 156 phosphosites with abundances that differed between wild type and mutant parasites under any of the three conditions; for 130 of these we also obtained protein-level data (i.e., SILAC ratios for nonphosphorylated peptides from the same protein; Figure 2A ad 2C1 and Table S2). Pearson-correlation analysis of the forward and reverse experiment for each condition showed significant correlation (p-values <0.0001 for all comparisons) (Figure 2B). These results allowed us to identify which differences in phosphopeptide abundance can be explained simply by a difference in the abundance of that protein (Figure 2C); only ∼14.9% of the 130 phosphosites where we also had proteome data correlated with protein abundance, showing that the vast majority of sites identified as different in the mutants are likely a result of differences in the degree of phosphorylation (Figure 2C2 and Supplemental Figure S2).
For a majority (∼58%) of the 156 phosphorylation sites discussed above we did not obtain high-confidence SILAC ratios for the untreated samples (“IC/ION−”; Figure 2A and 2C3), precluding any conclusion about their regulation in the absence of ionophore. In the 65 phosphosites where we did obtain such data, however, we observed the following: only 2 of the proteins on which one or more of these phosphosites were detected showed a difference in protein levels and both these were in the set that was different in the “IC/ION−” conditions. 51 of the 65 phosphosites showed a significant difference between wild type and mutant parasites even in the absence of ionophore (Figure 2C3) and only 14 were not significantly different in these latter conditions. These results indicate that TgCDPK3 likely regulates biological processes during the normal function of intracellular parasites, independent of egress and ionophore treatment. Among the phosphorylation sites that are already different in the absence of ionophore are some on proteins important for ion-homeostasis (P-type ATPase, putative, TGGT1_103910) and a dense granule protein GRA22 (TGGT1_125960) that has recently been shown to play a role in egress [21]. Importantly, many of the differences were observed in the two independently derived TgCDPK3 mutant lines (MBE1.1 and RHΔcdpk3) indicating that these changes are a consequence of TgCDPK3 inactivation and not a consequence of the genetic modification of the parasites independent of TgCDPK3 function (Figure 2D).
The two datasets for “EC/ION+” contained most of the 156 differing phosphorylation sites, but these experiments involved comparisons between RH (wild type) and the same mutant strain (MBE1.1). Thus, we could not exclude the possibility that some of the observed differences in phosphorylation state are due to secondary mutations carried by this strain (i.e., are not dependent on TgCDPK3). To address this, we made use of a complemented MBE1.1 cell line which expresses wild type TgCDPK3 under its endogenous promoter [8] and compared the phosphoproteome and proteome of WT and MBE1.1::CDPK3 from extracellular, ionophore-treated parasites in a new experiment “COMP/EC/ION+” using the methods described above. We then retrieved all SILAC ratios of this experiment for the 156 phosphorylation sites that we identified as different between WT and CDPK3 mutant parasites (Supplemental Table S2).
As expected, we observed a significant difference (P-value <0.0011 Kolmogorov-Smirnov test) between the distributions of SILAC ratios of all 156 differing phosphorylation sites of the “EC/ION+” datasets for the mutant (MBE_RH_EC_ION+_forward or reverse) when compared to the SILAC ratios observed for “COMP/EC/ION+” (Figure 3A). Overall, 68 (72.3%) of the 94 phosphorylation sites for which we obtained SILAC ratios from the complementation experiments showed complementation (Figure 3B). Of the 26 sites that were not rescued in the complemented strain, the vast majority (80.8%) appeared different because of differences in protein-levels compared to only 1.5% of the sites that were complemented. These data show that differences in the efficiency of phosphorylation at a given site in the TgCDPK3 mutant are largely complemented and differences that occur at the protein-level are not. While these latter differences in protein abundance could be an indicator of MBE1.1-specific effects due to undetected mutations or passage history, we observed that ∼25% of the sites that were not complemented were also identified as differentially phosphorylated in RHΔcdpk3 vs. WT parasites, indicating a dependence of TgCDPK3 function.
A primary aim of this study was to identify phosphorylation sites that differ in their usage between the WT and TgCDPK3 mutant parasites upon ionophore treatment. Within this set of 156 changing sites, two classes are most interesting: “Class A” where the phosphosite was detected in both WT and mutant parasites in the absence of ionophore-treatment but whose usage did not differ between these two strains; and “Class B” where the phosphosite was not detected in one or other or both of the strains in the absence of ionophore-treatment. This latter class likely represents phosphorylation sites that are substantially phosphorylated only during ionophore-induced egress, assuming the absence of detection in the “IC/ION-“ condition is not caused by technical reasons, as explained below. In Class A, we identified 14 phosphorylation sites on a total of 11 proteins (Figure 4A). Seven of the 11 proteins are more phosphorylated in the mutants relative to WT in the presence of ionophore and 5 of these were predicted or shown to be secreted into the host cell [18]. The preponderance of secreted proteins in this group is discussed further below.
To understand more about TgCDPK3's role, we looked for proteins in our dataset that were already known or predicted to play a role in egress or motility (Figure 4B). Among the identified proteins in Class B are several that are associated with actin regulation (cyclase associated protein (CAP, TGGT1_086070) [22]–[24]), putative motor-proteins (Myosin A (TGGT1_070410), F (TGGT1_103490) and G (TGGT1_092070) [25]), proteins of the inner membrane complex (IMC [26]) and a recently discovered protein that associates with cortical microtubules (TrxL-1 (TGGT1_115220) [27]). Although IMC and microtubule-associated proteins are not directly implicated in egress or motility, rapid rearrangements of the cytoskeleton prior to egress could be part of such a process. Also two uncharacterized kinases (TGGT1_043160, TGME49_053450) were identified, although nothing about their function is known.
Whether TgCDPK3 inactivation leads to a motility phenotype in the presence of ionophore remains unclear. Whereas Lourido and colleagues reported significant differences in the types of motility that TgCDPK3 mutants could perform, McCoy and colleagues observed only a slight, but not significant trend toward such differences [6], [7]. Both groups reported no differences in the speed of the parasites while gliding over a surface. Several members of the motor-complex (MyoA, GAP45 (TGGT1_078320) and MLC1 (TGGT1_013010)) are known to be phosphorylated in a calcium-dependent manner in Toxoplasma and Plasmodium [28]–[33]. Thus, we specifically looked for differences in the relative abundance of phosphopeptides in such proteins as well as other members of the motor complex: GAP40, GAP50, GAP70 and aldolase [34], [35]. Many previously identified phosphosites in these proteins were identified in our datasets but only MyoA and Aldolase (TGGT1_069710) showed a difference in relative phosphorylation between the TgCDPK3 mutants and WT parasites: on MyoA we observed increased phosphorylation of S518 and decreased phosphorylation of S20/S21 in extracellular, ionophore-treated mutants vs. WT (we cannot differentiate which of the two adjacent serines is phosphorylated based on the spectra). The latter site was less phosphorylated in the mutants in the absence of ionophore but this difference decreased in the “EC/ION+” condition. These data suggest that while TgCDPK3 activity is essential for processes that regulate egress-related events, it is not the main kinase regulating phosphorylation of the motor complex components.
Among the proteins that were differentially phosphorylated between wild type and mutant parasites were some that contain EF-hands, proteins that are regulated directly by calcium (Figure 4B). We identified a differentially regulated phosphorylation site on a small EF-Hand protein that contains no other recognizable domain and is annotated as a putative calmodulin (TGGT1_042450). The phosphorylation site was identified just adjacent to the EF-hands themselves.
In addition to the putative calmodulin, we identified two calcium-dependent kinases (TgCDPK2a (TGGT1_062170) and TgCDPK3 itself) as differentially phosphorylated in the MBE1.1 mutant vs. WT. For both kinases, the sites were hyperphosphorylated in the mutant parasites and were located within the ATP-binding loop of the kinase domain, a region where phosphorylation can play a regulatory role [36], [37]. It is worth noting that the mutation (T239I) in the activation loop of TgCDPK3 in the mutant MBE1.1 [8] functionally inactivates the kinase and so a different kinase must be involved in the hyperphosphorylation.
Interestingly, our dataset also indicated that two of several phosphorylation sites on proteins recently identified as targets of TgCDPK1 (TGGT1_059880). PRP (TGME49_005320) and DRPB (TGGT1_064650) [38] are, surprisingly, more phosphorylated in TgCDPK3 mutant vs. WT parasites upon ionophore treatment. Although TgCDPK1 itself did not emerge from our stringently filtered datasets as differentially phosphorylated in the mutants vs. WT, these results indicate that there might be an increase in activity of TgCDPK1 in the TgCDPK3 mutants. Hence, we specifically looked for evidence of differences in the phosphorylation state of TgCDPK1 in our datasets and found several phosphopeptides with high-quality quantifications in the “EC/ION+_FW” condition that corresponds to the activation loop threonine (T200) of TgCDPK1 [39], [40]. The phosphorylation site identified in TgCDPK1 showed a higher level of phosphorylation (2.9-fold) in the mutant vs. WT parasites (Figure 4c), while the protein levels of TgCDPK1 appear similar between WT and TgCDPK3 mutants in our proteomic dataset (Supplemental Table S1) and by Western blot (data not shown). This phosphorylation is a prerequisite for activity of the kinase and supports the model that the TgCDPK3 mutants have a higher level of activated TgCDPK1. It was identified in no other dataset which is why it is not included in the set of 156 phosphorylation sites discussed above (all of which were seen in at least two datasets), but all peptides containing quantitative information for TgCDPK1:T200 were identified in the heavy and the light version in two different fractions giving confidence in its identification and quantification. All phosphopeptides were identified as a missed cleavage form, which is often the case for phosphopeptides [41]. We did not identify the activation-loop phosphosite in the dataset from the complemented mutants and so could not assess whether such complementation rescued its phosphorylation; however, we did observe hyperphosphorylation of the above-mentioned targets of TgCDPK1 in the MBE1.1 mutant and all showed phenotypic rescue upon TgCDPK3 complementation. Hence, TgCDPK3 plays a role in the phosphorylation of these proteins through another kinase, presumably, TgCDPK1.
TgCDPK3 mutants egress in the presence of ionophore with much slower kinetics than WT parasites [8], [13]: whereas after 2 minutes 100% of WT parasites have exited in the presence of ionophore, TgCDPK3 mutants only start to slowly egress after 3–4 minutes, reaching near 100% egress levels by 10 minutes [13]. The increased amount of phosphorylated TgCDPK1 and its targets in the TgCDPK3 mutants prompted us to test whether accumulation of active TgCDPK1 might eventually reach levels necessary for this enzyme to take the place of TgCDPK3, thereby explaining the ability of the TgCDPK3 mutant parasites to respond, albeit slowly, to the ionophore. To test this hypothesis, we over-expressed TgCDPK1 in MBE1.1 parasites using a strong promoter (from the GRA2 gene) and measured the ability of MBE1.1::TgCDPK1 parasites to egress in the presence of ionophore. We tested the level of overexpression using an antibody specific to TgCDPK1, which recognizes a single band at around 55kD in MBE1.1 parasites (Figure 5A). The product of the introduced TgCDPK1 transgene is HA-tagged and showed the expected size-shift on gel electrophoresis with an expression level that was ∼10× higher than the slower migrating, endogenous TgCDPK1 (Figure 5A). Immunofluorescence imaging of TgCDPK1::HA showed a concentration of TgCDPK1::HA toward the periphery of MBE1.1::TgCDPK1 parasites whereas WT parasites showed a more general cytosolic staining; biochemical fractionation assays, however, revealed no difference in membrane association of TgCDPK1 in the two strains (data not shown). While we saw no difference in growth or number of parasites/vacuole in the MBE1.1::TgCDPK1 parasites (data not shown), they were substantially but not fully rescued in their ionophore-induced egress phenotype (Figure 5B): at 2 minutes, most WT (RH) parasites had egressed but MBE1.1 and MBE1.1::TgCDPK1 remained mainly inside; however, at 6 minutes ∼50% of MBE1.1::TgCDPK1 had egressed while MBE1.1 were still almost entirely intracellular. This shows that over-expression of TgCDPK1 can partially overcome the block seen in mutants lacking active TgCDPK3, supporting the implication of the SILAC data that the function of these two kinases is directly or indirectly linked.
We have previously shown that the egress phenotype of TgCDPK3 mutants can be rescued by overexpressing an engineered form of TgCDPK3 that has mutations in its predicted myristoylation and pamitoylation sites but only when using a strong promoter (SAG1), not when using the endogenous promoter [8]. This was likely because overexpressing TgCDPK3 allows a fraction of it to reach the plasma membrane [8] where it can phosphorylate its targets. To test whether the increased levels of activated TgCDPK1 observed in MBE1.1 mutants might similarly be rescuing the egress delay observed by phosphorylating TgCDPK3 targets, we directly compared the TgCDPK1 and TgCDPK3 substrate specificity. We incubated peptide microarrays spotted with ∼500 defined but random 13-mer peptides containing a central serine residue with recombinant TgCDPK1 or recombinant TgCDPK3 and compared their ability to phosphorylate the arrayed peptides (Figure 5C). We ranked the peptides according to the phosphorylation intensity with the highest phosphorylated peptide being 1st. A comparison of the ranking for each given peptide incubated with either CDPK1 or CDPK3 shows that a majority of the peptides are equally well phosphorylated by the two kinases although some are predominantly phosphorylated by one, but not the other. We did not obtain significantly enriched amino acids in any position for either TgCDPK1 or TgCDPK3, but that might be related to the technical limitation of these arrays. They sometimes contain more than 1 phosphorylatable residue per peptide; i.e., an additional serine, threonine or tyrosine residue in addition to the central serine. Since we cannot distinguish whether the central serine, or another phosphorylatable residue is phosphorylated, they have limited value for a motif analysis. However, it still allowed us to directly compare TgCDPK1 and TgCDPK3 in their linear motif analysis with the result that TgCDPK1 appears likely able to phosphorylate at least a subset of TgCDPK3 targets.
In addition to the calcium-dependent proteins mentioned earlier, we identified a putative calcium-transporting ATPase (TGGT1_103910) that showed differential phosphorylation between mutant and WT parasites in the “IC/ION−” condition. These results suggested a putative role for TgCDPK3 in regulating calcium levels. To test this, we measured calcium levels in the absence of ionophore in WT, MBE1.1, MBE1.1 complemented with a functional copy of TgCDPK3 (MBE1.1::CDPK3) and, as a control, MBE1.1 complemented with a nonfunctional copy of TgCDPK3 (MBE1.1::CDPK3(T239I)). Both, MBE1.1 and MBE1.1::CDPK3(T239I) showed elevated calcium levels (∼150%) compared to MBE1.1 complemented with a functional copy of CDPK3 (100%), supporting the notion that TgCDPK3 plays a role in maintaining normal intracellular calcium levels (Figure 6). To rule out that integration of the TgCDPK3 WT copy into MBE1.1 lowered basal calcium levels because of an off- target effect, we also measured basal calcium levels in MBE1.1 parasites complemented with the Plasmodium falciparum orthologue of TgCDPK3, PfCDPK1 (PF3D7_0217500), that has recently been shown to complement the egress phenotype of MBE1.1 [42]. Complementation with the WT PfCDPK1 version, but not a kinase dead mutant (T231I) decreases calcium levels similar to those of MBE1.1::CDPK3. No significant differences in the change of calcium levels were observed in the response to the calcium ionophore itself (data not shown). This suggests that inactivation of CDPK3 causes a reproducible elevation of basal calcium levels that are directly dependent on TgCDPK3.
The aim of this study was to identify phosphorylation events that are dependent on TgCDPK3. Based on previous publications, we hypothesized that the signaling pathways controlled by TgCDPK3 might be most easily detected during ionophore-induced egress. Among the proteins that fulfilled this prediction were several that are known to be secreted into the parasitophorous vacuole (PV), the PV-membrane (PVM) or into the host cell. Some of the phosphopeptides in these proteins showed a decreased abundance in the WT samples which could be due to dephosphorylation or degradation of these proteins resulting from breakdown of the PVM. We have not further investigated these events here, but they are consistent with the fact that breakdown of the PVM, normally an early event in egress, is defective in the TgCDPK3 mutant parasites [6]–[8], [13].
Despite the low number of phosphorylation sites for which we obtained SILAC ratios in untreated samples as discussed above, we identified several significant differential phosphorylation events on proteins that are either known or predicted to play a role in egress or associated processes, and these are discussed further below. In addition to sites that are differentially phosphorylated, those that show no change in phosphorylation status in the mutants relative to WT can be equally informative. For example, our results indicate that TgCDPK3 is not the major regulator of key phosphorylation sites observed on the components of the machinery that drives parasite motility including GAP45, one of the key “glideosome” proteins known to be regulated by phosphorylation [28], [30]. The only known part of the motor for which we confidently saw differences in the phosphorylation state between WT and TgCDPK3 mutants was MyoA. The differences observed for MyoA could explain some of the observed phenotypic differences in TgCDPK3 mutants with regards to motility, but given that at least one site (S20/S21) was already differentially phosphorylated in intracellular parasites not treated with ionophore (“IC/IONO−”), it appears that this phosphorylation site has a function independent of, or in addition to, egress. A recent report on a conditional knock-out of MyoA, where loss of this protein has no effect on parasite egress (or invasion [43]) supports the notion that phosphorylation of MyoA and other proteins by TgCDPK3 can serve functions other than egress. Interestingly, two other myosin isoforms (MyoG and MyoF) show up as differentially phosphorylated in the TgCDPK3 mutants vs. WT, one of which, MyoF, was recently described as playing a role in apicoplast segregation [44]. The role that these myosin isoforms play during ionophore-induced egress requires further analysis.
One protein strongly indicated as a potential regulator of motility is CAP, a regulator of actin dynamics [45], which we found to be ∼2-fold less phosphorylated in the TgCDPK3 mutants relative to WT in the presence of ionophore. In tachyzoites, CAP has been shown to be localized in the apical end, rapidly redistributing into the cytosol when becoming extracellular [46]. Thus, CAP could regulate actin dynamics during egress and motility in a location-dependent manner. In Plasmodium berghei, deletion of CAP showed a defect in oocyst development but the function of CAP and actin regulation in this process is not understood [23].
The identification of TgCDPK3-dependent phosphorylation of calcium-regulated proteins, including TgCDPK1, TgCDPK2a, and an EF-hand containing protein is a strong indicator that TgCDPK3, in addition to being regulated by calcium, controls other calcium-regulated processes. Furthermore, the identification of two proteins that are known phosphorylation targets of TgCDPK1, and indications that loss of TgCDPK3 may result in a more active TgCDPK1 itself, indicate that the pathways controlled by TgCDPK1 are, at least in part, activated in TgCDPK3 mutants treated with ionophore. While this suggests that TgCDPK3 might be a negative regulator of TgCDPK1, we have no direct evidence for this and understanding their precise roles will require further investigation.
Overexpression of TgCDPK1 partially rescues the egress phenotype. This indicates that 1) either active TgCDPK1 can phosphorylate targets of TgCDPK3, or 2) that active TgCDPK1 can activate egress and microneme secretion independent of TgCDPK3, but with much slower kinetics. Both kinases appear to have overlapping substrate specificity and the dominant localization of the overexpressed TgCDPK1 at the periphery and the increased rescue of egress supports the first option where TgCDPK1 may be phosphorylating TgCDPK3 targets at the plasma membrane. However, we cannot exclude the alternative, in which egress in the TgCDPK3 mutants is facilitated via a TgCDPK3-independent pathway. In both scenarios, the elevated levels of TgCDPK1 we observed in MBE1.1 parasites that are phosphorylated in the autophosphorylation loop might explain how TgCDPK3 mutant parasites egress slowly over time when treated with ionophore.
A possible explanation for how CDPK1 might be activated is via a PKG (TGGT1_087710) controlled pathway. Lourido et al., have shown that activation of PKG is partially CDPK3 dependent as inhibition of CDPK3 decreases egress triggered by Zaprinast [6]. PKG has also been shown to be important for egress in Plasmodium falciparum as a key-regulator for calcium levels thought to be important for regulation of the TgCDPK1 orthologue in P. berghei, PbCDPK3 [47], [48]. While we consistently identified the phosphorylated activation loop of TgPKG (Threonine 837, Supplementary table S1) under all conditions, we did not observe CDPK3- dependent differences. But it is possible that other phosphorylation sites of the protein, which we might not have detected for technical reasons, or other regulatory mechanisms are mainly involved in regulation of PKG.
Whatever the mechanism by which TgCDPK3 and PKG are connected, our data support a broader role of TgCDPK3 and the pathways it controls. This is evident from the proteins that are differentially phosphorylated, including proteins that are important for metabolism, transcription and ion homeostasis in addition to the proteins important for egress, several of which are differentially phosphorylated even in the absence of ionophore. This is in line with our observation that basal calcium levels are elevated in TgCDPK3 mutants in the absence of ionophore. This allows for a model in which altered calcium fluxes in TgCDPK3 mutants have a profound effect on the homeostasis of the cell, which could dictate how a cell behaves under conditions of stress. The fact that CDPK3 mutant parasites don't show a measurable phenotype in cell culture points towards compensatory mechanisms that allow for normal growth. The fine balance the parasites have to strike might be easily tipped as in the case of calcium ionophore treatment, when the egress phenotype becomes evident. While this phenotype was only observed in vitro, CDPK3 mutant parasites also have a phenotype in vivo.
Type I Toxoplasma parasite strains lacking TgCDPK3 (as used in this study) are still highly virulent in mice whereas Type II strains lacking TgCDPK3 activity are attenuated with severely reduced latent stages (tissue cysts) being found in the brain of chronically infected animals [8], [49]. We previously hypothesized that this could be due to an egress phenotype but the data we present here would suggest that the phenotypic changes are due to alterations in TgCDPK3-dependent functions unrelated to egress.
Our observation that there are differences in protein abundance in the TgCDPK3 mutants relative to WT resembles a recent study of the closest orthologue of TgCDPK3 in the malaria parasite Plasmodium berghei, PbCDPK1, that showed a role for PbCDPK1 in translational repression [10]. However, our results differ in two ways: 1) we observe differences in the absence of environmental triggers and 2) complementation of the TgCDPK3 mutants with a WT copy of the gene restores protein levels in only one case. The fact that we identified a substantial fraction of non-complemented phosphorylation sites in both the TgCDPK3 point mutant (MBE1.1) and knock-out clearly indicates that the effect is due to TgCDPK3. Several of these sites can be explained by changes in protein level indicating that phosphorylation state and proteome changes occur in the absence of TgCDPK3 and their non-complementation suggests that an epigenetic event “locks in” some of the effects. While it is possible that genetic effects in the chemical mutants are partially responsible for the observed static changes of protein levels, an epigenetic effect seems more likely as the only protein coding mutation identified in MBE1.1 using whole genome sequencing was in TgCDPK3. In addition to that, several of the changes that do not revert are identified in the chemical mutant and the TgCDPK3 knock-out strain, strongly arguing for a TgCDPK3 specific effect. Among the proteins and phosphorylation sites that are not complemented are some that are important regulators in glycolysis and other metabolic pathways, suggesting loss of TgCDPK3 could lead to long-lasting phenotypic changes that will not be reversed upon complementation. The results presented here, therefore, present a new insight into how the protein kinases of Toxoplasma interact to regulate several key functions, extending well beyond ionophore-induced egress.
Parasites lines and labeling of parasites was achieved as previously described [15], [16]. Briefly: parasites were grown in heavy (146 mg/l 13C6-, 15N2-L-lysine, 84 mg/l 13C6-L-arginine, 40 mg/L unlabeled L-Proline) or light media (146 mg/l unlabeled L-lysine, 84 mg/l unlabeled L-arginine, 40 mg/l unlabeled L-Proline). After 4 complete lytic cycles, parasites incorporated between 96% and 98% of heavy amino acids. For the comparative analysis parasites were seeded onto confluent human foreskin fibroblast in 150 mm dishes with an MOI of 5 in either heavy or light media. 24 hours post-infection the cells were washed once with fresh media and incubated in the presence of 1 uM A23187 or DMSO for 30 seconds. After the incubation time, the parasites were immediately placed on wet ice and quickly washed once with pre-chilled, ice-cold PBS prior to lysis in ice-cold 8 M urea containing protein and phosphatase inhibitors (Roche). We performed each experiment using 5 individual 15 cm dishes in order to monitor an average of the signaling events that take place during the first 30 seconds of ionophore treatment.
Peptide and phosphopeptide samples were prepared as previously described [17], [18] using SCX and IMAC chromatography for phosphopeptide enrichment. Briefly, samples were lysed, reduced, alkylated, and digested with trypsin. After desalting, the peptides were fractionated using strong cation exchange chromatography (SCX) and phosphopeptides were further enriched using IMAC (immobilized metal affinity chromatography) and the phosphorylated and non-phosphorylated flow-through peptides were analyzed by LC_MS/MS on a LTQ-Velos Orbitrap in technical duplicates in a total of 196 MS/MS runs. In addition, we analyzed the phosphoproteome and proteome of RH vs. the MBE1.1 mutant complemented with a WT copy of TgCDPK3 (MBE1.1::CDPK3; [8]) in technical duplicate in comprised of a total of 36 runs. This was done using parasites “EC/ION+”.
Phosphorylated and non-phosphorylated (flow-through) peptides were resuspended in 4% formic acid, 5% acetonitrile and analyzed by LC-MS/MS in technical duplicate on a system consisting of a MicroAS autosampler (Thermo Scientific), binary HPLC pumps (Agilent 1200 series) with flow-splitting, an in-house built nanospray source, and an LTQ Orbitrap Velos (Thermo Scientific). 2 µg of sample was loaded onto a 100 µm ID fused silica capillary packed with 18 cm of 5 µm Magic C18AQ resin (Michrome Bioresources). Peptides were eluted using a gradient of water:acetonitrile with 0.1% formic acid from 7% to 25% acetonitrile over 120 min, and then 25–40% B over 30 minutes. A top 10 method was run consisting of one MS1 scan (resolution: 6×104 AGC: 5×105, maximum ion time: 500 ms) followed by ten data dependent MS2 scans (AGC: 1×10, maximum ion time: 100 ms) of the most abundant ions. Dependent scans were configured with the following settings: 2.0 m/z isolation width, dynamic exclusion width: −0.52, 2.02, exclusion duration: 60 seconds, normalized collision energy: 35, activation time: 5 ms. Charge state screening was employed to reject ions with unassigned or +1 charge states.
HFFs were cultivated in 96 well plates and each well infected with 500 parasites. After 24 hours of growth, the parasites were incubated in HBSS containing 1 µM A23187 calcium ionophore for time periods ranging from 0 to 10 minutes, in triplicates, after which the cells were fixed in 100% methanol and subsequently stained with Giemsa. All assays have been done on at least three independent occasions.
Spectra were searched against a concatenated database of Human (IPI, version 3.66) and Toxoplasma (toxoDB, release 6.1) proteins using SEQUEST [50], with 15 ppm precursor mass tolerance, trypsin specificity with up to two missed cleavages, static modification of cysteine (carbamidomethylation, +57.0215) and variable modification of serine, threonine, and tyrosine (phosphorylation, +79.9663), methionine (oxidation, +15.9949) lysine (SILAC 13C(6)15N(2), +8.0142), and arginine (SILAC 13C(6), +6.0201). Phosphorylation site localization was assessed using the Ascore algorithm [20]. All datasets were filtered using the target-decoy method [19], [51] to a false discovery rate (FDR) of <1% on the peptide and <3% on the protein level. Phosphopeptides were combined into phosphosites based on their localization probabilities, and phosphosites were further filtered to an FDR of <1% for each phosphorylatable residue (S,T,Y) using the peptide score provided by Ascore. All peptides matching to the human proteome were removed to exclude peptides from the parasite that are identical with the host and where quantification is thus not reliable. SILAC quantification was achieved by analysis of the MS1-intensity peaks using the VISTA algorithm [52]. Quantifications were scored using: closeness of log2 H/L to 1∶1, signal to noise of each isotopic partner, and a VISTA confidence metric that accounts for chromatography quality. Weighted averages were calculated using these scores for sites and proteins for which more than one identification was made. These weighted averages are intentionally conservative, in an attempt to eliminate false-positives from the tails of the distribution. Unweighted average and standard deviation calculations have been included as well. Phosphopeptides were categorized as either mono-, bis-, or tris-phosphorylated, and separate averages were calculated for sites found in peptides of each type. The resulting data was further filtered for a minimum of 2 quantifications in each respective experiment, a minimum VISTA confidence score of 88 for the best quantification and a minimum signal to noise ratio for the best peptide of 8.
SILAC log2 ratios were centered on “0” based on the median SILAC ratio for each dataset.
Differentially phosphorylated sites were identified by using the following criteria:
a minimum log2 fold change of 0.75 (+ or −), which is ∼1.5 times the standard deviation across the experimental datasets, and a consistent change of phosphorylation site abundance in one or more of the conditions. Note, one mismatch was allowed to capture phosphorylation sites that are just at the threshold, or missing in a single sample due to a bad SILAC quantification, to be included in the dataset. Phosphorylation site quantifications were manually curated by analyzing the MS1 elution profiles and removed if they originated from low-quality quantifications.
Protein SILAC ratios were calculated by using the median SILAC ratio for all identified peptides. Pearson correlation was calculated using a two-tailed test with a 95% confidence interval (Prism6).
The open reading frame of TgCDPK1 was cloned into pGRA [53] using the restriction sites NsiI and NcoI. The plasmid was linearized using HindIII and transfected into MBE1.1 parasites and selected with MPA/XAN as previously described [53].
Random peptide library kinase arrays (KIN-MA-RLYS, JPT, Germany) with a central serine were incubated according to the manufacturer's instructions. Briefly, 100 nM recombinant CDPK1 or CDPK3 was incubated with 50 µM Ca and 10 µM cold ATP to allow for autophosphorylation and activation of the kinase. This mix (400 µL) was then supplemented with 10 µCi of 32P-ATP and added to the peptide arrays for 4 h at 30 degrees Celsius. After washing and drying a high-resolution phosphor-imaging screen (Fujifilm, BAS-SR) was exposed. Images were acquired using a Typhoon scanner using 25 µM resolution and spot intensity values were analyzed using microarray software. Median spot intensities were generated after manual verification of each spot. Contaminations were manually removed and excluded from the analysis.
Each median intensity was plotted as a rank from highest (1st rank) to the lowest observed phosphorylation.
Naturally egressed parasites were resuspended in intracellular buffer (5 mM NaCl, 142 mM KCl, 2 mM EGTA, 1 mM MgCl2, 5.6 mM glucose, 25 mM HEPES, pH 7.2) and labeled with Fluo-4 AM (2.5 µM) at room temperature with shaking for 30 minutes. Parasites were then pelleted, washed and resuspended in intracellular buffer followed by incubation at room temperature with shaking for 30 minutes to allow de-esterification. The parasites were then loaded onto 96-well plates (107 per well) and fluorescence (excitation 496 nm and emission 516 nm) was quantified using a Synergy H1 plate reader (BioTek).
Recombinant full length TgCDPK1-HIS6 was generated by cloning the open reading frame into the expression vector pET28. Recombinant TgCDPK1 was purified using the HIS-tag with NI-NTA agarose (GE-Healthcare). 200 µg of purified TgCDPK1 was injected with an equal volume of Freunds incomplete adjuvant into 6–8 week old pre-screened BALB/c mice from Charles Rivers. None of the mice showed reactivity in the pre-bleeds. Mice were boosted 3 times with 100 µg of recombinant protein every 3 weeks. Test bleeds were taken at week 6 and all mice were sacrificed at week 10 for the terminal bleed. Final bleed serum was screened for selectivity by Western blot and used in this study.
Statistical analysis and graphs were made using GraphPad Prism6. Data handling was performed using Excel and MySQL using Python scripts. Heat-maps were generated using Cluster 3 (unclustered, k-means), and visualized by Treeview.
|
10.1371/journal.pgen.1007151 | Nuclear re-localization of Dicer in primary mouse embryonic fibroblast nuclei following DNA damage | Dicer is a key component of RNA interference (RNAi) and well-known for its role in biogenesis of micro (mi)RNA in the cytoplasm. Increasing evidence suggests that mammalian Dicer is also present and active in the nucleus. We have previously shown that phosphorylated human Dicer associates with chromatin in response to DNA damage and processes double-stranded (ds)RNA in the nucleus. However, a recent study by Much et al. investigated endogenously tagged HA-Dicer both in primary mouse embryonic fibroblasts (PMEFs) as well as adult homozygous viable and fertile HA-Dicer mice under physiological conditions and concluded that murine Dicer is exclusively cytoplasmic. The authors challenged several findings, reporting functions of Dicer in mammalian nuclei. We have re-investigated this issue by applying subcellular fractionation, super-resolution microscopy followed by 3D reconstitution, and phospho-Dicer-specific antibodies using the same HA-Dicer PMEF cell line. Our data show that a small fraction of the murine HA-Dicer pool, approximately 5%, localises in the nucleus and is phosphorylated upon DNA damage. We propose that Dicer localisation is dynamic and not exclusively cytoplasmic, particularly in cells exposed to DNA damage.
| Cytoplasmic Dicer is a key component of the canonical micro (mi)RNA biogenesis pathway. However, a growing body of evidence points toward localisation and activity of mammalian Dicer in the nucleus. A recent study by Much et al., employed an endogenously HA-tagged Dicer knock-in mouse cell line to show that Dicer is exclusively cytoplasmic. This paper challenges several studies reporting various RNA metabolic functions of Dicer in human nuclei. Given the controversy about Dicer’s subcellular localisation, it is essential to address this issue. Employing the same cells as used by Much and colleagues, we combined super-resolution microscopy followed by 3D reconstitution and biochemical assays to show that endogenously tagged HA-Dicer prominently localises in the nucleus under physiological conditions. We demonstrate that DNA damage triggers accumulation of phosphorylated HA-Dicer in the nucleus, confirming previous observations in human cells. Our data indicate evolutionary conservation of nuclear Dicer localisation and function in mammals in response to DNA damage.
| The endoribonuclease Dicer recognises and processes double-stranded (ds)RNA substrates of various origins into small non-coding (nc)RNA [1]. Dicer activity generates 20–25 nt long micro (mi)RNA from precursors and modulates gene expression by post-transcriptional gene silencing (PTGS) in the cytoplasm (reviewed in [2]). A growing body of evidence suggests that additional functions for Dicer may exist in many species, including mammals, which are potentially independent of miRNA biogenesis and may involve non-canonical modes of RNAi in the nucleus (reviewed in[3]). Conditional depletion of Dicer in mouse embryonic stem cells, for instance, compromises centromere silencing and impairs expression of homologous endogenous dsRNA loci [4, 5]. A series of studies imply nuclear localisation of mammalian Dicer and association with chromatin. The Filipowicz lab reported enrichment of mammalian Dicer at ribosomal RNA loci, suggesting a possible role for Dicer in maintaining integrity of ribosomal DNA arrays. However, the authors could not describe a direct function for Dicer in nucleoli [6]. Human Dicer may also interact with the nuclear pore complex component NUP153 [7]. Interestingly, Dicer depletion in human cells caused defects in precursor messenger (pre-m)RNA processing [8]. Catalytically active Dicer was purified from human nuclei and shown to promote processing of dsRNA hairpin structures [9] and stimulate initiation of RNAPII transcription at hormone-responsive genes [10]. In addition, nuclear Dicer fosters termination of RNAPII transcription [11] and alternative polyadenylation at a subset of protein-coding genes [12]. The latter two studies conclude that Dicer association with chromatin may be mediated by the localised production of dsRNA, which is processed into endogenous small interfering (endo-si)RNA to mediate heterochromatin formation by recruitment of G9a methyltransferase in a Dicer-dependent manner. These findings are in line with previous studies, reporting the existence of nuclear RNAi in human cells [13]. The authors showed that transfection of exogenous small interfering (exo-si)RNA triggers silencing of a subset of protein-coding gene promoters. More recently, two studies point toward Dicer-dependent nuclear RNAi in mammals by demonstrating that nuclear, chromatin-associated Dicer impairs expression of the microtubule-binding protein Doublecortin in mouse adult neural stem cells [14] and transactivation of the human secreted frizzled-related protein 1 promoter in cholangiocarcinoma cells [15]. Collectively, these data indicate that Dicer may be present and active in mammalian nuclei to regulate expression of protein-coding genes by both miRNA-dependent and -independent mechanisms.
However, mechanistic insight in mammalian nuclear Dicer localisation remains largely inconclusive. Analysis of ectopically expressed human Dicer mutants suggest that the dsRNA binding domain (dsRBD) may harbour a cryptic nuclear localisation signal, which is potentially occluded by the helicase domain in the full-length Dicer protein [16]. Indeed, lack of the helicase domain or duplication of the dsRBD trigger nuclear accumulation of ectopically expressed Dicer mutants [16]. Confusingly, the authors could not detect nuclear localisation of full length Dicer under physiological conditions. The N-terminal Dicer helicase domain forms a clamp-like structure adjacent to the RNase III active site in the base of the Dicer enzyme [17]. Truncation of the helicase domain or alterations of C-terminal domains, such as introduction of post-translational modifications, may cause structural rearrangements that ‘unfold’ the helicase domain, potentially exposing an ‘unmasked’ C-terminal domain for increased dsRNA binding affinity and catalytic activity. However, recent data demonstrate that a different cytoplasmic N-terminal deletion mutant of human Dicer efficiently processes exogenous dsRNA substrates in HEK293-derived Dicer knockout cells, but fails to accumulate to the nucleus [18], indicating that rearrangement of the Dicer helicase domain is necessary, but not sufficient for nuclear accumulation.
Moreover, recent work by Much et al. challenged the existence of mammalian Dicer in the nucleus per se [19]. Using PMEF::HA-Dicer cells, a primary mouse embryonic fibroblast cell line, which expresses a catalytically active, endogenously HA-tagged Dicer (HA-Dicer) at physiological levels [20], the authors failed to detect any evidence for nuclear HA-Dicer localisation under conditions that were previously reported to trigger nuclear Dicer accumulation, such as treatment with the nuclear export inhibitor Leptomycin B (LMB), stimulation of mitogen-activated protein kinase (MAPK) signalling or DNA damage-inducing γ-irradiation. These findings seem to contradict various other subcellular localisation studies, which apparently detect a fraction of Dicer in the nucleus of human cells [21, 22] and in purified nuclei [23]. In this regard, we have recently shown that a subset of the endogenous human Dicer pool is phosphorylated in response to DNA damage and associates with DNA double-strand breaks (DSBs) on chromatin to process damage-induced dsRNA [24]. Similarly, human Dicer may be recruited to DNA lesions to mediate chromatin decondensation during nucleotide excision repair in response to UV irradiation [25]. These findings suggest a functional link between nuclear Dicer accumulation and the DNA damage response (DDR).
Here, we provide evidence for existence of HA-Dicer in murine nuclei under physiological conditions and involvement of nuclear phosphorylated HA-Dicer in the DDR. Using subcellular fractionation, super-resolution microscopy followed by 3D reconstitution and phospho-Dicer-specific antibodies, we demonstrate that a small fraction of HA-Dicer localises to nuclei of unperturbed cells. Following DNA damage, phosphorylated HA-Dicer accumulates in the nucleus in a phosphatidylinositol-3-kinase (PI3K)-dependent manner. We propose that a subset of the mammalian Dicer pool relocalises to the nucleus rather than being exclusively restricted to the cytoplasm.
Comprehensive assessment of subcellular localisation of endogenously tagged HA-Dicer in PMEF::HA-Dicer cells (Fig 1A) critically relies on avoidance of non-physiological artefacts, usage of adequate culture conditions of primary cells and both reliable and sensitive detection methodology. We noticed that low-passaged PMEF::HA-Dicer cells were not dividing as rapidly as wild type PMEF cells. To monitor for potentially elevated levels of senescent cells in our PMEF::HA-Dicer culture, we assessed expression of several proliferation markers following either starvation or serum stimulation (S1 Fig). Expression of cyclin E, cyclin B1, c-Myc and phosphorylation of MAPK effectors ERK1/2 as well as p38 was markedly decreased in PMEF::HA-Dicer cells starved with media containing 0.1% fetal bovine serum (FBS) and restored upon stimulation of starved cells with 20% FBS. In contrast, we could not detected elevated levels of cell cycle inhibitors p21 or p16, irrespective of changes in culture conditions. We conclude that predominantly non-senescent PMEF::HA-Dicer cells were cultured.
Previous investigations of HA-Dicer in PMEF::HA-Dicer cells excluded any nuclear localisation or activity of HA-Dicer [19]. We noticed that some imaging data presented in this study displayed a spotted and sporadic distribution of HA antibody signals, not only in the cytoplasm of PMEF::HA-Dicer cells, but also in wild type PMEF nuclei, contrary to a rather homogenous cytoplasmic HA staining of testis, thymus and uterus samples (Figs 1–3 in[19]). Moreover, close inspection of mass spectrometry data provided by Much and colleagues (S1 Table, https://doi.org/10.1371/journal.pgen.1006095.s003) indicates that several factors involved in nuclear RNA metabolism, such as RNA polymerase II co-activators p15 and TIF1B or the pre-mRNA processing factor Fip1 may potentially be overrepresented in HA immunoprecipitations from PMEF::HA-Dicer cells compared to controls from wild type PMEFs, arguably reflecting false-positive enrichment due to non-specific HA antibody reactivity. To optimise conditions for HA-Dicer analysis, control for false-positive data and allow flexibility in antibody combination, we initially tested three different HA epitope tag antibodies, namely rabbit monoclonal C29F4 and mouse monoclonal HA.11, which were both used by Much and colleagues, as well as rat monoclonal 3F10 by immunoblotting. We incubated each antibody with whole cell extracts from either wild type PMEFs, PMEF::HA- Dicer cells or Dicer-/- knockout MEFs (S2A Fig). Each HA antibody generated a prominent band migrating at 250 kD in presence of PMEF::HA-Dicer, but not wild type or Dicer-/- MEF extracts. However, C29F4 generated an additional band, migrating at 130 kD when incubated with either of the extracts. Next, we incubated each HA antibody with serial dilutions of identical PMEF::HA-Dicer extract (S2B Fig). We detected a prominent signal migrating at 250 kD, which was sensitive to dilution, with all three HA antibodies. C29F4 reactivity was lost after diluting the extract 4-fold, whereas HA.11 and 3F10 signals remained detectable at 4-fold dilution. Unlike 3F10, both C29F4 and HA.11 generated additional signals migrating at 130 kD and 80 kD, respectively. To quantify the sensitivity of each HA antibody and visualise HA-Dicer detection thresholds for each HA antibody, we calculated loss of HA reactivity as ratio of relative HA signal normalised to Ponceau S signal from whole cell extract, which we defined as deltaHA (ΔHA), and plotted values over serial dilution steps. We found that at dilution steps 1:2 and 1:4, which allow quantification of signals in the linear range, ΔHA values are highest for C29F4 and lowest for 3F10, indicating increased sensitivity of HA.11 and 3F10 antibodies and compromised specificity of C29F4 and HA.11. We further incubated each HA antibody with whole cell extracts from PMEF::HA-Dicer following starvation or serum stimulation (S2C Fig). Again, C29F4 and HA.11, but not 3F10, generated aberrant bands of high molecular weight, which were sensitive to starvation and induced by serum stimulation. Of note, it is currently unclear, whether aberrant bands detected with HA antibodies C29F4 and HA.11 reflect predominantly unspecific signals or may also display Dicer cleavage products with potential relevance for Dicer localisation and function.
Next, we tested for specificity and sensitivity of HA antibodies in immunofluorescence microscopy. We co-cultured wild type PMEFs and PMEF::HA-Dicer cells in absence or presence of nuclear export inhibitor LMB or DNA damage-inducing Topoisomerase II inhibitor Etoposide prior to HA staining with either C29F4, HA.11 or 3F10 antibody (S3A Fig.). Each HA antibody generated prominent cytoplasmic reactivity in a subset of untreated control cells, arguably reflecting HA-Dicer expressed in PMEF::HA-Dicer cells. Cytoplasmic HA reactivity was accompanied by increased nuclear HA staining upon treatment with LMB or Etoposide. Importantly, treatment with LMB or Etoposide neither caused significant onset of nuclear HA reactivity in cells without cytoplasmic HA staining, indicating specificity of each individual HA antibody in confocal imaging (S3A Fig). Drug treatments also did not alter expression of full length HA-Dicer (S3B Fig). Again, we detected several aberrant signals when probing with C29F4 and HA.11, but not 3F10 HA antibody. We conclude that the 3F10 HA antibody is most specific and sensitive for detection of HA-Dicer by immunoblotting, whereas differences among HA antibodies seem marginal when used in confocal imaging.
To reassess the subcellular localisation of HA-Dicer in unperturbed PMEF::HA-Dicer cells, we combined the 3F10 HA antibody with highly sensitive super-resolution microscopy. We detected clear nuclear 3F10 staining, in addition to prominent reactivity in the cytoplasm (Fig 1B). Incubation with LMB enhanced nuclear HA-Dicer 3F10 signals 2-3-fold (Fig 1C). To substantiate nuclear HA-Dicer localisation, we applied subcellular fractionation and probed for HA-Dicer with each individual HA antibody (Fig 1D). Although the bulk of HA-Dicer was present in the cytoplasmic fraction, we detected a clear HA-Dicer signal in the nuclear fraction using C29F4 or HA.11 antibodies. The 3F10 antibody generated the strongest nuclear HA-Dicer signal. Importantly, cytoplasmic and endoplasmatic reticulum (ER) membrane-associated contaminants were not detectable in nuclear fractions.
To estimate the amount of nuclear HA-Dicer in the nucleus of unperturbed cells quantitatively, we measured nuclear HA-Dicer band intensities relative to cytoplasmic levels and calculated the amount of nuclear Dicer in % normalised to a cytoplasmic-to-nucleoplasmic input ratio of 1:3, reflecting a 3-fold concentrated nuclear fraction. Values for nuclear HA-Dicer varied between 3.7% and 9.7%, depending on the HA antibody. We conclude that a subset of approximately 5% of the total HA-Dicer pool localises to the nucleus of PMEF::HA-Dicer cells under physiological conditions.
We have recently discovered that a subset of human Dicer is phosphorylated in response to DNA damage to process damage-induced dsRNA in the nucleus [24] and wished to assess murine HA-Dicer subcellular localisation in context of DNA damage. First, we monitored that PMEF::HA-Dicer cells are responsive to DNA damage. We detected elevated levels Ataxia telangiectasia mutated (ATM)/ATM-related (ATR) kinase substrates, increased phosphorylation of histone variant H2A.X (γH2A.X) and induction of downstream effector p21, indicating onset of DNA damage signalling upon treatment with γ-irradiation and Etoposide, but not with LMB (S4A Fig). Next, we made use of two recently described phospho-specific Dicer antibodies, which were raised against two conserved carboxy-terminal murine Dicer phospho-serine residues Ser1712 and Ser1836 [26]. We used a mixture of both phospho-antibodies, hereinafter p-DCR-1 antibodies, for comprehensive detection of phosphorylated HA-Dicer. Of note, we have previously confirmed the specificity of p-DCR-1 antibodies by mutation of both phospho-serine epitopes into alanine residues in human HEK293 cells [24]. When incubating p-DCR-1 antibodies with PMEF::HA-Dicer whole cell extracts, we detected increased p-DCR-1 reactivity following preincubation with Etoposide, but not LMB (Fig 2A). To further validate specificity of p-DCR-1 antibodies, we used the 3F10 HA antibody to immunoprecipitate comparable amounts of total HA-Dicer from PMEF::HA-Dicer, but not wild type PMEF cells cultured in presence or absence of Etoposide (S4B Fig.). Elevated levels of γH2A.X were indicative for Etoposide-induced DNA damage. When probing immuno-selected samples with p-DCR-1 antibodies, we detected faint, but modestly elevated reactivity upon Etoposide treatment (S4C Fig). Moreover, p-DCR-1 reactivity was detectable when probing whole cell extracts from wild type PMEF and PMEF::HA-Dicer cells, but not Dicer-/- MEFs upon incubation with Etoposide (S4D Fig.). We conclude that p-DCR-1 antibodies specifically detect damage-induced Dicer phosphorylation on immunoblots.
Next, we preformed subcellular fractionation of PMEF::HA-Dicer cells cultured in presence or absence of Etoposide (Fig 2B). Using HA antibodies C29F4 or HA.11, we found the bulk of the total HA-Dicer pool localising in the cytoplasm, with a small fraction present in nuclei, irrespective of DNA damage. In contrast, signals for both total and phosphorylated HA-Dicer were elevated 2-3-fold in damaged nuclei upon detection with 3F10 and p-DCR-1 antibodies. We could, however, not detect a shift in migration of HA-Dicer, indicating that the DDR targets a relatively small number of Dicer molecules. We noticed that Much and colleagues failed to detect nuclear HA-Dicer in subcellular fractions and speculated that this may be due to different amounts of material loaded as nuclear input, which does not include 3-fold concentrated nuclear samples. Indeed, when loading fractions in a 1:1 ratio, we could not detect clear HA-Dicer signals in nuclear fractions on blots displaying prominent cytoplasmic HA-Dicer levels, irrespective of Etoposide-induced DNA damage (S4E Fig).
To visualise phosphorylated HA-Dicer, we performed confocal imaging of PMEF::HA-Dicer cells stained with p-DCR-1 antibodies in absence or presence of Etoposide (Fig 2C). Strikingly, we detected formation of nuclear, p-DCR-1-positive foci in >90% of cells incubated with Etoposide, but not in undamaged control nuclei and confirmed partial colocalisation of p-DCR-1 signals with γH2A.X-positive damage foci using RGB profiler. To confirm that Etoposide-induced nuclear HA-Dicer localisation is primarily caused by induction of DSBs, we tested for formation of p53 binding protein 1 (53BP1)-positive damage foci, a hallmark of DSB repair [27]. Indeed, we detected strong nuclear 53BP1 staining in damaged cells and partial colocalisation of 53BP1 signal with nuclear HA-Dicer upon Etoposide incubation (S4F and S4G Fig). Next, we wished to control for specific detection of total and phosphorylated HA-Dicer in immunofluorescence microscopy. Therefore, we co-cultured a mixture of both wild type PMEFs, PMEF::HA-Dicer cells and Dicer-/- MEFs in absence or presence of Etoposide and co-stained cells with HA (3F10) and p- DCR-1 antibodies (S4H Fig). In absence of Etoposide, strong reactivity of HA, but not p- DCR-1 antibodies was detected predominantly in the cytoplasm in a subset of cells. Treatment with Etoposide, however, induced additional nuclear HA staining and onset of nuclear p-DCR-1 reactivity in cells both positive and negative for HA staining. Importantly, a subset of cells remained negative for both HA and p-DCR-1 reactivity, indicating that p-DCR-1 antibodies detected phosphorylated HA-Dicer in damaged wild type PMEF and PMEF::HA-Dicer, but not Dicer-/- MEF cells.
Three members of the phosphatidylinositol-3-kinase (PI3K) family—ATM, ATR and DNA-dependent protein kinase (DNA-PK)—govern the response to DNA damage by phosphorylating hundreds of substrates [28–31]. To investigate the contribution of PI3Ks to damage-induced phosphorylation of murine HA-Dicer, we treated PMEF::HA-Dicer cells with Etoposide in absence or presence of PI3K inhibitors and imaged cells using p-DCR-1 and HA antibodies (S5A Fig). Again, we detected prominent nuclear localisation of p-DCR-1 foci and HA-Dicer in the vast majority of damaged nuclei. In contrast, preincubation of PMEF::HA-Dicer cells with PI3K inhibitors prior to Etoposide treatment attenuated both p-DCR-1 and HA staining in the nucleus in 80–90% of imaged cells. When monitoring for HA-Dicer expression levels in absence or presence of Etoposide or upon preincubation with PI3K inhibitors we could not detect significant alterations in HA-Dicer levels (S5B Fig). To control for activity of PI3K inhibitors, we probed for phosphorylation of downstream targets checkpoint kinase 1 (Chk1) and γH2A.X. We confirmed that inhibition of ATR or ATM impairs Etoposide-induced phosphorylation of Chk1 or H2A.X, respectively. We conclude that phosphorylated HA-Dicer accumulates in the nucleus in response to DNA damage in a PI3K-dependent manner.
To further assess the subcellular localisation of phosphorylated HA-Dicer in PMEF::HA-Dicer cells in response to DNA damage, we preformed γ-irradiation. We have recently described damage-induced nuclear localisation of phosphorylated human Dicer 2–3 hours after γ-irradiation with a total dose of 10 Gray (Gy) [24]. To assess HA-Dicer localisation in response to γ-irradiation under optimised conditions, we employed a time course experiment using a total dose of 10 Gy γ-irradiation followed by up to 5 hours recovery (Fig 2D). Formation of 53BP1-positive foci was used as a marker for DSBs. We detected a wave of p-DCR-1 reactivity, which was detectable concomitantly with induction and clearance of 53BP1-positive damage foci. We found that 53BP1 foci formation was induced 30 minutes after irradiation modestly, but peaked 2 hours after irradiation and was reduced after 5 hours. The p-DCR-1 signal was also modestly detectable after 30 minutes, but most clearly detectable 2 hours after irradiation. We further confirmed formation of DSBs by detection of prominent γH2A.X -positive foci, which partially co-localised with p-DCR-1 staining (S6 Fig). Moreover, prominent nuclear localisation of total HA-Dicer was also detectable 2 hours after irradiation using 3F10 HA antibody, suggesting recruitment of phosphorylated HA-Dicer to close proximity of DSBs.
To further validate specificity of 3F10 HA and p-DCR-1 antibodies in confocal imaging, we repeated γ-irradiation in both wild type PMEFs and Dicer-/- MEFs at optimised conditions. As expected, γ-irradiation neither induced 3F10 reactivity in wild type PMEFs (S7A Fig), nor p-DCR-1 reactivity in Dicer-/- MEFs (S7B Fig), nor reactivity of secondary Alexa Fluor antibodies (S7C Fig). In analogy to Etoposide treatment, we repeated co-culture of wild type PMEF wild type PMEFs, PMEF::HA-Dicer cells and Dicer-/- MEFs and stained for HA-Dicer using HA and p-DCR-1 in presence of γ-irradiation (S7D Fig.). Again, we detected increased nuclear HA reactivity in a subset of irradiated cells and onset of nuclear p-DCR-1 reactivity in cells both positive and negative for HA staining as well as cells, which were negative for pDCR-1 and HA signals, despite being irradiated. Much and colleagues used γ-irradiation to study HA-Dicer subcellular localisation in response to DNA damage at a single time point, namely 30 minutes after irradiation with a total dose of 20 Gy, but failed to detect nuclear HA-Dicer localisation. We speculated that suboptimal conditions were used to induce phosphorylation of HA-Dicer by DNA damage signalling in PMEF::HA-Dicer cells. Surprisingly, we detected prominent reactivity for both total and phosphorylated HA-Dicer in damaged nuclei as little as 30 minutes after high dose irradiation (S7D Fig.). Importantly, we further confirmed induction of HA-Dicer phosphorylation by γ-irradiation and specificity of p-DCR-1 antibodies by (i) immunoblotting with p-DCR-1 antibodies with irradiated wild type PMEF, PMEF::HA-Dicer and Dicer-/- MEF whole cell extracts (S8A Fig.), (ii) partial colocalisation of p-DCR-1 and HA signals upon γ-irradiation using RGB profiling (S8B Fig.) and (iii) by super-resolution imaging followed by 3D reconstitution (S8C Fig.). Modest colocalisation of p-DCR-1 and HA signals could also be observed in cells analysed immediately after irradiation (30 minutes), but not after prolonged incubation (5 hours) or non-irradiated control cells, suggesting that a threshold of DNA damage limits detection of nuclear phosphorylated HA-Dicer following irradiation with 10 Gy. Reassuringly, γ-irradiation induced phosphorylation of DNA damage signalling components Chk1 and H2A.X, but did not significantly alter HA-Dicer levels over time (Fig 2E). Of note, detection of nuclear phosphorylated HA-Dicer appeared to be clearer in immunofluorescence microscopy than on immunoblots. The reason for this apparent discrepancy is currently unclear, but might at least in part be due to intrinsic differences in p-DCR-1 antibody sensitivities, as p-DCR-1 antibodies are primarily suited for immunofluorescence microscopy and comprise limited performance in immunoblotting (Swathi Arur, personal communication).
For proof of principle, we performed an interspecies heterokaryon experiment (Fig 3). Sporadic formation of interspecies heterokaryons containing both murine and human nuclei was confirmed by DAPI staining, displaying typical spotted murine nuclei, and staining with Phalloidin, displaying a cellular continuum. Strikingly, we detected strong, nuclear 3F10 HA antibody reactivity in HEK293 cells fused to PMEF::HA-Dicer, but not wild type PMEF cells in Etoposide-treated cells. Induction of nuclear 3F10 reactivity occurred concomitant with formation of γH2A.X-positive DNA damage foci and partial colocalisation with γH2A.X. As only PMEF::HA-Dicer cells express HA-tagged Dicer, we conclude that the HA signal detectable in HEK293 nuclei originated in PMEF::HA-Dicer cells. Minor nuclear HA reactivity was also detectable in HEK293 cells fused to PMEF::HA-Dicer in absence of Etoposide, arguably reflecting non-specific accumulation of HA-Dicer in nucleoli after permeabilisation and displacement of ribosomal RNA. Importantly, however, no HA signal was detectable in HEK293 cells fused to wild type PMEFs, irrespective of Etoposide treatment, underscoring specificity of the 3F10 HA antibody toward HA-Dicer and damage-induced nuclear accumulation of HA-Dicer.
Taken together, our data indicate that a subset of HA-Dicer localises to nuclei of unperturbed cells, with increased levels of phosphorylated HA-Dicer being detectable in damaged nuclei. We show that phosphorylation of HA-Dicer in response to DNA damage triggers accumulation of HA-Dicer in the nucleus. Phosphorylated nuclear HA-Dicer arguably reflects a minor fraction of the HA-Dicer pool, which we estimate to be 5%. Importantly, no adverse impact on viability and fertility of mice homozygous for the DcrFH allele has been reported, suggesting that HA-Dicer is not compromised in its localisation or function by introduction of the HA epitope tag per se [20]. Advancing on previous Dicer localisation studies, our findings provide insight into nuclear Dicer localisation and function under physiological conditions. Since we monitor for specificity and sensitivity of HA-Dicer detection using various controls, we conclude that Much and colleagues may have failed to detect changes in the subcellular localisation of murine HA-Dicer upon DNA damage or other stimuli due to technical limitations, such as lack of antibody sensitivity.
Recent evidence from C. elegans Dicer, DCR-1, demonstrates that phosphorylated DCR-1 is accumulating and functional in nuclei. In the adult worm germ line, DCR-1 localises in uniformly distributed cytoplasmic and nuclear foci and on the inner side of nuclear pores [32]. During development, the Ras-dependent MAP kinase MPK-1, a homologue of ERK kinases in mammals, phosphorylates cytoplasmic DCR-1 at two serine residues in the C-terminal RNase III and dsRBD domains, which triggers nuclear translocation of phosphorylated DCR-1 [26]. A similar translocation phenotype has been observed in human HEK293 cells. Importantly, phosphorylation of conserved C-terminal residues by MAPK signalling is conserved in mammalian Dicer, as demonstrated by in vitro kinase assays and fibroblast growth factor (FGF) stimulation [26]. Along the same lines, we showed accumulation of human phosphorylated Dicer in damaged nuclei, and discovered the damage-inducible Dicer residue serine-1016, which facilitates accumulation of Dicer in the nucleus and processing of nuclear, damage-induced dsRNA [24].
We conclude that Dicer proteins are found in the nuclei of the vast majority of studied eukaryotes, including mammals. The cytoplasm remains the main compartment for Dicer localisation. However, during development or stress, a subset of the cytoplasmic Dicer pool may be altered either genetically, by proteolysis, heat shock, or by PTMs to adjust for Dicer subcellular localisation or activity, suggesting structural and functional distinct nuclear Dicer subpopulations. Our findings point toward additional layers of complexity in the regulation of RNAi components and underscore the relevance of studying mechanisms of non-canonical RNAi in mammals.
Wild type or HA epitope-tagged primary mouse embryonic fibroblasts (PMEF wt and PMEF::HA-Dicer, female, a kind gift from the O’Carroll Lab), or Dicer knockout mouse embryonic fibroblasts (MEF Dicer-/-, clone [1A11], [4] a kind gift from the Heissmeyer Lab), or wild type human HEK293 cells were cultured in Dulbecco's modified Eagle complete medium (DMEM, Sigma) supplemented with 10% fetal bovine serum (FBS, Life Technologies), 2 mM L-glutamine, 1x non-essential amino acids and 100 units/ml penicillin/streptomycin at 37°C and 5% CO2 at low passages (<20 passages). Non-immortalized PMEFs were derived from E13.5 embryos of DcrFH/+ intercrosses according to standard protocols. For serum starvation, PMEF::HA-Dicer cells were shifted to DMEM containing 0.1% FBS for 24 hours. For serum stimulation, PMEF::HA-Dicer cells were starved in DMEM containing 0.1% FBS for 20 hours, followed by incubation with DMEM containing 20% FBS for 4 hours prior to lysis. Nuclear export was inhibited with CRM1/exportin1-inhibitor Leptomycin B (Cayman, 20 nM, 16 hours). DNA damage was induced with the Topoisomerase II inhibitor Etoposide (Sigma, 25 μM, 2 hours) or γ-irradiation using at total dose of 10 or 20 Gy. Small-molecule inhibitors KU-55933 (ATM inhibitor, 5 μM, Sigma); VE-821 (ATR inhibitor, 1 μM, Sigma); LY294002 (PI3K inhibitor, 5 μM, NEB) were used for 1 hour prior to induction of DNA damage.
Subcellular fractionation was performed as described [10]. Approximately 3x 106 PMEF::HA-Dicer cells grown on 10 cm dishes (Corning), were trypsinised, washed in cold 1x PBS and centrifuged (1200rpm, 5 min). Pellets were lysed in five volumes (i.e. 300 μl) of hypotonic lysis buffer (10 mM HEPES pH 7.9, 60 mM KCl, 1.5 mM MgCl2, 1 mM EDTA, 1 mM DTT, 0.075% NP-40, 1x protease/phosphatase inhibitor cocktails, Roche) and incubated for 10 minutes at 4°C with rotation. Nuclei were pelleted by centrifugation (1200 rpm, 4°C) for 10 minutes. The cytoplasm was collected from the supernatant. Nuclei were washed five times in 800 μl hypotonic lysis buffer without NP-40 and lysed in 1 volume (i.e. 33 μl) of nuclear lysis buffer (20 mM HEPES pH 7.9, 400 mM NaCl, 1.5 mM MgCl2, 0.2 mM EDTA, 1 mM DTT, 5% Glycerol, 1x protease/phosphatase inhibitor cocktails, Roche). Lysates were diluted with two volumes (i.e. 66 μl) of dilution buffer (20 mM HEPES pH 7.9, 1.6% Triton- X-100, 0.2% Sodium deoxycholate, 1x protease/phosphatase inhibitor cocktails, Roche), followed by 10 sec sonication with a Bioruptor (Diagenode) at low energy and incubation with 10 U Benzonase (Sigma) for 5 min. Lysates were centrifuged (13500 rpm, 4°C, 10 minutes) and the supernatant (i.e. 100 μl) was collected as soluble nuclear fraction. 10% of subcellular fractions (i.e. 30 μl) of cytoplasmic and 10 μl of nuclear fraction were boiled in 0.25x volume (i.e. 30 μl for cytoplasm or 3.33 μl for nuclei) of 4 x SDS-PAGE sample buffer (12% SDS, 40 mM Tris HCl pH 7.4, 40% glycerol, 3% beta-Mercaptoethanol, 1% Bromophenol Blue) at 95°C for 5 minutes, respectively. Samples were sonicated and 10 μl of either cytoplasmic or nuclear fractions (i.e. 25% or 75% of boiled samples, respectively), were analysed by Western Blot using precast gels (Mini-PROTEAN TGX, BioRad). Each gel lane was loaded with 1/40 (10% x 25%, i.e. 2.5% of lysed cytoplasm) or 3/40 (10% x 75%, i.e. 7.5% of lysed nuclei), respectively, in a 1:3 ratio, unless stated differently.
The amount of nuclear Dicer relative to cytoplasmic levels and normalised to the 1:3 input ratio was calculated in % using the following equation: [(HA Ab signal in nuclear fraction / HA Ab signal in cytoplasmic fraction) / (7.5% of lysed nuclei / 2.5% of lysed cytoplasm)] x 100%. Intensities of bands were quantified using ImageJ and values for nuclear Dicer were plotted as relative signals normalised to signals from cytoplasmic fractions or non-damaged nuclear fractions, respectively. For example, calculation of amount of nuclear Dicer using 3F10 Ab as shown in Fig 1D: [(0.29/1) / (7.5 / 2.5)] x 100% = 9.7%. Values for nuclear Dicer using C29F4 or HA.11 were 8.0% or 3.7%, respectively.
Whole cell extracts from approximately 5x 105 cells grown on 6-well multi-well dishes (Corning) were lysed directly in 100 μl 4 x SDS-PAGE sample buffer, 10 μl of lysate was loaded, separated and stained with Ponceau S (Sigma) prior to antibody hybridisation. For semi-quantitative analysis of detection thresholds of HA antibodies, 10 μl lysate of PMEF::HA-Dicer cells, which were grown on 6-well multi-well dishes and lysed in 100 μl 4 x SDS-PAGE sample buffer, was diluted 5 times in a 2-fold serial dilution series. HA signal intensities of 10 μl of either non-diluted sample, i.e. input, or diluted samples, or Ponceau S stainings thereof, were quantified using ImageJ and plotted as relative normalised signals. Signals were plotted, loss of reactivity of HA antibodies was visualised as gap and quantified as delta (Δ = Rel. norm. Ponc. S signal—Rel. norm. HA Ab signal). For each HA antibody, a Δ was calculated at dilutions steps 1:2 or 1:4 to quantify signals in the near-linear range of sensitivity.
For immunoprecipitation, approximately 3x 106 wild type PMEF or PMEF::HA-Dicer cells grown on 10 cm dishes (Corning) were trypsinised, washed in cold 1x PBS and centrifuged (1200rpm, 5 min). Pellets were lysed in 5 volumes WCE lysis buffer (20 mM Tris pH 7.5, 150 mM NaCl, 0.1% NP-40, 2 mM MgCl2, 50 mM NaF, 1 x protease/phosphatase inhibitor cocktails, Roche) for 20 minutes on ice, sonicated and Benzonase digested as described above. WCE lysates were precleared with protein G agarose beads (Merck Millipore) for 30 min. Samples were incubated with 5 ug primary HA antibodies for 4 hours and pulled down using protein G agarose beads for 45 min. IP samples were washed three times for 10 min with 800 μl WCE lysis buffer (20 mM Tris pH 7.5, 150 mM NaCl, 0.1% NP-40, 2 mM MgCl2, 50 mM NaF, 1 x protease/phosphatase inhibitor cocktails, Roche), and eluted with SDS-PAGE sample buffer.
The following primary antibodies were used: anti-α-Tubulin (Abcam, [YL1/2], ab6160); anti-Rad21 (Merck Millipore, 05–908); anti-RNAPII-CTD S2P (Abcam, ab5095); anti-histone H3 (Abcam, ab1791); anti-HA (Roche, [3F10], 11867423001); anti-HA (BioLegend, [16B12], 901501, previously Covance MMS-101P); anti-HA (CST, [C29F4], 3724); anti-Lamin C (Novus, [EM11], NBP1-50051); anti-KDEL motif (Abcam, [10C3], ab12223); anti-γH2A.X (Ser139, Merck Millipore, 05–636); anti-Drosha (Abcam, ab12286); anti-Cyclin E (Santa Cruz, [M-20], sc-481); anti-Cyclin B1 (Abcam, ab2949); anti-c-Myc (Clontech, 631206); anti-53BP1 (Santa Cruz, [H-300], sc-22760); anti-pATM/ATR substrates mix (CST, [SxQ, D23H2/D69H5], 9670); anti-pChk1 (Ser345, CST, 133D3); anti-pChk1 (Ser317, CST, D12H3); anti-pERK1/2 (Thr202/Tyr204, CST, 9101); anti-ERK1/2 (CST, 9102); anti-p-p38 (Thr180/Tyr182, CST, 9211); anti-p21 (Santa Cruz, [C-19], sc-397); anti-p16 (Santa Cruz, [F-12], sc-1661); and anti-p-DCR-1 (Ser1712/Ser1836, a kind gift from the Arur Lab) [26].
The p-DCR-1 signals represent a mixture of two individual antibodies, raised against carboxy-terminal murine Dicer epitopes phospho-Ser1712 and phospho-Ser1836 individually in separate rabbits. Murine epitopes Ser1712/Ser1836 are equivalent to human epitopes Ser1728/Ser1852 and C. elegans epitopes Ser1705/Ser1833. However, human and mouse epitopes differ by one amino acid relative to the original epitope in C. elegans.
Approximately 3x 105 wild type PMEFs, Dicer-/- MEFs clone [1A11], and PMEF::HA-Dicer cells grown on 6-well multi-well dishes (Corning) were washed in 1x PBS, fixed on coverslips with 3% Paraformaldehyde in PBS for 10 min, washed and incubated with 50 mM Ammonium chloride in PBS for 10 min, washed in 1x PBS, permeabilised with PBS/0.1% Tween for 5 min and blocked with PBS/10% FBS for 2 hours at 4°C. Primary antibodies were incubated overnight at 4°C in PBS/0.15% FBS. Alexa Flour 488-, 555-, or 647-conjugated secondary antibodies (Invitrogen) were incubated in PBS/0.15% FBS at room temperature for 1.5 hours in a humidified chamber. Cells were washed 3 times for 5 minutes with PBS/0.1% Triton-X 100 between antibody incubations. Nuclei were counterstained and mounted with 6-diamidino-2-phenylindole (DAPI)-containing Mowiol (Merck Millipore).
For confocal imaging, slides were processed on an Olympus microscope, using 60x lens. Samples with 1.5-mm coverslips were imaged using a an FV1000 confocal system on an Olympus IX-81 microscope with photomultiplier tube detectors and Olympus PlanApo N, 60×/1.35NA lens at RT. DAPI-containing Mowiol (EMD Millipore) was used as the imaging medium. DAPI; Alexa Fluor 488, 539, and 635 (Thermo Fisher Scientific) channels were used for acquisition with Olympus Fluoview software. Experimental settings including values of the laser power for each channel, HV, gain and offset parameters were determined at the beginning of each individual imaging session (by assessing background reactivity and saturation levels of each channel) and kept constant over the entire imaging session. ImageJ software (NIH) was used for further processing of the images. All quantifications represent a number of cells that have shown phenotype or % of positive cells, see figure legends for details. n = number of cells. Signals were quantified using RGB profiler (ImageJ, NIH). For super-resolution microscopy, cells were imaged at room temperature using an inverted Zeiss 880 microscope fitted with an airy-scan detector. The system was equipped with Plan- Apochromat 63x/1.4 NA oil lens. 488 nm argon and 405nm, and 633 nm diode lasers were used to excite GFP, DAPI and Alexa Fluor 633, respectively. Sequential excitation of each wavelength was switched per line to ensure blue, green and red channels were aligned. Sections of 20 slices with 0.5 μm thick intervals were collected with a zoom value of 600 pixels/μm. Images were processed using Airyscan processing (Zeiss 880 Airyscan: Airyscan is a special detector added to the coupling port of LSM 880. It is more light efficient than a standard confocal point detector. The extra on photons can be used to increase the sensitivity of the image, to scan faster or to improve the resolution) in 3D with a strength value of Auto (~6). 3D representation were generated using Imaris 8.4.1 (Bitplane, Oxford Instruments).
Approximately 3x 105 murine wild type PMEF or PMEF::HA-Dicer cells were grown to 70–80% confluency on 6-well multi-well dishes (Corning). Approximately 2x 105 wild type human HEK293 cells were seeded on top of the PMEF layer prior to membrane fusion. Mixed cell populations were grown in presence of Cycloheximide (50 μg/ml) for 4 hours prior to fusion. For heterokaryon formation, cells were washed with warm 1x PBS, incubated with 100 μl warm PEG-3000 solution (50% w/v in PBS) for 2 minutes and washed with 1x PBS five times. Heterokaryons were cultured for additional 2 hours in Cycloheximide- containing medium in presence or absence of Etoposide (25 μM) prior to fixation. Alexa Fluor 647-conjugated Phalloidin (Life Technology) was used to stain the cytoskeleton. Preparation of immunofluorescence slides was performed as described in section imaging analysis.
|
10.1371/journal.pcbi.1004345 | Deciphering Transcriptional Dynamics In Vivo by Counting Nascent RNA Molecules | Deciphering how the regulatory DNA sequence of a gene dictates its expression in response to intra and extracellular cues is one of the leading challenges in modern genomics. The development of novel single-cell sequencing and imaging techniques, as well as a better exploitation of currently available single-molecule imaging techniques, provides an avenue to interrogate the process of transcription and its dynamics in cells by quantifying the number of RNA polymerases engaged in the transcription of a gene (or equivalently the number of nascent RNAs) at a given moment in time. In this paper, we propose that measurements of the cell-to-cell variability in the number of nascent RNAs provide a mostly unexplored method for deciphering mechanisms of transcription initiation in cells. We propose a simple kinetic model of transcription initiation and elongation from which we calculate nascent RNA copy-number fluctuations. To demonstrate the usefulness of this approach, we test our theory against published nascent RNA data for twelve constitutively expressed yeast genes. Rather than transcription being initiated through a single rate limiting step, as it had been previously proposed, our single-cell analysis reveals the presence of at least two rate limiting steps. Surprisingly, half of the genes analyzed have nearly identical rates of transcription initiation, suggesting a common mechanism. Our analytical framework can be used to extract quantitative information about dynamics of transcription from single-cell sequencing data, as well as from single-molecule imaging and electron micrographs of fixed cells, and provides the mathematical means to exploit the quantitative power of these technologies.
| Gene expression starts with transcription, a multi-step process that produces an RNA molecule that is complementary to the gene. Cells often control the amount of gene expression by controlling the amount of RNA produced through interactions between regulatory DNA and proteins involved in transcription. While the identity of the molecules that take part in this regulatory process is known for a number of different genes, their dynamics in cells is still poorly understood. We show theoretically that the cell-to-cell variability in the number of nascent RNA molecules, those still in the process of being synthesized by the RNA polymerase, carries the signature of transcriptional dynamics in cells. We analyze published nascent RNA distributions for a set of yeast genes and show that the data is inconsistent with a single-step model of transcription initiation. Instead we propose a coarse-grained model where initiation happens not in one but in two sequential steps. Our analytical framework can be used to extract quantitative information about the dynamics of transcription from single-cell sequencing data, as well as from single-molecule imaging and electron micrographs of fixed cells.
| Transcription is a multi-step process that leads to the production of messenger RNA (mRNA) molecules from its DNA template. Genetic experiments on cells have identified the key molecular components of transcription, while biochemical studies with purified components have uncovered the basic mechanisms governing their dynamics and interactions in vitro. Still an important question that remains is whether the same mechanisms are also operational in cells. One approach to unraveling the mechanisms of transcription in cells is to measure the outputs of this process, either the proteins that correspond to the genes being transcribed, or the actual mRNA molecules. This idea has motivated numerous experiments that count protein [1–3], and mRNA [4–6] molecules in single cells. The measured steady state distribution of these molecules in a clonal cell population can then be used to infer the dynamics of transcription [4,5]. For instance, analysis of the steady state distributions of cytoplasmic mRNA in yeast for a number of different genes, have suggested that yeast genes may fall into two different classes: those that are transcribed in random uncorrelated events clearly separated in time and without any transcriptional memory [4] (this is often referred to as Poissonian transcription), and those that are transcribed in bursts caused by the promoter switching slowly between an active state and an inactive state (this is often referred to as Bursty-transcription [7,8]).
While this approach to deciphering transcriptional dynamics in vivo by counting cytoplasmic RNA in single cells has led to important insights, a key limitation is that processes that are downstream from transcription initiation can mask the signature of transcriptional dynamics in measurements of the cell-to-cell variability of mRNA and protein abundances. A striking example of this is the recent finding that spatial and temporal averaging, i.e., the process of accumulation and diffusion of mRNA transcripts during nuclear cycles, significantly reduces the variability in mRNA copy number expected from stochastic transcription initiation [9]. In addition, effects such as mRNA transport out of the nucleus, mRNA processing, and nonlinear mRNA degradation [10–14] can also in principle affect the level of variability of cytoplasmic mRNA. All of these non-transcriptional sources of variability may propagate to the protein level as well, affecting the cell-to-cell fluctuations in protein copy number, which is also affected by the stochastic nature of translation. Finally, it has been recently shown that partitioning of both mRNA and protein molecules during cell division [15–17] can generate distributions in their abundances similar to those that would be generated by stochastic transcription and translation. Therefore, the cell-to-cell variability of both protein and cytoplasmic mRNA copy number do not necessarily reflect transcriptional dynamics alone but are determined by a combination of stochastic processes of which transcriptional dynamics is just one component [18].
One alternative to analyzing steady state mRNA and protein distributions, has been to directly image transcription in real time using fluorescently labeled RNA-binding proteins that associate with nascent RNA, which is still in the process of being assembled at the gene by the RNA polymerase [19–23]. When applied to E. coli, Dyctostelium or animal cells this technique revealed widespread transcriptional bursting consistent with the mechanism of transcription initiation where the promoter switches between an active state and an inactive state [7]. In contrast, in experiments on two constitutive and cell-cycle activated genes in S.cerevisiae, Larson et al. [23] found that the transcription initiation process is dominated by one rate limiting step. In spite of the great promises of this approach, it is technically challenging and still remains in its infancy.
Lately, a score of experimental papers have reported measurements of distributions across a clonal cell population of nascent RNA transcripts at a single gene, using single-molecule fluorescence in-situ hybridization (FISH) [4,24–26] (Fig 1B). These experiments reveal the number of RNA polymerases engaged in transcribing a single gene in a single cell at a specific instant in time. This information can also be obtained from so-called Miller spreads (electron micrographs of intact chromosomes extracted from cells) which provide images of transcribing polymerases along a gene [27–31] (Fig 1A). Perhaps more importantly, single-cell whole genome RNA sequencing is slowly but steadily being developed and turned into a quantitative technique, one which will be able to provide a snapshot of the number of RNA polymerase molecules engaged in the transcription of every gene in the cell at a given instant in time (Fig 1C) [32–36]. Counting nascent RNAs (or the number of transcribing polymerases) provides a more direct readout for the transcriptional dynamics at the promoter within the short window of time required for an RNA polymerase molecule to complete elongation (for a typical gene in yeast the elongation time is of the order of few minutes [4]). As such, this experimental approach is not affected by the aforementioned stochastic processes that contribute to cytoplasmic mRNA and protein fluctuations. Indeed, as mentioned above, strong discrepancies between cytoplasmic and nascent mRNA distributions have been recently found in Drosophila embryos [9]. Below we also demonstrate similar discrepancies in yeast by analyzing published data obtained from counting nascent and cytoplasmic mRNA in single cells.
It is thus starting to become possible to obtain quantitative measurements of the distribution of nascent RNA or, what is the same, of the number of transcribing polymerases per gene. In spite of its many advantages, the potential of nascent RNA distributions has not been fully exploited (for a notable exception see [37]) due to the lack of mathematical formalisms that allows one to connect molecular mechanisms of transcription initiation and elongation with measured nascent RNA distributions. One of the key results that we report here is the development of such formalism. In particular, we show how to compute the mean and variance of the distribution of nascent RNAs for an arbitrary mechanism of transcription initiation and stochastic elongation. The results of these calculations provide the tools to extract information about transcriptional dynamics from experimentally determined nascent RNA distributions. We demonstrate the usefulness of our method by analyzing published nascent RNA distributions for a set of constitutively expressed yeast genes [25]. We find that all of these yeast genes have similar average initiation rates. We also find that initiation of transcription of these yeast genes is a two-step process, where the average durations of the two steps are equal. This is in sharp contrast to the conclusion that was reached for some of these genes by counting cytoplasmic mRNAs, namely that transcription initiation is dominated by one rate limiting step [25]. By analyzing the nascent RNA distribution, we are able to reach a level of kinetic detail, particularly fast processes, which are obscured at the level of cytoplasmic RNA. While the molecular identities of the two steps leading to transcription initiation remain unknown, our results point to the existence of multiple transcription initiation steps in vivo. It is worth emphasizing that multiple initiation steps of similar duration lead to a reduction of fluctuations in the number of nascent RNAs in a cell, when compared to those produced by single-step initiation.
In order to connect mechanisms of transcription initiation with nascent RNA distributions, we consider a model of transcriptional dynamics with an arbitrarily complex initiation mechanism followed by an elongation process. We describe both processes using chemical master equations. This approach is inspired by the work of Kepler et al. [38] who computed the moments of the mRNA distribution for a promoter, where it switches between an active and an inactive state. We have previously developed this method further to compute the moments of mRNA and protein distributions for arbitrarily complex promoters that can switch between multiple states, each state leading to transcript production at a particular rate [8,39–42]. Here we implement the same master equation approach to compute the first and second moments of the nascent RNA distribution. A new element in our analysis is the explicit inclusion of the stochastic elongation process, which predicts that the nascent RNA distributions depend on the length of the gene being transcribed, for which we find confirmation in published data. This dependence of the distribution of nascent RNAs on gene length has also been described recently in [37]. Our theory also suggests new experimental approaches to deciphering the dynamics of transcription initiation in vivo, in which the length of the transcribed gene is varied and the effect on the number of nascent RNAs is measured.
To describe the transcription initiation process we focus on promoter dynamics. (Here we use the term promoter to denote the stretch of regulatory DNA that controls the initiation of transcription of a specific gene.) The promoter switches between different states as different transcription factors bind and fall off their respective binding sites, causing the effective initiation rate to fluctuate. We assume that after initiation, each RNA polymerase (RNAP) moves along the gene by stochastically hopping from one to the next base at a constant probability per unit time (Fig 2A). Our model assumes that transcription initiation timescales are much slower than the elongation timescale and hence RNAPs do not interfere with each other while moving along the gene. This approximation is reasonable for all but the strongest promoters characterized by very fast initiation [43,44]. We demonstrate this explicitly using numerical simulations [45,46] which include a detailed model of transcription elongation that takes into account excluded-volume interaction between adjacent polymerases (i.e. “traffic” as defined in previous work [43]), as well as ubiquitous RNAP pausing [43,47] (please see S1A Fig). The agreement between analytical results based on our simple model and the stochastic simulations of the more realistic model that incorporates traffic jams and pausing of RNAPs only starts to break down when the initiation time scales become comparable to the elongation time scales (please see S1C and S1D Fig). We conclude that for typical rates reported for RNAP elongation and pausing the simple model of transcription adopted here reproduces the first two moments of the nascent RNA distribution with deviations from those obtained from the more realistic model that are less than 10% as long as initiation of transcription is slower than 30 initiations/min. All the initiation rates that have been reported so far from in vivo measurements are slower [4,19,23], with important exceptions such as the ribosomal promoters [43,44].
Our model does not explicitly include the rate of termination at which the RNAP departs the last base of the gene. The genes [25] that we analyze have an average initiation rate of the order of kINI = 0.145±0.025/min. Hence even for a termination rate of the order of 1/min [23], the variance and mean of the nascent distribution won’t be affected for these genes. Another simplifying assumption that we make is that we place no restriction on the number of transcribing RNAPs that can occupy a given base (in reality at any given instant the number is zero or one). This is equivalent to assuming that the occupancy of any given base of the gene by a transcribing polymerase is much less than one, which holds when the initiation time scale is much slower than the elongation time scale. Hence, despite its simplicity, the model of transcription initiation and elongation we adopt here should apply to most genes.
In order to compute the first two moments of the nascent RNA distribution for an arbitrary transcription initiation mechanism, we consider a promoter that can exist in N possible states. The rate of transition from the s-th to the q-th state is ks,q, and the rate at which RNAP initiates transcription from the s-th promoter state is ks,ini. Following the initiation process, every RNAP moves along the gene (elongates) by hopping from one base to the next with a probability per unit time k, which is equal to the average rate of elongation. The number of RNAP molecules, which is the same as the number of nascent RNAs, at the i-th base pair is denoted by mi. Hence the number of nascent RNAs (M) along a gene whose length is L bases, is given by,
M=∑i=1Lmi.
As remarked earlier we do not consider the processes of transcription termination and mRNA release, as they tend to be fast on the time scales set by initiation and elongation. However these can be easily incorporated into the model. (For the mathematical details please see the S1 Text.) The state of the combined promoter+RNA system is described by (L+1) stochastic variables: the number of nascent RNAs (m1,…,mL) at every base along the gene, and the label s, characterizing the state of the promoter. Hence, the probability distribution function that characterizes the promoter+RNA system is given by P(s,m1,…,mL). To stream-line the mathematics we define the following probability vector:
P→(m1,…,mL)=(P(1,m1,…,mL),P(2,m1,…,mL),…,P(s,m1,…,mL)).
(1)
The time evolution for this probability vector can be described by a set of chemical master equations, which can be written in compact, matrix form as
dP→(m1,…,mL)dt=(K^−R^−Γ^∑i=1Lmi)P→(m1,.,mi,.,mL)+R^P→(m1−1,…,mL)+∑i=1L−1k(mi+1)Γ^P→(m1,.,mi+1,mi+1−1,.,mL)+k(mL+1)Γ^P→(m1,…,mL+1).
(2)
In Eq (2), we define the following matrices:
K^
, which describes the transition between different promoter states, and whose elements are Kqs = kq,s if q≠s and
Kss=−∑qkq,s
.
R^
is a matrix that contains the rates of initiation from different promoter states. In the case of one-step initiation it is diagonal with the diagonal elements equal to the rates of initiation from different promoter states. In the case of two-step initiation this matrix is off-diagonal owing to the fact that the promoter state changes after initiation (for details please see the S1 Text).
Γ^
is also diagonal and its elements represent the hopping rate for the polymerase from one base pair to the next, i.e., Γsq = k δs,q.
We limit our calculation to the steady state nascent RNA distribution for which the left hand side of Eq (2) is set to zero. To obtain the first and second moments of the number of nascent RNAs,
M=∑i=1Lmi.
in steady state we use Eq (2) to compute the quantities 〈mi〉 and 〈mimj〉 for all i,j ≤ L. Even though the random variables mi for different bases i on the gene are mutually dependent, we end up deriving a set of linear equations for 〈mi〉 and 〈mimj〉 (Please see the S1 Text.) We find that these equations for the moments close, in other words they do not depend on any further, higher moments of the mi’s. These linear equations can then be solved to obtain exact expressions for the first two moments of M as a function of all the rates that define the molecular mechanism of initiation under investigation. (For the mathematical details please see the S1 Text.)
In order to demonstrate how the distribution of nascent RNAs at the transcription site can be used to extract dynamical information about the process of transcription initiation in vivo, we consider the canonical model of transcription shown in Fig 2A [38]. The gene can switch between two states: an active state, from which transcription initiation can occur, and an inactive state from which initiation does not occur. The two states might correspond to a free promoter and one bound by a repressor protein, or a promoter occluded by nucleosomes. In most theoretical studies to date transcription initiation from the active state was assumed to be characterized by a single rate-limiting step. Instead of initiation being a one-step process we consider the possibility that there are two rate-limiting steps involved in transcription initiation from the active state. These could represent the loading of the transcriptional machinery at the promoter [48,49] (in prokaryotes, this would correspond to the formation of open complex by RNAP [50–52]), which occurs with a rate kLOAD followed by the RNA polymerase escaping the promoter into an elongation state (with rate kESC).
Three different limits of our model correspond to the various scenarios that have been previously explored in the literature [4,19,53–55]. First we consider the limit when the promoter is always active (kOFF → 0 in Fig 2A) and initiation is governed by a single rate-limiting step. This is a situation when one of the two kinetic steps leading up to initiation (either the assembly of the transcriptional machinery or the escape of RNA polymerase from the promoter) is much slower than the other. In this case we find that the nascent RNA distribution is characterized by a variance that is equal to the mean. In other words the Fano factor, defined as the variance divided by the mean, to characterize cell-to-cell variability is 1.The second limit of interest is when the rates of assembly of the transcriptional machinery (kLOAD) and promoter escape (kESC) have comparable magnitudes, i.e., transcription initiation is a two-step process. In this limit, transcription initiation events are anti-correlated due to the presence of a “dead-time” or refractory period in between subsequent initiation events. The third limit of interest is the “transcriptional bursting”, when the promoter is not always active, but is slowly switching between the active and inactive states [7]
A key prediction of our model of stochastic transcription initiation and elongation, which is described in Fig 2A, is how the cell-to-cell fluctuations of the nascent RNA number depend on the length of the gene being transcribed. This dependence was also explored in [37] where the importance of the elongation rate and gene length in determining nascent RNA distributions was described, and nascent RNA distributions were used to infer the kinetic rates. Gene length is an interesting quantity to consider from the point of view of experiments, both due to the natural variation in gene length, and the ability to synthetically alter the length of the gene being expressed from a promoter of interest by genetic manipulation. Calculations of the Fano factor as a function of gene length (Fig 2B) reveal that this quantity easily discriminates between the three models of transcription initiation described above. When the gene length is small the Fano factor is close to one for all three models of initiation. As the gene length increases, the Fano factor increases above one for the “bursting” scenario, due to slow switching of the promoter between an active and an inactive state, but it decreases below one when the promoter is always active and there are two rate limiting steps leading up to elongation. Finally, in the case when initiation is dominated by one rate-limiting step and the promoter is always active, the Fano factor is equal to one, independent of gene length.
Qualitatively these results can be understood by recalling that with a single initiation step, the waiting time between initiation events is exponentially distributed. In this case the number of initiation events to occur in a time interval set by the elongation time (which is roughly equal to the number of nascent transcripts) is given by a Poisson distribution [56], for which the Fano factor is one. For two or more rate limiting steps leading to initiation, the waiting time between successive initiation events is gamma distributed [55]. As a result the distribution of nascent RNAs is expected to be narrower than Poisson with a Fano factor less than one. The presence of transcriptionally inactive states on the other hand has the effect of broadening the distribution of nascent RNAs, and should lead to a Fano factor greater than one in the case when initiation from the active state is a one-step process.
For bursty promoters that switch between an active and an inactive state (for example the PDR5 gene in yeast [4]) the nascent RNA distribution can also be used to discriminate between different mechanisms of regulation. Recent experiments [57,58] have suggested that transcriptional regulation may be achieved by either modulation of the burst size (given by kINI/kOFF, where kINI = kLOAD × kESC/ (kLOAD + kESC) is the average rate of initiation), or by modulating the burst frequency (kON); it is also possible that both are tuned [7]. In Fig 2C we show the results of our calculations of the Fano factor for the nascent RNA distribution, using parameters that are characteristic of the PDR5 gene and assuming that transcriptional regulation is achieved either by tuning the burst size or the burst frequency. We see that even though both mechanisms of regulation produce Fano factors larger than one, they make qualitatively different predictions for the functional dependence of the Fano factor on the mean number of nascent RNAs.
Nascent RNA distribution is determined by stochastic initiation and elongation. However for a long gene, the elongation process becomes practically deterministic due to the law of large numbers. Assuming as we do in our model that each elongation step is a stochastic process, with the same rate k,the elongation time will be Gaussian distributed with a mean and variance that are proportional to the length of the gene. Therefore the deviation of the elongation time away from the mean compared to the mean will decrease as the square root of the gene length. A recently published paper by Senecal et al.[37] has explored how stochastic initiation and deterministic elongation processes affect nascent RNA distribution.
However, and although this is indeed the case for FISH data such as the one analyzed in Fig 3, in the paper, an important application of our method will be the analysis of single-cell sequencing data, where the positions of every polymerase along the gene can be determined. In addition, and as shown in Fig 1, we anticipate that our method can be applied to electron micrograph data (e.g. such as those reported in [29,31]), a method that also allows one to measure the position of each polymerase along a transcribed gene. Using this technique, the number of polymerases in the first L nucleotides of the gene can also be determined, and statistics (mean and variance) can be computed and compared to experimental results. This will allow us to computationally bin a gene into smaller chunks of arbitrary length, and use length as a “data analysis” turning knob that we can tune computationally to investigate how it affects the noise(Fano factor). To be able to do this, a fluctuating elongation rate is essential, since in principle the gene length can be made as short as desired during data analysis.
In the limit of a long gene, when the residence time of the RNAP on the gene is practically deterministic, we can use queuing theory to compute closed form expressions for the first and second moments of the nascent RNA distribution [59,60]. For the one-step model, the promoter is always active (kOFF → 0 in Fig 2A) and there is a single rate-limiting step leading up to initiation. In this case the nascent RNA distribution is characterized by a variance that is equal to the mean, which is what we computed for stochastic elongation as well. The second limit of interest is when the rates of assembly of the transcriptional machinery (kLOAD) and promoter escape (kESC) have comparable magnitudes, i.e., transcription initiation is a two-step process. For an elongation time T = L/k (where L is the number of bases along a gene and k is the average rate of elongation) the mean and variance are given by
⟨M⟩=kLOADkESCTkLOAD+kESCVariance=⟨M⟩[1−kESCkLOADT(kESC+kLOAD)+kESCkLOADT(kESC+kLOAD)2[2(1−exp(−(kESC+kLOAD)T)(kESC+kLOAD))−2T+T2(kESC+kLOAD)]].
The Fano factor can hence be computed very easily by taking the ratio of the variance and mean. It is to be noted that when one of the rates that describe the two steps leading to initiation (kLOAD and kESC) becomes much smaller than the other, we are back to the case of one rate-limiting step. In the case of one rate-limiting step the Fano factor becomes one, which is the signature of a Poisson initiation process. However when the rates kLOAD and kESC become comparable, the Fano factor is reduced from 1 and attains a minimum value when kLOAD is equal to kESC. This also follows from the equations for the variance and the mean and can be intuited by noting that the two rates appear in the equations in a symmetrical fashion.
The third limit of interest is the ON-OFF model of initiation which is characterized by kOFF (the rate of promoter switching from the ON to OFF state), kON (the rate of promoter switching from the OFF to ON state), kESC (the rate of escape, which we assume is much higher than the rate of assembly of the transcriptional machinery), and time of elongation T. The mean and variance are given respectively by,
⟨M⟩=kONkESCTkON+kOFFVariance=⟨M⟩[1+2kESCkOFF(kON+kOFF)2+2kESCkOFF(kON+kOFF)3(exp(−(kON+kOFF)T)−1T)].
As shown in S4 Fig, these formulas give almost identical results to those we obtain when taking into account stochastic elongation, when the gene length is of the order of few thousand bases.
The theoretical results described above can be used as a mathematical tool to extract information about transcription initiation dynamics from nascent RNA distributions, which have been measured in a series of recent experiments [4,25]. To demonstrate the utility of this approach, we analyze a set of nascent RNA distributions for twelve different constitutively expressed genes in yeast [25]. We find that for six of these twelve genes (RPB2, RPB3, TAF5, TAF6, TAF12, KAP104), the mean number of nascent RNAs scales linearly with the gene length, as shown in Fig 3A. If we assume that all of these genes have comparable elongation rates (k = 0.8 kb/min (4)), then the linear relationship between the mean nascent RNA number and gene length implies that the average initiation rates of these genes are all roughly the same and equal to kINI = 0.145±0.025/min.
In addition to the mean, our model allows us to investigate the behavior of the variance of the nascent mRNA distribution with gene length, and compare it to the predictions from different models of transcription initiation. Given that the Fano factors of the nascent RNA distribution for the six genes, RPB2, RPB3, TAF5, TAF6, TAF12, and KAP104, are all less than one, the simplest model consistent with the data is one where the promoter is always active and transcription initiation is a two-step process (see Fig 2A).
This model is parameterized by the rates kLOAD and kESC. The Fano factor of the nascent RNA distribution depends on the ratio of these two rates. Our model makes prediction for how the Fano factor changes with the gene length when the rates kLOAD and kESC are tuned, consistent with the mean initiation rate being kINI = kLOAD × kESC/ (kLOAD + kESC) = 0.145±0.025/min. As shown in Fig 3B, this value of the initiation rate defines a region in the Fano factor–Gene length phase (light-blue shaded area in Fig 3B). This region is bounded on its upper side by the limit when one of the rates (either kLOAD or kESC) is much larger than the other one (which turns initiation into a one-step process with a Fano factor equal to one), and on the lower side, by the limit in which the two rates are identical. The limit of identical rates gives the minimum Fano factor attainable when the average initiation rate is 0.145±0.025/min. Remarkably, we find that the six genes in question have the lowest possible Fano factor. In principle, the six genes shown in Fig 3A could have ended up anywhere within the shaded region in Fig 3B. The fact that they all follow the lower boundary of the allowed region suggests that these genes, which have varying length, have not only the same average initiation rate, but also that they have identical promoter cycling kinetics, with roughly the same values of kLOAD and kESC (kLOAD = kESC = 0.29±0.013/min).
It is to be noted that a multi-step initiation model, where initiation happens in more than two sequential steps (this can also include bursting kinetics) can also account for the Fano factor for the six different genes being less than one. However a more complicated model with more than two sequential steps will have many free parameters e.g. for a three step model we will have three sequential steps to initiation characterized by three different rates. Although we cannot rule out such possibilities, the two-step model in spite being the simplest possible scenario explains the data well and provides mechanistic insight into the dynamics of initiation for the six different genes. The key result here is that by analyzing nascent RNA distributions, we can exclude the one-step and ON-OFF models of initiation.
The remaining six (RPB1, MDN1, PUP1, PRE7, PRE3, PRP8) constitutive genes of the twelve studied [25] initiate at rates that are different than the rate of initiation that we found for the six genes discussed above (see S2 Fig). All but one of these six genes have nascent RNA Fano factors that are less than one, consistent with two or more steps leading up to initiation. This second set of genes thus acts as a control group that, as expected for a set of genes having different gene-specific rates of transcription, occupies the allowed region in the Fano factor-Gene length phase space without clustering at the lower boundary of this region, like we found for the six genes discussed above (S3 Fig).
Direct imaging of transcriptional dynamics in real time [19–23] at the molecular scale and in individual cells still remains challenging. As an alternative, a number of recent studies have tried to decipher the dynamics of transcription initiation using the measured cell-to-cell variability of transcriptional outputs (cytoplasmic messenger RNA or protein molecules) at the single cell level [1,3,4,25,61]. These measurements of transcriptional cell-to-cell variability have been interpreted in the context of a classification scheme for promoters, which are characterized by either a Poisson or a Gamma distribution of their outputs. These differences have then been taken to indicate a difference in the mechanism of transcription. A Poisson distribution is taken as evidence that the promoter transcribes at a constant rate, i.e., initiation is a one-step process. The Gamma distribution on the other hand is indicative of bursty promoter dynamics [4,19]. In practice, the distribution of cytoplasmic mRNA or proteins obtained from a population of cells is fitted to a mathematical model that incorporates the stochastic kinetics of transcription (and translation in the case of proteins), and the fitting parameters are interpreted as representative of the kinetic properties of stochastic gene expression (e.g., burst size, burst frequency, average transcription rate, etc.) [5]. Even though in some cases this approach has produced kinetic parameters whose values are consistent with direct measurements of the same parameters [1], the interpretation of the kinetic parameters can be difficult given that the distributions of mRNA and protein may be affected by stochastic processes that occur downstream of transcription initiation. Examples of these processes include the non-linear degradation of mRNA and proteins [14], maturation time of fluorescent reporters [62], transport of mRNA out of the nucleus [10,11], mRNA splicing [12,13] and small RNA regulation [14,63]. Furthermore, recent theoretical results [15,16] indicate that fluctuations due to random partitioning of molecules during cell division may yield the same mathematical dependence between variance and mean of protein and mRNA copy number in clonal cell populations, as would a stochastic model of transcription initiation and linear degradation.
In order to demonstrate that the distribution of mRNAs can be affected by stochastic processes that occur downstream of transcription, thereby obscuring the signature of transcription initiation dynamics, we compare the nascent RNA and cytoplasmic mRNA distributions for the twelve yeast genes analyzed in Fig 4. First, we compute the Fano factor of the cytoplasmic mRNA distribution predicted by the initiation mechanism inferred from the measured nascent RNA distribution for all twelve genes studied (23). (See the S1 Text for details of the calculation.) Then we compare the results of our calculations with the experimentally determined distributions obtained by counting cytoplasmic mRNA. We find that for all of the yeast genes examined the predicted Fano factors for the cytoplasmic mRNA distributions are less than the measured ones, as shown in Fig 4. In other words the signature of two-step initiation observed in the nascent RNA distribution is washed out at the cytoplasmic mRNA level due to other sources of noise. It remains unclear what processes are responsible for these differences. In a recent study of transcription in fly embryos, it was also found that the variability of nascent and cytoplasmic mRNA could differ more than six fold [9]. In this case, the reason for this difference is spatial and temporal averaging of mRNA by diffusion and accumulation of mRNA transcripts during nuclear cycles. The yeast and fly examples demonstrate that the relationship between nascent and cytoplasmic RNA distributions is complex and context dependent.
An alternative to counting cytoplasmic proteins or mRNA is to count the number of transcribing polymerases [27–30], or nascent RNAs [4,25,37] on the gene being transcribed, using electron micrographs and fluorescence in situ hybridization, respectively. These measurements are not affected by post-transcriptional processes and are more direct readouts of transcriptional dynamics. To date, these distributions of nascent RNAs have been used mostly in a qualitative manner, due to the lack of mathematical models that connect these distributions with the underlying mechanisms of transcription, apart from the recent paper by Senecal et al. [37]. For instance, distributions of nascent RNAs (or of transcribing RNA polymerases) have been recently reported in yeast [4,23,25], fly embryos [9,49,64], and bacteria [29,65]. The model of transcription initiation and elongation developed here offers a way to quantitatively analyze these measured nascent RNA distributions, and connect them to molecular mechanisms of transcription. In particular, when we consider three different models of transcription initiation that incorporate three broad classes of initiation mechanisms, we find that they make qualitatively different predictions for nascent RNA distributions.
Analyzing the nascent RNA distributions for twelve constitutively expressed genes in yeast [25], we find that all but one of these distributions have a Fano factor less than one. This observation is consistent with a simple model in which initiation proceeds in two-steps (for some of the genes more than two steps are implicated by the data; see S3 Fig), which are of similar duration. The two rate limiting steps can arise from a number of different sources. For the genes in yeast considered in the paper the initiation complex is formed by assembly of multiple transcription factors and co-factors [66]. After the formation of the initiation complex, an RNAP molecule initiates transcription by escaping the promoter. The two-step model we consider in the paper would be realized if out of all these steps leading up to initiation any two steps become rate-limiting. The most surprising finding when analyzing these twelve genes was that six of them have not only the same average initiation rate, but also the same rates of loading of the transcriptional machinery, and of promoter escape. We do not have a mechanistic interpretation for this finding, but the data suggests the existence of a common molecular mechanism of initiation for these six genes, and given that they represent half of the genes in the data set we have analyzed here, it is tempting to speculate that other yeast genes may share the same kinetics. More experiments are clearly needed to test this hypothesis, ideally ones where the dynamics of transcription are followed directly [23].
Our findings for the yeast promoters, highlight the utility of our theory for deciphering transcriptional dynamics in vivo from nascent RNA distributions. In addition, counting nascent RNAs, mRNAs and proteins simultaneously will undoubtedly further enhance our understanding of how the central dogma of molecular biology plays out in individual cells.
To compute the first two moments of the nascent RNA distribution for the canonical model of transcriptional regulation shown in Fig 2A we apply the general method of deriving moment equations from the master equation, Eq (2). The rate matrices that define the master equation, Eq (2), are in this case:
K^=[−kONkOFF0kON−(kOFF+kLOAD)kESC0kLOAD−kESC],R^=[00000kESC000],Γ^=[k000k000k].
Here
K^
, is the transition matrix, which describes promoter switching between the three possible states shown in Fig 2A. When an RNA polymerase initiates transcription from the state in which the polymerase is bound to the promoter, the state of the promoter changes to the state in which the promoter does not have a bound polymerase. This accounts for the rate of escape appearing in the transition matrix and also explains why
R^
(the initiation rate matrix) is not diagonal. Using these matrices in the master equation for the nascent RNA distribution (Eq (1)) we compute analytically the mean and the variance of the distribution as a function of the gene length L. These results were used to make the plots in Fig 2B and 2C.
Our model (S1B Fig) makes the assumptions that RNAP molecules do not pause and do not collide with other RNAP molecules, while moving along the gene. We also take the size of the RNAP footprint to be one base, and we do not restrict the number of RNAPs at each base along the gene. These assumptions are equivalent to the assumption that the average number of transcribing RNA polymerases is much less than one per base. If we consider a constitutive (one-step) promoter with an initiation rate kESC, and the rate at which the RNAP translocates from one base to the next is k, then the number m of RNAP molecules on the first base of the gene would be Poisson distributed [40] and given by,
P(m)=(kESCk)mm!e−kESCk.
As the above equation demonstrates, if the ratio of initiation rate and hopping rate kESC/k is of the order of 0.01 (characteristic of MDN1 promoter [25)], the probability of finding two or more RNAP molecules at the first base of the gene would be 5×10−5. This justifies one of the main assumptions of our model, namely that we can ignore the constraint that no base can be occupied by more than one polymerase. The assumption will be valid as long as the initiation time scale is slower than the elongation time scale, and it makes the model analytically tractable. As described in the results section, this assumption leads to simple formulas for the first two moments of the nascent RNA distribution in the large gene-length limit, and to a set of L2 linear equations in the case of stochastic elongation. As shown in the S1 Text, these linear equations are readily solved to obtain the moments of the nascent RNA distribution using standard computing tools such as Mathematica or Matlab. This is important in order to test many parameter sets without having to run a new Gillespie simulation for every set which can be impractical for complex kinetic mechanisms of transcription initiation.
As argued above, we expect the approximations made in our model to be reasonable for all but the strongest promoters characterized by very fast initiation [43,44].
In order to test this intuition, we compare the analytic predictions of our model with numerical simulations of a more realistic one (referred to as the traffic model in S1A Fig), which properly accounts for the footprint of a transcribing RNAP molecule on the DNA, ubiquitous pausing of the polymerase, and excluded volume interactions between adjacent polymerases along the gene. In particular we compare the mean and the Fano factor of nascent RNA distributions, as predicted by our model of transcription for the case when initiation occurs via a single rate limiting step, with those obtained from numerical simulation of the traffic model obtained using the Gillespie algorithm [45,46].
A single time step of the simulation is performed in the following way: one of the set of all possible reactions is chosen at random according to its relative weight, which is proportional to the rate of the reaction, and the state of the system is updated by implementing the change described by the chosen reaction. The time elapsed since the last step is drawn from an exponential distribution, the rate parameter of which equals the sum of all the rates of the possible reactions at that time. This process is repeated for a long enough time such that the number of RNAP molecules along the gene (which is the same as the number of nascent RNAs) reaches steady state.
We consider four different transcription initiation rates, spanning the typically observed values in E. coli and yeast cells [4,19,23,25], and we observe in the simulations how the mean and Fano factor of the nascent RNA distribution are affected by RNAP pausing and road blocking (S1A and S1B Fig). We find that for initiation rates slower than 30 initiations/min, both the mean and the Fano factor extracted from the simulations are in good agreement (less than 10% difference) with the analytical results (S1C and S1D Fig). In simulations we used the following parameters to describe RNAP elongation: kP- = 4/sec, kP+ = 0.01/sec, k = 80 bp/sec, as was reported for ribosomal promoters in E.Coli [43]. We also use a gene of length L = 2000 bases and a polymerase whose DNA footprint is 30 bases.
We generated the plots for the Fano factor versus gene length (Fig 2B), for the three limits of the model in Fig 2A using the parameters listed below. For the bursty promoter, where the promoter slowly switches between inactive and inactive states, we use kOFF = 5/min, kON = 0.435/min, k = 0.8kb/min, kLOAD = 5/min and kESC = 0/min; kOFF, kON, k, and kLOAD are the characteristic rates for the PDR5 promoter, as reported in [4]. For the two-step initiation model, where the promoter does not switch between an active and an inactive state but has two rate limiting steps leading up to initiation, we use kLOAD = 0.14/min, kESC = 0.14/min, kOFF = 0/min, kON = 0/min, k = 0.8kb/min; these are characteristics of yeast genes, such as MDN1 [25]. For the one-step model, there is one rate limiting step leading up to transcription elongation and we choose kLOAD = 0.09/min, kESC = 0/min, kOFF = 0/min, kON = 0/min, k = 0.8kb/min, which are characteristics of the yeast gene RPB1 [25].
Genes that are transcribed from a promoter that switches between an active and an inactive state can be regulated by changing the rates of switching between these two states, either by modulating the burst size (given by kINI/kOFF, where kINI = kLOAD kESC/ (kLOAD + kESC) is the average rate of initiation), or by modulating the burst frequency (kON), (it is also possible that both are modulated) [57,58]. In order to compute the predictions for the nascent RNA distribution for these two mechanisms of regulation in Fig 2C, we change burst size and burst frequency by changing kOFF and kON. In the first case, we change the burst size by changing kOFF and taking the other parameter values to be, kON = 0.435/min, k = 0.8kb/min, L = 4436 bps, kINI = 5/min as reported for PDR5[4]. Then we change burst frequency by changing kON, where the other parameters are, kOFF = 5/min, k = 0.8kb/min, L = 4436 bps, kINI = 5/min as reported for PDR5 [4]. In Fig 2C the Fano factor of the nascent RNA distribution is plotted as a function of its mean normalized by meanmax, where meanmax is the maximum of the mean number of nascent RNAs which is obtained when there is no transcriptional regulation and the promoter is always active.
We analyze the measured nascent RNA distributions for twelve different constitutively expressed yeast genes reported in reference [25]. By applying our theoretical results to the published data, we find that the average initiation rates of six (KAP104, TAF5, TAF6, TAF12, RPB2, RPB3) of these twelve genes are all roughly the same, and equal to 0.145±0.025/min. However the other six genes (RPB1, MDN1, PUP1, PRE7, PRE3, PRP8) initiate transcription at different rates. This we conclude from S2 Fig, where the mean number of nascent RNAs is plotted against the gene length for all the twelve genes.
When considering experiments that count nascent RNAPs it is important to be mindful of the fact that the number of RNAP molecules along a gene is not necessarily equal to the nascent RNA counts. Transcribing RNAPs have partial nascent transcripts attached to them depending on how far along the gene they have progressed (as indicated in Fig 1C). In a single molecule FISH experiment, the RNA sequence that is targeted by the fluorescent probes determines if these transcripts are detected or not. Probes against the 5’ end detect transcripts early on, while probes against the 3’ end will detect only almost finished transcripts. [9]. However, as long as there is a way to correctly extract the RNAP number distribution from nascent RNA intensity, our model can accurately transform this data into information about the transcriptional dynamics.
In addition to the mean, we analyze the Fano factor of the nascent mRNA distributions as well, and compare it with the prediction from our model of transcriptional regulation (Fig 2A). It is to be noted that for 9 of the different genes we consider, the number of nascent RNAs does not exceed 2. These distributions can be described by the probabilities of having 0,1 and 2 nascent RNAs. Still the Fano factor is a useful metric for analyzing these distributions and quantifying how much they differ from a Poisson distribution. We find the Fano factors of the nascent mRNA distribution for all of the twelve genes in the data set to be less than (or at most equal to) one. Hence the simplest model consistent with the published data for these twelve yeast genes is one where the promoter is always active and transcription initiation is a two-step process (see Fig 1A) parameterized by the kinetic rates kLOAD and kESC. We find that Fano factors for the six genes: KAP104, TAF5, TAF6, TAF12, RPB2, RPB3, which all initiate transcription at the same average rate, follow precisely the trend-line expected when kLOAD and kESC are equal (kLOAD = kESC = 0.29±0.013/min), as shown in Fig 2B.
We also analyze the Fano factor of the nascent mRNA distributions of the other six genes: RPB1, MDN1, PUP1, PRE7, PRE3, PRP8, which initiate transcription at a different mean rate (S2 Fig). Their location in the phase space defined by the gene length and Fano factor (as shown in S3 Fig) indicates that they all have at least two rate limiting steps leading up to initiation and that these steps are likely parameterized by different rates, unlike what we observe for the other six genes.
It is to be noted that it might be difficult to experimentally tell the difference between a mature RNA or a single nascent RNA, there are 12 different genes for which data is available [25] and our conclusions are based on examining the whole set. As pointed out earlier 9 of these genes have up to 2 nascent RNA molecules. Hence for these genes the distributions of nascent RNA molecules have three bins (for 0,1 and 2 mRNA molecules respectively). However our analysis also includes genes for which there are more than 1 or 2 nascent RNAs, such as RPB1 (up to 3), MDN1 (up to 5), PRP8 (up to 3). We find all of these genes to have a Fano factor of less than one indicative of two or more steps leading to initiation.
In order to compare the nascent RNA and cytoplasmic mRNA distributions, we compute the Fano factor of the cytoplasmic mRNA distribution, predicted by the two-step mechanism of initiation for the seven genes (KAP104, TAF5, TAF6, TAF12, RPB2, RPB3, MDN1). In other words an mRNA molecule is produced in two sequential steps, e.g., by first assembling the transcriptional machinery at the promoter DNA, followed by RNA polymerase escaping the promoter. These two steps are parameterized by the kinetic rates kLOAD and kESC, respectively. For the gene RPB1 initiation is a one-step process, while for others (PUP1, PRE3, PRE7, PRP8) three steps are required to account for the measured nascent RNA distribution. We further assume that mRNA is degraded with a constant probability γ per unit time per molecule. The degradation rates of the twelve genes used in the calculation are those reported in reference [25]. Given these assumptions about mRNA production and degradation, we compute the Fano factor (ratio of variance and mean) of the mRNA distribution using the approach developed previously in order to find the moments of mRNA distribution [8,38–41]. The computed Fano factor is a measure of the expected cell-to-cell variability in the number of cytoplasmic mRNAs if only the initiation process was contributing to this variability. The fact that we observe a discrepancy between the mRNA variability calculated in this way and the measured mRNA variability (Fig 4) is indicative of the presence of significant sources of noise that are downstream of transcription.
|
10.1371/journal.pgen.1001355 | The SUMO Isopeptidase Ulp2p Is Required to Prevent Recombination-Induced Chromosome Segregation Lethality following DNA Replication Stress | SUMO conjugation is a key regulator of the cellular response to DNA replication stress, acting in part to control recombination at stalled DNA replication forks. Here we examine recombination-related phenotypes in yeast mutants defective for the SUMO de-conjugating/chain-editing enzyme Ulp2p. We find that spontaneous recombination is elevated in ulp2 strains and that recombination DNA repair is essential for ulp2 survival. In contrast to other SUMO pathway mutants, however, the frequency of spontaneous chromosome rearrangements is markedly reduced in ulp2 strains, and some types of rearrangements arising through recombination can apparently not be tolerated. In investigating the basis for this, we find DNA repair foci do not disassemble in ulp2 cells during recovery from the replication fork-blocking drug methyl methanesulfonate (MMS), corresponding with an accumulation of X-shaped recombination intermediates. ulp2 cells satisfy the DNA damage checkpoint during MMS recovery and commit to chromosome segregation with similar kinetics to wild-type cells. However, sister chromatids fail to disjoin, resulting in abortive chromosome segregation and cell lethality. This chromosome segregation defect can be rescued by overproducing the anti-recombinase Srs2p, indicating that recombination plays an underlying causal role in blocking chromatid separation. Overall, our results are consistent with a role for Ulp2p in preventing the formation of DNA lesions that must be repaired through recombination. At the same time, Ulp2p is also required to either suppress or resolve recombination-induced attachments between sister chromatids. These opposing defects may synergize to greatly increase the toxicity of DNA replication stress.
| DNA damage, arising from environmental stress or errors in DNA metabolism, can interfere with DNA replication. Cells respond by using homologous recombination to bypass the damage, resulting in DNA strand linkages between the replicated chromosomes. It is crucial to undo these linkages so chromosomes can segregate properly. Previously, a regulatory mechanism known as SUMO modification was shown to be important in controlling recombination following replication interference by the DNA damaging agent MMS. We show that mutations in a yeast enzyme called Ulp2p, which reverses SUMO modification, increase recombination and impose a requirement for recombination to maintain survival. MMS–treated ulp2 mutants also accumulate recombination intermediates and fail to separate their chromosomes, leading to a permanent block to cell division. Further analysis suggests this block may not simply be due to a failure to resolve recombination intermediates, but may reflect a role for Ulp2p in undoing additional chromosome attachments that accompany recombination. In sum, our data indicate that cells defective for Ulp2p develop a love/hate relationship with recombination, requiring recombination for viability while failing to resolve chromosome attachments induced by recombination repair. Identification of Ulp2p substrates that ensure chromosome separation following recombination will shed light on how SUMO modification maintains genome stability.
| As part of the DNA damage response, homologous recombination (HR), particularly template switch recombination through the post-replication DNA repair pathway (PRR), provides an important mechanism for restarting stalled replication forks and filling in un-replicated gaps in DNA (reviewed in [1], [2]). These recombination events must be managed carefully, however. DNA strand exchange during HR, followed by re-initiating replication using the nascent sister chromatid as a template, can result in the formation of DNA linkages between daughter chromosomes. Failure to resolve these linkages, called sister chromatid junctions (SCJs), leads to chromosome breakage or aneuploidy, and may contribute to genome instability in many forms of cancer (reviewed in [3]).
A variety of studies implicate SUMO post-translational modification as an important regulator of HR in response to replication stress. Following activation of the SUMO precursor protein, SUMO modification is catalyzed by the E2 conjugating enzyme Ubc9p, which typically acts through one of several E3 ligases to covalently join SUMO moieties to lysine residues on substrate proteins (reviewed in [4]). One SUMO substrate that plays an especially prominent role in controlling HR at replication forks is Pol30p/PCNA, which is modified to recruit different activities to the replisome. During S phase, Ubc9p works through the E3 ligase Siz1p to sumoylate PCNA on K164 and K127 [5]. SUMO modified PCNA recruits the Srs2p helicase [6], [7], which suppresses unscheduled HR by disassembling Rad51p nucleoprotein filaments [8]-[10]. Following replication fork stalling at MMS-induced DNA lesions, however, PRR proteins catalyze either mono- or poly-ubiquitinylation of PCNA K164 [5]. These modifications recruit trans-lesion bypass polymerases or induce template switching HR, respectively, providing alternative mechanisms to bypass the lesion and restart replication [5], . The existence of additional SUMO substrates that control HR is suggested by the observations that mutations affecting both Ubc9p and the E3 ligase Mms21p, which is not required for PCNA sumoylation, confer sensitivity to the replication impeding drugs hydroxyurea (HU) and methyl methansulfonate (MMS) [5], [14]-[19]. Mms21p exists in a complex with two members of the structural maintenance of chromosomes family of proteins, Smc5p and Smc6p, which are also required for HU and MMS-resistance [15], [16], [20]-[22]. Notably, in response to MMS, ubc9, mms21, smc5 and smc6 mutants show an accumulation of X-shaped DNA structures that are thought to represent either regressed forks-a possible intermediate in fork restart-or hemi-catenate SCJs [17], [19], [23]. In this sense, they resemble mutants defective for the Sgs1p/Top3p/Rmi1p complex, which, through concerted helicase/topoisomerase activities, catalyzes the dissolution of hemi-catenates and other DNA linkages [24]-[27]. These findings suggest complex roles for sumoylation in either preventing excessive/improper HR at stalled replication forks and/or mediating the active dissolution of SCJs.
As with the forward SUMO pathway, SUMO de-conjugation is also required to tolerate replication stress. Budding yeast contains two members of the SENP/Ulp family of SUMO isopeptidases, Ulp1p and Ulp2p, which catalyze removal of SUMO [28]-[30]. Ulp1p is an essential enzyme that is preferentially localized to the nuclear pore [28]-[32], whereas Ulp2p is distributed throughout the nucleus [29], [30], [33]. Ulp2p (first identified as Smt4p; [34]) is not essential, but ulp2 mutants grow poorly and exhibit a complex assortment of phenotypes, including chromosome segregation and cell division defects. [29], [30], [35]-[42]. Ulp1p and Ulp2p also mediate functions that promote SUMO modification. Ulp1p is required to cleave the SUMO precursor to expose a glycine residue necessary for conjugation [28], while Ulp2p possesses a chain editing activity that prevents formation of aberrantly poly-sumoylated substrates [43]. Poly-sumoylation has the potential to interfere with the functional role of SUMO addition. Moreover, recent evidence has revealed that some poly-sumoylated substrates are targeted for degradation by the SUMO-targeted Slx5p-Slx8p ubiquitin ligase [44]-[46].
Although Ulp1p and Ulp2p play different roles in the SUMO pathway, one trait shared by ulp1 and ulp2 strains is that both exhibit sensitivity to HU and MMS [29], [30], [43]. Previously, a ulp1-I615N mutant was shown to accumulate single-stranded gaps during DNA replication, to exhibit increased spontaneous recombination, and to become dependent on Srs2p and HR for viability, suggesting a role for Ulp1p in suppressing replication errors that induce HR [47]. Insight into the replication stress sensitivity of ulp2 mutants has come from the important finding that Ulp2p is required for cells to complete mitosis following DNA damage checkpoint arrest [41]. From this, de-sumoylation of Ulp2p substrates may be necessary to restart the chromosome segregation machinery once the checkpoint block to mitosis has been relieved [41], [48]. But whether Ulp2p, like other components of the SUMO pathway, is also involved in controlling HR during DNA damage or replication stress has not yet been examined. In this study, we find that, following replication fork stalling by MMS, ulp2 mutants accumulate persistent recombination intermediates that are likely to correspond to SCJs. This mis-regulation is accompanied by a severe, recombination-dependent, block to chromosome segregation, revealing a critical role for Ulp2p in allowing sister chromatids to disjoin following HR DNA repair.
We initially set out to determine if ulp2 mutants displayed a similar dependency on recombination as ulp1-I615N strains [47]. A ulp2 deletion mutant (ulp2Δ) was mated to rad52Δ, rad51Δ and rad6Δ strains. Rad51p and Rad52p are required for most forms of HR [2], while Rad6p controls trans-lesion synthesis and template switching PRR [1]. ulp2Δ rad52Δ, ulp2Δ rad51Δ and ulp2Δ rad6Δ double mutants were either not obtained or were obtained at lower than expected frequencies from these crosses (Table 1, Table 2, Table 3). For rad52Δ, we examined this apparent synthetic lethality further by isolating ulp2Δ rad52Δ segregants harboring a wild type (WT) copy of RAD52 on a URA3 minichromosome (pRAD52). ulp2Δ rad52Δ/pRAD52 mutants grew weakly, if at all, on media containing 5-FOA, a drug that only allows growth if cells are capable of losing pRAD52 (Figure 1A). Thus, Rad52p is essential for proliferation of ulp2Δ cells.
The essential role of Rad52p prompted us to examine whether HR was elevated in the absence of Ulp2p. Yeast cells exhibit a uniform nuclear distribution of fluorescent Rad52p-GFP in the absence of DNA damage (Figure 1B, [49]), but Rad52p-GFP rapidly assembles into intra-nuclear foci during HR DNA repair [49]. We found that an average of 17% of ulp2Δ cells in mid-logarithmic phase cultures displayed Rad52p-GFP foci, a significant increase (p = 0.0074) compared to less than 1% in WT cells. (Figure 1B). As a second assay, we utilized a reporter in which recombination events between direct repeats on chromosome XV can be selected because they restore an intact HIS3 gene [50]. ulp2Δ cells exhibited a 4.7-fold increase in the median frequency of this form of recombination (Figure 1C; p = 0.044), indicating spontaneous HR at this genomic locus is significantly increased in ulp2Δ mutants.
DNA replication errors are potent inducers of HR and can initiate chromosome rearrangements [51], [52]. Based on this, we used a yeast artificial chromosome (YAC) assay to examine the frequency of spontaneous gross chromosomal rearrangements (GCRs) in ulp2Δ cells ([53]; Figure 2A). For comparison, we also measured GCR frequencies in ulp1-333, smt3-331 and ubc9-1 SUMO pathway mutants (SMT3 encodes the single SUMO isoform in budding yeast). Using the YAC system, we obtained median GCR frequencies of 252×10−7 for WT cells, 5490×10−7 for smt3-331 cells, 6109×10−7 for ubc9-1 cells, and 2617×10−7 for ulp1-333 cells (Figure 2B), representing 22-, 24-, and 10-fold increases, respectively, compared to WT controls. In contrast, and counter to initial expectations, it proved difficult to recover spontaneous GCRs in ulp2Δ mutants, with a median GCR frequency of 56×10−7 (Figure 2C). This represents a significant (p = 0.025) 4.5-fold decrease compared to WT.
To further monitor chromosome rearrangements we examined two circular dicentric minichromosomes. In one (p2XCENdirect), two copies of a CEN sequence were oriented in a direct repeat configuration. In the other (p2XCENinvert), the same CEN duplication was oriented as inverted repeats. Previous studies have shown that both direct and inverted repeat dicentrics can be efficiently transformed into yeast, and are initially retained through a combination of co-orientation of the two CENs on the spindle and non-disjunction following dicentric bridging [54], [55]. During outgrowth, however, rearranged minichromosomes that have deleted one of the CENs accumulate. For direct CEN repeats these deletions tend to arise through loop out events, whereas inverted CEN repeats are resolved through more complex re-arrangements. Consistent with this characterization, in WT transformants p2XCENdirect and p2XCENinvert exhibited similar mitotic stabilities to p1XCEN controls (Figure 2D). Analysis of minichromosomes rescued from these cells revealed precise CEN1 excision for p2XCENdirect and a diversity of plasmid species for p2XCENinvert (not shown). In ulp2Δ mutants, p1XCEN was only retained in ∼30% of the cells; this result is in keeping with previous studies showing reduced minichromosome stability in the absence of Ulp2p [29]. p2XCENdirect demonstrated a similar stability to p1XCEN (Figure 2D), and underwent the same precise CEN deletions observed in WT (not shown). In contrast, p2XCENinvert proved extremely unstable, with less than 1% of ulp2Δ cells maintaining the mini-chromosome. These results suggest that some chromosome re-arrangements either fail to occur or cannot be tolerated in ulp2 mutants.
In order to more directly examine the consequences of HR in ulp2Δ mutants, we used MMS to induce recombination. As an initial experiment, we examined chromosome integrity following exposure to MMS by pulse-field gel electrophoresis. WT and ulp2Δ cells were arrested in G1, released into media containing 0.01% MMS for 2 hr, and then allowed to recover in MMS-free media. Following MMS treatment a lower molecular weight DNA smear was observed in both WT and ulp2Δ strains (Figure 3A), reflecting MMS-induced chromosome breakage [17]. For both strains, a one hr recovery largely restored the normal chromosome banding pattern. This suggests Ulp2p is not obviously required for healing MMS-induced DNA breaks.
We next examined processing of MMS-induced DNA lesions. In the experiment shown in Figure 3B, WT cells and ulp2Δ mutants expressing RAD52-GFP were treated with 0.01% MMS and allowed to recover. After a 2 hr recovery, ∼30% of WT cells accumulated Rad52p-GFP foci (Figure 3B). By 6 hr, however, the percentage of cells with Rad52p-GPF foci had substantially declined and many cells were proceeding with the next round of cell division. In contrast, ulp2Δ mutants showed a much stronger accumulation of Rad52p-GFP foci, reaching a maximum of ∼60% (Figure 3B), and these foci tended to persist for the duration of the recovery period. We also examined Rad52p-GFP foci in ulp2Δ cells treated with 200 mM HU. HU does not normally induce Rad52p foci because the integrity of the replisome is maintained by the S phase checkpoint (Figure 3B, [56]). HU treated ulp2Δ cells, however, exhibited a strong induction of Rad52p-GFP foci.
In response to MMS, proper regulation of HR is required to prevent X-shaped recombination intermediates from accumulating in the vicinity of origins of replication [17], [19], [23]. On two-dimensional gels these structures migrate as a “X-spike” that is distinct from replication forks and bubbles [57], [58]. To determine whether ulp2Δ mutants accumulated this type of HR intermediate, ulp2Δ cells, along with WT and sgs1Δ controls, were released from a G2/M nocodazole block and treated with 0.033% MMS for 3 hr as previously described [17]. Genomic DNAs were fractionated on two-dimensional gels, and probed with a DNA fragment corresponding to ARS305. A prominent X-spike signal was observed in sgs1Δ and ulp2Δ samples (Figure 3D). Thus, Ulp2p deconjugating and/or chain editing activities are required to prevent accumulation of MMS-induced HR intermediates.
Based on current evidence, Sgs1p is one SUMO target that could be connected to Ulp2p's role in HR. In particular, a recent study has shown that a single prominent SUMO species of Sgs1p accumulates after MMS exposure, and K621 has been identified as the acceptor lysine that is responsible for this modification [59]. We were able to confirm that treatment with 0.3% MMS resulted in a substantial fraction of Sgs1p-myc shifting into a reduced mobility species (Figure 4A and Figure S1), and that a decreased amount of this form was observed following treatment with a lower MMS concentration (0.033%; Figure 4B, 4C). The appearance of this form was abolished in ubc9-1 strains (Figure 4B) and a sgs1-K621R mutant (Figure S1), indicating it is likely to correspond to the previously reported K621 conjugate. In ulp2Δ strains, however, a marked increase in this putative Sgs1p SUMO species was observed (Figure 4B, 4C), which persisted for at least 3 hr after removal of MMS (Figure 4C). In sum, these results suggest that sumoylation of Sgs1p is likely to be regulated by Ulp2p.
If failure to properly control Sgs1p sumoylation was responsible for ulp2Δ HR defects, SUMO-resistant Sgs1p might ameliorate these phenotypes. We therefore examined whether a plasmid-born copy of the sgs1-K621R allele could prevent Rad52p foci accumulation. Following a two hr treatment with 0.010% MMS, however, no significant reduction in ulp2Δ sgs1-K621R cells displaying Rad52p-GFP foci was observed (Figure 4D). Previous studies have shown that a form of Smt3 (smt3-3KR) that cannot form polymeric SUMO chains can rescue the HU and MMS sensitivity of ulp2 mutants [43], leading us to test whether smt3-3KR could prevent Rad52p foci accumulation. This proved to be the case, as smt3-3KR ulp2Δ double mutants did in fact show a substantial reduction in the accumulation of both spontaneous and MMS-induced Rad52p foci (Figure 4F). Thus, proper SUMO chain editing through Ulp2p is likely to be important in controlling HR.
In our experiments, it was apparent that ulp2Δ cells frequently remained blocked in the cell cycle during recovery from MMS, similar to previous results examining ulp2 recovery following HU treatment and in response to an irreparable DNA double strand break [41]. We took four experimental approaches to investigate the basis for the apparent MMS recovery defect of ulp2Δ cells. First, phospho-activation of the Rad53p checkpoint kinase during the DNA damage response results in a series of slower migrating gel mobility variants [60], and collapse of these forms provides a means to assess silencing of the checkpoint. In WT cells, Rad53p phospho-variants almost completely disappeared during a 2–4 hr recovery after treatment with 0.01% MMS (Figure 5A). A similar pattern was observed in ulp2Δ strains, although the accumulation and disappearance of shifted Rad53p appeared to be slightly delayed.
Second, we examined degradation of Pds1p/securin. Pds1p is a downstream target of the DNA damage checkpoint that is stabilized to block cohesin proteolysis and anaphase entry [61], [62]. The kinetics of Pds1p degradation therefore provides a read-out of commitment to anaphase. In these experiments, we used the cdc14-1 allele to block Pds1p re-synthesis once cells recovered from the checkpoint. cdc14-1 PDS1-myc and cdc14-1 ulp2Δ PDS1-myc cells were treated with 0.001%, 0.005% and 0.01% MMS for 2 hr, allowed to recover at a cdc14-1 non-permissive temperature, and Pds1p-myc abundance was monitored over a 24 hr period. In cdc14-1 cells, Pds1p-myc degradation proceeded in a dose-dependent manner until 10 hr post-treatment (Figure 5B, 5C). At this point, Pds1p started to increase in the 0.001% and 0.005% MMS cultures, probably reflecting leakage through the cdc14-1 arrest. These degradation kinetics were virtually indistinguishable in cdc14-1 ulp2Δ cells, although re-synthesis of Pds1p was not observed (Figure 5B, 5C). These results suggest that MMS treated ulp2Δ cells can terminate checkpoint signaling and commit to anaphase.
Third, we used micro-colony analysis to determine whether getting rid of the checkpoint relieved the restraint on cell division. Cells from MMS treated and untreated cultures were positioned on agar plates, and the appearance of cell bodies was examined over time. A budded yeast cell arrested at the DNA damage checkpoint consists of two cell bodies. If this cell completes mitosis and one of progeny cells sends forth a bud, the microcolony now contains three cell bodies, and the number of cell bodies increases exponentially with continued division. We found that an average of 68% of WT cells were able to form microcolonies containing ≥ 16 cell bodies within a 3 day period after transient exposure to MMS, indicating the majority recovered efficiently (Figure 6). In comparison, even in the absence of MMS, 20% of ulp2Δ cells remained blocked at the 2–3 cell body stage. This lethality was strongly exacerbated by MMS treatment, with 64% of ulp2Δ cells failing to proliferate beyond 2–3 cell bodies. Inactivating the DNA damage checkpoint in rad9Δ ulp2Δ mutants, or both the DNA damage and S phase checkpoints in mec1Δ ulp2Δ mutants, did not relieve the ulp2Δ block to cell division (Figure 6). ulp2 cells fail to maintain chromatid cohesion at centromeric regions during DNA damage checkpoint arrest [36], [42], which could potentially activate the spindle assembly checkpoint (SAC). We therefore tested whether abolishing the SAC could restore ulp2Δ division. However, ∼60% of MMS treated ulp2Δ mad2Δ mutants still remained blocked with 2–3 cell bodies (Figure S2). We further generated a ulp2Δ rad9Δ mad2Δ triple mutant to abolish both DNA damage and SAC checkpoint responses. This triple mutant grew extremely poorly, and, following exposure to MMS, ∼90% of the cells failed to recover (Figure 6). Thus, MMS treated ulp2Δ mutants experience a terminal block to cell division even in the absence of pre-anaphase checkpoint controls.
Fourth, we examined mitotic progression in ulp2Δ cells by cytology and flow cytometry. Following a 2 hr treatment with 0.01% MMS, WT cells arrested at the DNA damage checkpoint typically showed short pre-anaphase spindles and an undivided mass of chromatin (Figure 7A, 7B). Completion of mitosis during recovery was characterized by normal spindle extension and chromosome transmission. As monitored by DAPI staining and a Lac operator-GFP chromosome tag (TRP1-GFP), ∼70% of cdc14-1 cells underwent chromosome separation and segregation during recovery (Figure 8A, 8B), and FACS analysis indicated that many cells proceeded with additional rounds of cell division (Figure S3). In contrast, many MMS treated ulp2Δ cells showed partial, incomplete spindle extension during recovery, accompanied by an apparent block to nuclear division (Figure 7A, 7B). In some cells it was possible to visualize chromatin fibers that appeared to be pulled away from an undivided mass of chromatin (Figure 7B iii; arrows). In others, chromosome separation appeared more complete, but chromatin was stretched to varying degrees along the spindle (Figure 7B iv). DAPI staining indicated less than 20% of cdc14-1 ulp2Δ cells successfully completed chromosome segregation (Figure 8A). ∼30% of cdc14-1 ulp2Δ cells underwent TRP1-GFP separation during recovery, but the separated foci largely failed to segregate (Figure 8B). FACS analysis suggested that MMS treated ulp2Δ cells potentially tried to proceed with a second round of DNA replication following this block chromosome segregation, although the FACS profiles were quite heterogeneous and did not clearly resolve into a peak of cells with a 4N content of DNA (Figure S3).
Since sgs1Δ and ulp2Δ mutants both accumulate HR intermediates that might be expected to link sister chromatids (Figure 3C), we additionally examined chromosome segregation during MMS recovery in cdc14-1 sgs1Δ cells. Compared to the ulp2Δ defect, the fraction of MMS treated cdc14-1 sgs1Δ cells that could segregate their chromosomes to an extent necessary to form two distinct nuclear masses was only slightly reduced compared to cdc14-1 controls (Figure 8A; see Figure S4 for a more complete description). Taken as a whole, these results allow us to conclude that, although they commit to anaphase, ulp2Δ mutants are unable to separate their chromosomes efficiently following MMS treatment. Furthermore, this non-disjunction defect appears more severe than that observed in a sgs1Δ strain.
If defective HR in MMS treated ulp2Δ cells is causally linked to the chromosome separation defect that we observed in our experiments, blocking recombination should restore chromosome segregation. Given that HR is essential in ulp2Δ mutants (Figure 1) our approach to test this was to overproduce (OP) the Srs2p helicase. In addition to antagonizing nucleoprotein filament assembly [8]-[10], Srs2p also appears to exert anti-recobinogenic activity by unwinding D-loop intermediates [63], [64]. Srs2p OP should therefore be an effective way to short circuit early stages of HR. cdc14-1, cdc14-1 rad9Δ, cdc14-1 ulp2Δ and cdc14-1 rad9Δ ulp2Δ strains were transformed with a vector control or a high copy plasmid in which SRS2 was expressed under control of its endogenous promoter (pSRS2). The transformants were then treated with 0.01% MMS for 2 hr and allowed to recover at a cdc14-1 non-permissive temperature. Compared to vector controls, cdc14-1/pSRS2 cells remained blocked in a pre-anaphase configuration for the duration of the recovery period (Figure 9A). This delay was abolished in cdc14-1 rad9Δ/pSRS2 transformants, suggesting Srs2p OP was able to prolong DNA damage checkpoint arrest. In the absence of Ulp2p, however, inactivating the checkpoint in the cdc14-1 rad9Δ ulp2Δ/vector strain was insufficient to allow cells to proceed with chromosome segregation (Figure 9A, 9B). Significantly, Srs2p OP demonstrated a remarkable ability to allow ulp2Δ strains to escape this mitotic block, with ∼50% of cdc14-1 rad9Δ ulp2Δ/pSRS2 cells now segregating their chromosomes in a seemingly normal anaphase (Figure 9A, 9B). Thus, Srs2p OP substantially relieves the block to chromosome separation in MMS treated ulp2Δ cells.
One principal finding of this study is that, even in the absence of exogenous DNA replication stress, spontaneous recombination is increased in ulp2Δ cells. This conclusion is based on two observations. First, by genetic criteria, spontaneous recombination at a genomic location on chromosome XV is elevated in ulp2Δ strains. Second, ulp2Δ mutants also display an increase in the frequency of spontaneous Rad52p DNA repair foci. A similar increase in Rad52p foci has been observed in a number of other SUMO pathway mutants, and has been shown to be largely attributable to a requirement for sumoylation in preventing inappropriate recombination events involving the 2 µm circle, an endogenous plasmid found in most S. cerevisiae strains [65]. Since we have not directly examined the effect of the 2 µm circle on recombination in ulp2 mutants, destabilization of this extrachromosomal element may well contribute to the ulp2Δ increase in Rad52p foci. However, as the 2 µm circle is not required for S. cerevisiae growth, our finding that HR DNA repair becomes essential in ulp2Δ strains strongly suggests that Ulp2p acts to suppress the formation of genomic DNA lesions that must be repaired through recombination. Previous analyses of the SUMO pathway support this possibility. For example, SUMO conjugation-defective ubc9-1 mutants exhibit synthetic growth defects in the absence of HR and, at the non-permissive temperature, accumulate DNA structures that activate Rad53p [17]. Furthermore, as described in the Introduction, ulp1-I615N mutants also show increased HR and require HR for viability; in this case, the requirement for HR was shown to correspond with single-stranded DNA gaps arising during S phase [47]. It is striking that perturbations to Ulp1p and Ulp2p, which appear to target largely distinct sets of SUMO substrates [29], impose such seemingly similar dependencies on HR. Another observation that lends credence to the idea that Ulp2p suppresses recombinogenic DNA lesions is that ulp2Δ mutants greatly induce the formation of Rad52p foci following HU treatment. Such foci are not observed in HU treated WT cells [56], consistent with an underlying replication problem in ulp2Δ mutants that is exacerbated by slowed fork progression.
In analyzing genome stability in ulp2Δ strains, we observed two interesting differences compared to other SUMO pathway mutants. First, whereas our data indicate that Rad6p-dependent PRR is essential in ulp2 mutants, mis-regulation of SUMO conjugation in ulp1-I165N rad18 [47], ubc9-1 rad18 [19], siz1 rad18 [11], pol30-K164R rad18 and pol30-K164R rad6 [5] mutants can actually compensate for defective PRR. One scenario that might account for this difference is if poly-sumoylation of a Ulp2p substrate(s) caused a distinct perturbation to DNA replication that was repaired through PRR-mediated HR. In keeping with this interpretation, we find that blocking poly-SUMO chain formation reduces the accumulation of both spontaneous and MMS-induced HR foci in ulp2Δ mutants.
A second apparent difference concerns the formation of GCRs. In contrast to smt3-331, ubc9-1 and ulp1-333 strains, where spontaneous GCRs are increased, ulp2Δ mutants show reduced GCRs. Formally, Ulp2p could promote GCR formation by stimulating error prone DNA repair. There is precedence for this, as a previous study found that, in the absence of template switch PRR, Siz1p-mediated sumoylation of PCNA was required to form certain types of GCRs [66]. Alternatively, Ulp2p could be required for cells that would give rise to GCRs to recover and propagate efficiently. Our observations with dicentric minichromosomes are consistent with the idea that repair events leading to some GCRs may not be tolerated in ulp2Δ strains. We were able to recover re-arranged dicentrics from ulp2Δ mutants when duplicated CEN sequences were present in a direct repeat configuration. Such deletions can occur through single-strand annealing, an intra-chromosomal form of recombination [67]. In contrast, CEN deletion GCRs were not recovered when the two CENs were oriented as inverted repeats. Recent studies have shown that faulty template switch PRR is frequently involved in initiating deletions between inverted repeats [68], [69]. As discussed below, one possibility is that such recombination events are accompanied by formation of SCJs or other types of chromatid attachments that fail to be resolved in ulp2 cells.
Our results led us to suspect that HR DNA repair, while required for viability, might at the same time be toxic to ulp2 cells, prompting us to examine processing of MMS-induced recombination events. From this analysis, one conclusion is that, similar to Ubc9p, Mms21p, Smc5p/Smc6p, and Sgs1p/Top3p [17], [19], Ulp2p is required to prevent X-shaped DNA structures from accumulating at sites of replication fork stalling/collapse. We also find that, whereas Rad52p foci disappear during MMS recovery in WT cells, the incidence of these foci remains elevated in ulp2Δ strains, suggesting a possible role for Ulp2p in terminating recombination. Determining the molecular basis for how Ulp2p prevents accumulation of HR intermediates, and whether this function is related to or separate from Ulp2p's role in Rad52p foci disassembly, are important future questions.
Based on current information, Ulp2p could be connected to HR through a number of different SUMO substrates. First, Mms21p-mediated sumoylation of unknown substrates, probably in conjunction with Smc5p/Smc6p [22], [70], has been proposed to prevent excessive template switch recombination through PRR [19]. Alternatively, more recent evidence suggests Smc5p/Smc6p may instead act downstream of PRR to facilitate the dissolution of HR intermediates [71]. Second, Sgs1p is sumoylated under conditions when it is active in SCJ dissolution [17], [59], although apparently through an Mms21-independent pathway [17]. Third, Ubc9p/Siz1p-controlled sumoylation of PCNA and recruitment of Srs2p may suppress PRR-independent recombination at replication forks [6], [7], [19]. Fourth, Srs2p has also been shown to be sumoylated, with poly-sumoylation being proposed to trigger Srs2p degradation through the Slx5p/Slx8p pathway [72]. Fifth, a fraction of Rad52p [73]-[75], and other HR proteins [76], are sumoylated in response to MMS, which may be involved in fine-tuning processing of broken DNA. Finally, a growing number of protein-protein interactions within HR foci have been found to be controlled by sumoylation (reviewed in [77]). As part of completion of repair, Ulp2p may catalyze the disassembly of these networks.
As a first step in placing Ulp2p in these pathways, we tested whether mis-regulation of Sgs1p sumoylation was connected to ulp2Δ HR defects. Overproduction of Ulp2p was recently shown to block Sgs1p sumoylation on K621 following MMS treatment [59], and, as we report here, MMS-induced sumoylation of Sgs1p is elevated in the absence of Ulp2p. It is therefore likely that Ulp2p acts as the SUMO deconjugating enzyme for Sgs1p. Despite this, short-circuiting Sgs1p sumoylation using the sgs1-K621R mutation did not reduce Rad52p foci accumulation in ulp2Δ cells, indicating mis-regulation of other Ulp2p substrates is likely to be involved in modulating HR.
The failure of ulp2 mutants to resume cell division following DNA damage is one of the most intriguing aspects of the ulp2 phenotype. The first study to document this phenomenon showed that, following adaptation to a persistent DNA break, only a fraction of ulp2 cells were able to proceed with nuclear division, frequently accompanied by abnormally extended or broken mitotic spindles [41]. Inactivating the DNA damage checkpoint rescued this defect, suggesting a critical role for Ulp2p in re-initiating chromosome segregation following completion of the checkpoint response [41], [48].
While our results are largely in accord with this study, we observed a potentially informative difference in the role of the checkpoint in manifesting the ulp2Δ recovery defect. During MMS recovery, ulp2Δ cells dephosphorylated Rad53p and degraded Pds1p on schedule, suggesting they were competent to silence the checkpoint and initiate anaphase. Despite this, sister chromatids failed to disjoin, resulting in a dramatic failure in chromosome segregation. OP of Srs2p, which antagonizes HR [8]-[10], [63], [64], was able to largely restore chromosome segregation. In addition to modulating nucleo-protein filament assembly, Srs2p has also been shown to be required for full activation of the DNA damage checkpoint and for recovery from DNA damage checkpoint arrest [78], [79]. In our experiments, we observed that Srs2p OP greatly extended DNA damage checkpoint arrest in MMS treated WT cells. Based on the above considerations, this extended arrest could presumably reflect either mis-regulation of the checkpoint pathway, or, by interfering with HR DNA repair, elevated Srs2p could simply prolong normal checkpoint signaling. While the effects of Srs2 OP on checkpoint signaling and HR may be multi-faceted, the key point we wish to emphasize here is that abolishing the DNA damage checkpoint (or the SAC) did not allow ulp2Δ cells to divide more times during recovery from MMS treatment. Furthermore, preventing DNA damage checkpoint arrest in MMS treated ulp2Δ rad9Δ cells was insufficient to relieve the block to chromosome separation; OP of Srs2 was also necessary. In sum, these findings strongly suggest that, following replication fork stalling by MMS, downstream events initiated through HR, rather than checkpoint arrest per se, appear to play a causal role in interfering with chromosome segregation.
A key question concerns how HR could have this effect. Perhaps the simplest idea is that unresolved SCJs block chromatid disjunction. Whether this is a sufficient explanation, however, is unclear. First, in the experiments examining ulp2 adaptation to a persistent, endonuclease-targeted DNA break, both chromatids would be expected to be cut, preventing HR strand exchange [41]. Thus, the only way in which DNA linkages could form between chromosomes in these cells would be if extensive resection during prolonged checkpoint arrest triggered illegitimate recombination events. Second, we show that MMS treated sgs1Δ mutants, which are clearly defective in the dissolution of SCJs [17], [25], [27], do not show as severe a block to chromosome separation as Ulp2p-deficient cells. This is consistent with a recent study that showed, from among a collection of helicase-, nuclease-, and topoisomerase-deficient mutants, only smc5, smc6 and mms21 strains showed chromosome segregation defects after a pulse of MMS delivered in G1 [71]. This suggests a role for Mms21p-mediated sumoylation and the Smc5p/Smc6p complex in resolving SCJs or other types of chromatid linkages outside the Sgs1p/Top3p pathway [71]. Along these lines, it is notable that Ulp2p has been implicated in multiple facets of chromatid separation, including controlling sumoylation of cohesin regulatory proteins [37], [42], condensin [35], [38], and DNA topoisomerase II [36], [40]. Speculatively, following induction of HR, there may be an increased requirement for Ulp2p in the vicinity of DNA lesions, not only to prevent accumulation of HR intermediates, but also to complete replication, to disentangle DNA or to release protein-based forms of cohesion. Given the dramatic way in which the absence of Ulp2p potentiates the ability of replication toxins to block cell proliferation, a further understanding of the ulp2 recovery defect could lead to insights that are relevant to cancer treatment.
All S. cerevisiae strains used in this study were derived from the W303-related CRY1 strain and are listed in Table S1. A description of how different genetic elements were introduced into the CRY1 background can be found in Text S1. For all experiments, cells were cultured in standard formulations of yeast extract/peptone/dextrose (YPD) and synthetic complete minimal (SC) media. For G1 synchronization, alpha factor (Bio-Synthesis Corp.) was used at 10 µg/ml. For arresting cells in G2/M, nocodazole (Sigma-Aldrich) was used at 15 µg/ml in YPD. MMS and HU were purchased from Sigma-Aldrich. 5-FOA was purchased from Biovectra/Fisher and used at 1 mg/ml. G418 was purchased from Mediatech/Fisher and used at 200 µg/ml in YPD.
pLAY202 ([50]; provided by A. Bailis, City of Hope National Medical Center, Duarte, CA) was linearized with BstXI and targeted to the HIS3 locus, placing a URA3 marker between partially duplicated HIS3 sequences. pLAY202 integrants were propagated in Ura−/SC media, and, following overnight incubation, cell density was quantified using a hemacytometer. Viable cell counts were determined by plating a defined number of cells onto YPD and counting the resulting colonies. Recombination events were selected by plating a larger number of cells onto His−/SC media, and replica plating colonies that arose onto 5-FOA. Colonies that reverted to a His+, Ura− phenotype were scored as recombinants.
YAC yWss1572-1 ([53]; provided by D. Koshland, Univ. of California at Berkeley, Berkeley, CA) was modified so that the TRP1 marker on the left arm of the YAC was replaced with kanMX. This was performed by PCR amplifying a trp1Δ::kanMX disruption cassette using the following primers
5′-GCATATAAAAATAGTTCAGGCACTCCGAAATACTTGGTTGGCGTGTTTC
GTCAGCTGAAGCTTCGTACGC (CO354)
5′-TCTGGCGTCAGTCCACCAGCTAACATAAAATGTAAGCTTTCGGGGCGCAT
AGGCCACTAGTGGATCTG (CO355)
and pFA6a/kanMX2 [80] as template. G418Res, Trp− transformants were analyzed by PCR to verify correct targeting. The resulting YAC, named yWss1572Δtrp1, was subsequently transferred between strains using cytotransduction [81] or standard genetic crosses. To isolate GCRs, strains containing yWss1572Δtrp1 were grow in Ura−/SC media at 30°C for WT, ulp2Δ, ubc9-1 and smt3-331 strains, and 34°C for ulp1-333 mutants; these represent semi-permissive temperatures for the ubc9-1, smt3-331 and ulp1-333 alleles. Cell densities were quantified using a hemacytometer, and dilutions of the cultures were plated onto YPD to monitor plating efficiency. Aliquots of 105, 106, 107 and 108 cells were plated on 5-FOA to select for loss of the URA3 marker on the YAC. Colonies arising on 5-FOA were replica plated to YPD/G418 and Ade−/SC media. Clones growing on 5-FOA and YPD/G418, but not on Ade−/SC (G418Res, 5-FOASen, Ade−) were considered to arise from GCRs deleting the right arm of the YAC. In contrast, clones that were able to grow on 5-FOA, but could not grow on YPD/G418 or Ade−/SC (G418Sen, 5-FOASen, Ade−) were considered to arise through YAC mis-segregation events. For each culture, the total number of GCR clones arising on all the assay plates was used to calculate GCR frequency.
To monitor the mitotic stability of dicentric minichromosomes, p2XCENdirect (pJBN152; a YRp14-derived minichromosome containing two copies of a 1.7 kb CEN1 DNA fragment in a direct repeat configuration, see Text S1) and p2XCENinvert (pJBN151; similar to pJBN152 but with the CEN1 duplication oriented as an inverted repeat) were transformed into WT and ulp2Δ strains and compared to p1XCEN (YRp14/CEN1) controls. Transformants were inoculated into parallel YPD and Ura−/SC cultures and incubated at 30°C. After ∼15 hr of outgrowth, appropriate dilutions were plated onto YPD and Ura−/SC media. Mitotic stability was calculated by dividing the number of Ura+ colonies by the total number of colonies obtained on YPD.
Cultures for microscopy were supplemented with 50 µg/ml adenine to quench auto-fluorescence. To visualize Rad52p-GFP and TRP1-GFP, cells were fixed in 1% formaldehyde for 1.5 min and washed into PBS. DAPI staining was performed using Vecta-Shield (Vector Laboratories) containing 10 µg/ml DAPI. TUB1-GFP and HHF2-YFP strains were visualized as live mounts. HHF2-YFP is typically propagated as a heterogyzous diploid (HHF2-YFP-HIS3/+) to minimize selective pressure for rearranged variants that lose the fluorescent marker. However, in order to compare the response of HHF2-YFP strains to MMS concentrations similar to those used in our other recovery experiments, we chose to examine HHF2-YFP haploid segregants that were generated on an experiment-by-experiment basis. This proved to allow propagation of haploid strains with robust Hhf2-YFP fluorescence. In all cases, cells were visualized on Nikon E-800 or Nikon Eclipse 80i microscopes equipped with florescence optics and 100X (1.4 NA) or 60X (1.4 NA) objectives. Rad52p-GFP foci were typically scored using a number 4 neutral density filter to minimize photobleaching. A Zeiss Axioskop 40 microscope equipped with a 25 µm diameter optical fiber dissection needle was used to micromanipulate yeast cells for microcolony analysis. FACS analysis was performed by staining ethanol fixed yeast cells with propidium iodide as previously described [82].
10 ml aliquots of OD600 0.8 cultures were harvested by centrifugation and concentrated into 400 µl cell suspension buffer (10 mM Tris, 20 mM NaCl, 50 mM EDTA, pH 7.2). The cell suspension was warmed to 55°C and mixed with 400 µl 2% low melting temperature agarose (SeaKem) dissolved in TBE gel electrophoresis buffer (kept molten at 55°C) containing lyticase (Sigma L4025; final concentration 1 mg/ml). The cell suspension was transferred into molds and allowed to solidify to form plugs (4°C, 15 min). Plugs were pushed out into 50 ml conical tubes and incubated with 5 ml 1 mg/ml lyticase dissolved in 10 mM Tris, 50 mM EDTA, pH 7.2 for one hr at 37°C, followed by treatment 1 mg/ml Proteinase K (Sigma) dissolved in 100 mM EDTA, 0.2% Na Deoxycholate, 1% Na lauryl sarcosine, pH 8.0 at 50°C overnight. Plugs were washed (20 mM Tris, 50 mM EDTA, pH 8.0) 4 times one hour each and stored in wash buffer. Prior to electrophoresis, plugs were placed on a glass plate and trimmed to fit electrophoresis wells. Samples were then fractionated on 1% agarose gels in TBE using a Bio-Rad CHEF-DR II pulsed field electrophoresis system at 6V/cm for 22 hrs with a switch ramp time ramped from 50 to 90 sec at 14°C. Gels were stained with ethidium bromide (0.5 µg/ml, 15 min) prior to photography.
Genomic DNA preparations and two-dimensional gel electrophoresis were performed according to detailed online methods available from the Brewer-Raghuraman laboratory:
(http://fangman-brewer.genetics.washington.edu/DNA_prep.html)
(http://fangman-brewer.genetics.washington.edu/2Dgel.html)
In brief, cells were grown in 500 ml YPD until the cultures reached an OD600 of 0.6. The cultures were synchronized in nocodazole for 2 hr, washed, and released into fresh YPD containing 0.033% MMS. After a 3 hr treatment, cells were harvested by centrifugation and stored in 5 ml of NIB buffer (17% glycerol, 50 mM MPOS free acid, 150 mM potassium acetate, 2 mM magnesium chloride, 150 µM spermine and 500 µM spermidine, pH 7.2). Cells were lysed by bead beating in NIB buffer, and genomic DNA was purified on cesium chloride density gradients. The resulting DNA samples were digested with HindIII and EcoRV. For first dimension separation, ∼30 µg of digested DNA was loaded onto 0.35% agarose gels and fractionated at 22 volts for 42–48 hr at room temperature. Gel slices containing DNA in the 3–10 kb range were excised and positioned onto a 0.95% agarose gel. Electrophoresis in the second dimension was performed at 4°C at 80 volts for 17 hr at room temperature and 130 volts for another 1.5 hr. Following transfer to nylon membranes (Hybond-XL, GE Healthcare), samples were hybridized with a 280 bp ARS305 DNA fragment PCR amplified from genomic DNA using the following primers:
5′-CTCCGTTTTTAGCCCCCCGTG-
5′-GATTGAGGCCACAGCAAGACCG
The PCR product was radio-labeled (Megaprime DNA labeling system, GE Healthcare) and hybridized using Southern blot procedures as previously described [83].
Protein extracts were prepared by mechanical beakage of cells in 20% TCA as previously described [36]. 6% SDS-PAGE gels were used to fractionate samples for analysis of Sgs1p-myc and Pds1p-myc, while 12% SDS-PAGE gels (acrylamide: bis = 30:0.39) were used to analyze phosphorylated species of Rad53p. α-myc (9E10, 1:1000, Covance), α-Rad53p (SC-6749, 1∶2000, Santa Cruz), and HRP conjugated secondary (1∶25,000; Jackson ImmunoResearch) antibodies were used for immunoblotting.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.